meta(backlog): causal map and execution order for Gentle Pi's open backlog

[barbatdev]

Gentle Pi Open Backlog — Causal Survey

• Snapshot: strict-v1-88113a3da8ce1ac3212facb363d91144184de689f094ec1ea8c249dca3045472
• Cutoff: 2026-09-01T13:25:12.459361Z
• Frozen main: 6c391ad9e6fc2e4ca5931a7c22eaf08b97412162
• Inventory: 85 issues / 32 PRs / 84 roots
• External mapping: every frozen issue is assigned exactly once; every frozen PR has exactly one action.

P0 queue

• None.

Open PR actions

• feature-review: 5
• rebase-resolve: 15
• retain-and-verify: 10
• split-scope: 2

Maintainer decisions

• D-001 (unresolved): Does the root cause of background-subagent capability require the reported prerequisites (first supported runtime, minimum adapter version, selection mechanism, task mode, worktree preflight), or is the current ready|absent capability sufficient for initial deployment?
• D-002 (unresolved): Should the inline fast-path classification for non-SDD verification commands use command presence as an unconditional trigger (as assets/orchestrator.md and assets/orchestrator-delegation.md define), or require a more discriminating predicate to avoid misclassifying small tasks?
• D-003 (unresolved): What are the four approval-blocking decisions and the one enable/disable barrier policy that must be frozen in root-419 before the task-activity subsystem can proceed (closed-code vocabulary, cancellation cut-off, shutdown bound, concurrency limit, and the enable/disable policy)?
• D-004 (unresolved): Should PR 39's investigation be abandoned due to staleness, or does post-v0.6.0 harness architecture still make its findings relevant?
• D-005 (unresolved): Should stale PRs 195, 281, 539 be closed due to lack of contributor engagement (14+ days since last activity), or should they be pinged and preserved?
• D-006 (unresolved): Should the root-370 fix (relay/handshake bug) be implemented as a single PR, or split given the suspected 400-line budget risk and the complexity of diffing relay/handshake code between versions?
• D-007 (unresolved): For PRs 420, 426, 435, 439, 457: should maintainers rebase and attempt resolution of each conflicted PR, or close them and require fresh implementations given the divergence?
• D-008 (unresolved): Should PR 387 be merged (deferring to fix(orchestrator): prevent unauthorized execution-surface escalation #379's selected-authority seam) or closed, given that fix(orchestrator): prevent unauthorized execution-surface escalation #379's state and the selected-authority seam ownership are unknown?
• D-009 (unresolved): Should PR 504's duplicate declaration be confirmed as a real bug before merging, or is it a diff-context misplacement that should be resolved during rebase?
• D-010 (unresolved): Should the five Open Decisions in root-405 (Bash tool_call protection design) be put to maintainer vote before any implementation branch is created?

Next 10 high-value actionable roots

1. root-107 · P1 — fix(quiet-tools): preserve Pi's configured Bash shellPath on Windows #107. Read extensions/quiet-tools.ts (or the equivalent extension entry point) to verify whether createBashTool receives and forwards settings.json.shellPath from the Pi configuration, and inspect TOOL_CREATORS registration.
2. root-160 · P1 — fix(codegraph): execute npm shim safely on Windows #160, bug(codegraph): MCP codegraph tool fails on Windows (execFile without shell cannot spawn .cmd shim) #500. Verify bug(codegraph): MCP codegraph tool fails on Windows (execFile without shell cannot spawn .cmd shim) #500's environment captures on frozen main reference the same codegraph-tools.ts file or equivalent MCP adapter to confirm exact line parity
3. root-230 · P1 — bug(review): corrected candidate drops staged initial tree and counts the full slice as the correction #230. Read the FINALIZE handler in the review integration codebase (lib/review-integration-v2.ts, runtime/native-review-cli.mjs, and the authority store write path); confirm whether the corrected candidate is composed from staged base + unstaged diff or taken as-is, and whether actual_correction_lines is computed from the diff or from the initial staged total.
4. root-327 · P1 — bug(agents): builtin agent source is permanently empty — builtinAgentDirs probes agents/ directories the subagents packages never ship #327. Read extensions/gentle-ai.ts around line 327 to inspect builtinAgentDirs and the comment documenting the expected agents/ path. Verify pi-subagents package contents for agents/ directory.
5. root-343 · P1 — Candidate views leak 32 MB per session into the gentle-ai store and are never swept #343. Read lib/review-candidate-view.ts around line 596 to inspect the session_shutdown listener registration and verify absence of a startup cleanup sweep for candidate views.
6. root-348 · P1 — fix(orchestrator): enforce inline skill output contracts #348. Identify and read the orchestrator module responsible for inline skill execution and output contract enforcement. Verify whether loaded skill contracts are applied consistently.
7. root-369 · P1 — fix(skill-registry): mirror Pi-resolved loaded skills #369. Locate registry generation entry-point (skill-registry.ts or atl/skill-registry.md); confirm it reads fixed loose roots instead of before_agent_start.systemPromptOptions.skills; patch to consume Pi-resolved loaded-skill records
8. root-379 · P1 — fix(orchestrator): prevent unauthorized execution-surface escalation #379. Read extensions/gentle-ai.ts at snapshot 6c391ad and locate the tool_call handler; verify whether openProjectPaneIfMissing is checked before UI/tool execution and whether gentle-ai.execution-surface.external-fallback-prohibited is returned.
9. root-455 · P1 — fix(review): preserve consent bindings across active Pi sessions #455. Read lib/review-integration-v2.ts and runtime/review-integration-v2.mjs for consent binding storage and answerConsentBinding.
10. root-464 · P1 — bug(release): latest v2.2.0 still strands corrected lineages before last-event migration #464. Inspect package-local .gentle-ai Go binary diff between shipped v2.4.0 and main for review_facade.go correction preflight and FINALIZE.

Recent movement

• PR docs(readme): document review-mode and sdd-preflight commands #428 updated 2026-09-01T13:20:55Z — docs(readme): document review-mode and sdd-preflight commands
• Issue feat(review): run reviewer roles as Pi subagents (fresh constrained executor per lens) #544 updated 2026-09-01T11:39:17Z — feat(review): run reviewer roles as Pi subagents (fresh constrained executor per lens)
• Issue review happy-path MCP traffic exhausts agent context: full collectBindings inlined in status and re-pasted twice per capture #543 updated 2026-09-01T11:39:16Z — review happy-path MCP traffic exhausts agent context: full collectBindings inlined in status and re-pasted twice per capture
• Issue gentle_review_capture rejects ordinary collect bindings — forecast ack resubmit and even first no-ack capture fail (capture-binding-rejected) #509 updated 2026-08-31T23:48:07Z — gentle_review_capture rejects ordinary collect bindings — forecast ack resubmit and even first no-ack capture fail (capture-binding-rejected)
• PR test/ci: harden Windows test baseline and enable Windows CI #539 updated 2026-08-31T23:44:48Z — test/ci: harden Windows test baseline and enable Windows CI
• PR fix(skills/review): dedup drive-less URLs and sanitize scoped git config #540 updated 2026-08-31T23:37:59Z — fix(skills/review): dedup drive-less URLs and sanitize scoped git config
• PR fix(review): support pre-start untracked selection #532 updated 2026-08-31T18:34:02Z — fix(review): support pre-start untracked selection
• Issue feat(review): launch 4R lens capture subprocesses concurrently in the host relay #533 updated 2026-08-31T18:05:11Z — feat(review): launch 4R lens capture subprocesses concurrently in the host relay
• Issue Windows: CodeGraph Pi tool reports ENOENT because execFile("codegraph") does not resolve npm/Scoop shims #531 updated 2026-08-31T16:53:15Z — Windows: CodeGraph Pi tool reports ENOENT because execFile("codegraph") does not resolve npm/Scoop shims
• Issue bug(orchestrator): command presence forces subagent verification for bounded checks #463 updated 2026-08-31T13:17:17Z — bug(orchestrator): command presence forces subagent verification for bounded checks

Audit health

• Snapshot stabilization: PASS.
• Exact issue partition: PASS.
• Exact PR action coverage: PASS.
• Conservative unsupported-grouping fallback: applied.
• Publication: baseline published as Gentle Pi issue meta(backlog): causal map and execution order for Gentle Pi's open backlog #545; future updates follow the immutable-baseline and append-only delta contract.

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

[barbatdev]

Immutable baseline survey shard

Snapshot strict-v1-88113a3da8ce1ac3212facb363d91144184de689f094ec1ea8c249dca3045472 · cutoff 2026-09-01T13:25:12.459361Z · main 6c391ad9e6fc2e4ca5931a7c22eaf08b97412162

root-37 · P2

• Members: Approve SDD workflow guardrails and reliability enforcement #37 Approve SDD workflow guardrails and reliability enforcement
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: medium
• Causal hypothesis: The blocker is procedural: repository policy requires status:approved, the issue remains unlabelled, and the authoring token cannot create or administer labels. The proposed guardrail implementation itself is not confirmed on current main by the supplied search evidence.
• Current-main evidence:
  • issue record labels: [] — The frozen issue snapshot shows no status:approved label.
  • search term: status:approved — The supplied current-main search did not surface evidence of the typed guardrail builders or enforcement gates claimed in the issue body.
  • assets/chains/sdd-full.chain.md:70 — Current main contains SDD lifecycle chain text referring to apply, verify, sync, archive, and workload handling, but that is not proof the new guardrail enforcement exists.
• Rollback boundary: Components: SDD workflow governance and PR approval process
• Evidence gap: No maintainer approval is present in the issue metadata, and the prepared commits or guardrail artifacts were not validated against the frozen main SHA by the supplied evidence.
• Exact next action: Have a maintainer review the listed scope and either add status:approved to issue 37 or request scope changes before any PR is opened.

root-39 · P1

• Members: feat(harness): strengthen harness persistence using Pi's context and compaction events #39 feat(harness): strengthen harness persistence using Pi's context and compaction events
• Active PRs: fix(harness): refresh orchestrator prompt from disk on each access #51, feat(harness): re-anchor discipline via context event on every LLM call #52, feat(harness): re-anchor after session compaction #53, feat(harness): apply slim rules layer to named and SDD agents #54
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: true
• Confidence: low
• Causal hypothesis: Hypothesis names specific mechanisms (before_agent_start missing per-LLM-call steering, orchestratorPromptCache read-once, missing context/session hooks, empty gentlePrompt for named/SDD agents) but provided snippets are entirely unrelated (lock-file deps, provider-relay regexes, theme color defs, capability fixture schemas, test fixture JSON, installer scripts).
• Current-main evidence:
  • No retained current-main proof; see evidence gap.
• Rollback boundary: Components: extensions/gentle-ai.ts prompt injection and harness lifecycle hooks
• Evidence gap: All 60 provided snippets are irrelevant to the issue's target files and code paths.
• Exact next action: Fetch and diff extensions/gentle-ai.ts:1600-1700 and 2200-2220 for per-call gentlePrompt assembly; read agent/src/agent-loop.ts:250-280 for before_agent_start event emission; grep for 'session_before_compact', 'session_compact', and 'gentlePrompt' to verify hook registration.

root-55 · P2

• Members: bug(persona): neutral mode still leaks voseo — negative constraint fails on regular-verb present indicative forms #55 bug(persona): neutral mode still leaks voseo — negative constraint fails on regular-verb present indicative forms
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: medium
• Causal hypothesis: The persona-single-channel design (design.md) consolidated persona constants into the wrapper block, making NEUTRAL_PERSONA_PROMPT self-contained and explicit about avoiding voseo. Tests verify neutral mode does not contain 'answer in natural Rioplatense Spanish with voseo' (tests/persona-neutral-voseo.test.ts), and runtime-harness.mjs assertions confirm neutral mode excludes voseo instructions even after /reload.
• Current-main evidence:
  • openspec/changes/persona-single-channel/design.md:2-8 — design.md confirms buildGentlePrompt wrapper is the single canonical home for identity + persona + reply-language STYLE, and both named/SDD agents receive empty gentle-ai.ts:2204-2208, meaning the persona prompt is the authoritative source.
  • openspec/changes/persona-single-channel/design.md:11-17 — design.md § Proposal correction confirms self-contained __testing.buildGentlePrompt(persona) export is the test seam, isolating persona rendering from orchestrator content.
  • tests/persona-neutral-voseo.test.ts:2-8 — tests/persona-neutral-voseo.test.ts:2-8 documents that buildGentlePrompt neutral output must never include voseo encouragement.
  • tests/persona-neutral-voseo.test.ts:8-14 — tests/persona-neutral-voseo.test.ts:14 shows the explicit regex assert.doesNotMatch(/answer in natural Rioplatense Spanish with voseo/i) against neutral prompt output.
  • tests/runtime-harness.mjs:424-430 — tests/runtime-harness.mjs:429-430 asserts neutral prompt contains 'Do not use slang or regional expressions' after reload.
  • tests/runtime-harness.mjs:429-435 — tests/runtime-harness.mjs:431-435 asserts neutral prompt must NOT match 'When the user writes Spanish, answer in natural Rioplatense Spanish with voseo' after reload, with the error message 'neutral persona prompt must not include unconditional voseo instructions after reload'.
• Rollback boundary: Components: neutral persona prompt injection
• Evidence gap: No snippet proves the exact NEUTRAL_PERSONA_PROMPT text (the actual voseo prohibition wording), only test assertions. A direct grep of the persona constant would close this gap completely.
• Exact next action: Print extensions/gentle-ai.ts NEUTRAL_PERSONA_PROMPT constant (approximately lines 156-164) and verify the voseo prohibition uses a morphological rule ('do not use voseo conjugations') rather than only irregular examples.

root-57 · P2

• Members: fix(compaction): preserve user language preference after context compaction #57 fix(compaction): preserve user language preference after context compaction
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: low
• Causal hypothesis: Hypothesis attributes language-preference loss to compaction summaries written in English, where the agent infers conversation language from the summary. No provided snippet references context compaction, compaction summary generation, language_preference_source, must_preserve_after_compaction, or agent-language recovery logic. All 60 snippets are lock-file entries, installer scripts, package-verify paths, provider-relay matrix regexes, and persona design docs—none touching agent-session.ts compaction event handling.
• Current-main evidence:
  • No retained current-main proof; see evidence gap.
• Rollback boundary: Components: session compaction and reply-language state preservation
• Evidence gap: Zero snippets cover compaction lifecycle hooks, session recovery prompts, language-pref preservation state, or agent response language inference after compaction. The issue specifically concerns Pi's agent-compiler behavior (contextBeforeCompact, contextCompact, session_compact) which is not present in the provided material.
• Exact next action: Obtain agent/src/agent-session.ts sections handling session_before_compact / session_compact hooks and the recovery prompt generation; grep for 'user_reply_language', 'language_preference_source', and 'must_preserve_after_compaction' across the repo to confirm whether a compaction-aware language-pref preservation mechanism exists on current-main.

root-58 · P2

• Members: feat(model-profiles): add reusable model configuration profiles #58 feat(model-profiles): add reusable model configuration profiles
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: Unproven design gap: /gentle:models appears oriented around saved model assignments rather than named profile bundles; no confirmed current-main profile CRUD surface exists.
• Current-main evidence:
  • assets/sdd-orchestrator-workflow.md:240 — Current-main documentation states /gentle:models writes .pi/gentle-ai/models.json and the package applies saved assignments, indicating model routing is persisted at assignment level.
  • search term: model-profiles — The supplied current-main search did not surface lib/model-profiles.ts, model-profile-panel, or profile CRUD evidence, though bounded absence is not proof.
• Rollback boundary: Components: model configuration UX and local settings persistence
• Evidence gap: Need current-main inspection of extensions/gentle-ai.ts and model configuration helpers to confirm absence of profile storage and decide whether profiles should include orchestrator model/thinking settings.
• Exact next action: Inspect current main /gentle:models and settings persistence to confirm what model assignments already exist, then define a small accepted first-slice profile schema and storage boundary.

root-59 · P2

• Members: feat(sdd): implement project-generated TDD rubric for Pi SDD init #59 feat(sdd): implement project-generated TDD rubric for Pi SDD init
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: Unproven implementation cause: extensions/sdd-init.ts may treat detection.testCommand presence as strictTdd Boolean, causing all projects with a runner to receive the same Strict TDD policy; current-main search does not show that source line.
• Current-main evidence:
  • scripts/verify-package-files.mjs:53 — Current main packages and validates extensions/sdd-init.ts, confirming the file is part of the SDD surface named in the issue.
  • assets/agents/sdd-apply.md:94 — Current-main apply instructions load Strict TDD behavior based on config declaration or parent prompt, not method-specific rubric rows.
  • assets/agents/sdd-verify.md:65 — Current-main verify instructions also key verification evidence from strict TDD activation rather than rubric methods.
• Rollback boundary: Components: SDD init, orchestrator routing, strict TDD forwarding
• Evidence gap: Need current-main code proof of the Boolean detection in extensions/sdd-init.ts and an upstream contract review of feat(sdd): add project-generated TDD rubric before strict_tdd enforcement gentle-ai#986 before designing Pi serialization.
• Exact next action: Read the frozen-main extensions/sdd-init.ts implementation around test-runner detection and inspect the referenced cross-agent feature contract before opening any Pi implementation PR.

root-62 · P2

• Members: bug(agents): YAML tools frontmatter leaves Gentle subagents without filesystem tools #62 bug(agents): YAML tools frontmatter leaves Gentle subagents without filesystem tools
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: high
• Causal hypothesis: All provided snippets are installer, package-verification, and contract-mirror scripts; none contain YAML frontmatter parsing, agent loading, or tool-registration code relevant to the reported mechanism.
• Current-main evidence:
  • No retained current-main proof; see evidence gap.
• Rollback boundary: Components: packaged agent tools runtime provisioning
• Evidence gap: Zero snippets from agent-loading, frontmatter parsing, or subagent tool-initialisation paths. All 62 fragments are from scripts/gentle-ai-installer.mjs, scripts/verify-package-files.mjs, scripts/measure-native-authority-slimming.mjs, scripts/install-gentle-ai.mjs, scripts/check-provider-contract.mjs, scripts/mirror-provider-contract.mjs, and scripts/test-packed-runner.mjs.
• Exact next action: Read assets/agents/review-reliability.md to inspect its frontmatter tools block, then read the agent-loading/frontmatter-parsing code (likely in src/ or assets/ agents loader) that converts YAML frontmatter tool lists to child-session tool registrations.

root-63 · P2

• Members: gentle-pi references worker/scout/context-builder subagents but does not package them #63 gentle-pi references worker/scout/context-builder subagents but does not package them
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: Documentation and the packaging validator evolved around named SDD/review/JD assets and a gentle-ai-worker asset, while older or aspirational generic subagent names were left in user-facing guidance without a packaging or availability guard.
• Current-main evidence:
  • skills/gentle-ai/SKILL.md:47 — Current main search shows a line recommending scout/context-builder when context-heavy and one worker.
  • skills/gentle-ai/SKILL.md:53 — Current main search shows the multi-file write rule tells the reader to use one worker.
  • scripts/verify-package-files.mjs:16 — Search output shows the required-path inventory enumerates SDD/review/JD agent assets, including gentle-ai-worker.md, but not worker.md, scout.md, or context-builder.md in the returned excerpt.
• Rollback boundary: Components: docs/package assets/subagent orchestration
• Evidence gap: Truncated current_main_search does not confirm README.md or assets/orchestrator.md still contain the named generic agents, does not provide a complete assets/agents and installer inventory, and does not provide clean-install subagent_list_agents output for frozen main.
• Exact next action: At frozen SHA a6a73fa, run a complete current-main inventory: git ls-files assets/agents scripts/gentle-ai-installer.mjs scripts/verify-package-files.mjs, rg -n 'worker|scout|context-builder' README.md assets/orchestrator.md skills/gentle-ai/SKILL.md assets/agents scripts, and capture subagent_list_agents from a clean install; then either add generic agents to packaging or update docs and add an availability check.

root-67 · P2

• Members: Improve prompt cache ordering for el Gentleman harness #67 Improve prompt cache ordering for el Gentleman harness
• Active PRs: perf(prompt): improve cache-friendly prompt ordering #68
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: The stable Gentle AI™ identity/orchestrator block appears after volatile runtime context, shortening the reusable prompt prefix; prepending it before event.systemPrompt for parent sessions should improve cache prefix stability while leaving named/SDD executor behavior unchanged.
• Current-main evidence:
  • current_main_search terms: before_agent_start, event.systemPrompt, prompt, runtime, stable, before, parent — The supplied bounded current-main search is truncated and visible matches point to installer/runtime package files, not the parent prompt assembly or ordering code; this cannot confirm whether current main still appends the stable prompt after the runtime prompt or already prepends it.
  • issue comment https://github.com/Gentleman-Programming/gentle-pi/issues/67#issuecomment-5299938262 — A maintainer comment confirms the cache-ordering hypothesis and cites PR perf(prompt): improve cache-friendly prompt ordering #68 as matching scope with an ordering regression test, but the supplied record does not show PR perf(prompt): improve cache-friendly prompt ordering #68 merge status or current-main code evidence.
• Rollback boundary: Components: prompt assembly / before_agent_start parent session integration
• Evidence gap: Need inspect current main before_agent_start prompt assembly and PR perf(prompt): improve cache-friendly prompt ordering #68 test/merge state to prove whether the ordering improvement is already present, still missing, or merely proposed.
• Exact next action: Inspect current main for the parent before_agent_start hook or equivalent prompt assembly that manipulates event.systemPrompt, confirm whether the stable Gentle AI™ block is prepended before volatile runtime context, and run or inspect the PR perf(prompt): improve cache-friendly prompt ordering #68 ordering regression test against frozen main a6a73fa.

root-83 · P2

• Members: feat(models): add provider-aware SDD model presets #83 feat(models): add provider-aware SDD model presets
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: low
• Causal hypothesis: Feature gap hypothesis: /gentle:models implements per-agent model/thinking assignment, bulk set-all, export and restore, but has no SDD-phase→role-tier→provider-profile resolver and no preset-drift detection. Hypothesis is drawn from the issue body; the supplied current-main search did not surface the models command implementation or any preset/profile module, so absence on frozen main is not yet proven.
• Current-main evidence:
  • scripts/verify-package-files.mjs:20-38 — Frozen main packages the exact agent assets the requested tiers target (assets/agents/jd-judge-a.md, jd-judge-b.md, sdd-design.md, sdd-proposal.md, sdd-verify.md, plus sdd-apply per the agent list), so the role classification in the proposal maps to real, shipped agents.
  • current_main_search terms '/gentle:models', 'model', 'thinking', 'Set all agents' (truncated: true) — Returned matches are installer/package/provider-matrix files (scripts/gentle-ai-installer.mjs, scripts/verify-package-files.mjs, scripts/maintainer/provider-relay-matrix.mjs); no preset or provider-profile artifact appears among the matches. Because the search is bounded and truncated, this is weak negative evidence only, not proof that no preset code exists.
  • scripts/verify-package-files.mjs:221-233 — Frozen main already enforces a single authoritative Gentle AI version constant shared between lib/gentle-ai-binary.ts GENTLE_AI_VERSION and scripts/gentle-ai-installer.mjs INSTALLER_VERSION, so any new preset/provider-profile data would need the same pinned-drift discipline.
• Rollback boundary: Components: /gentle:models command + models.json routing config (extensions/gentle-ai.ts / lib models module)
• Evidence gap: No direct read at a6a73fa of the /gentle:models handler, the models.json reader/writer, or extensions/gentle-ai.ts to confirm that (a) presets and provider profiles are absent, (b) the per-agent {model, thinking} shape is exactly as described, and (c) 'Set all agents'/export/restore exist as claimed. Also no evidence of where a provider-detection signal would come from (available provider/model catalog) to resolve tiers to concrete model IDs, nor any test baseline for models.json application paths.
• Exact next action: At frozen SHA a6a73fa, open the /gentle:models implementation and the models.json load/apply module (extensions/gentle-ai.ts and any lib/models routing file) and grep for 'preset', 'providerProfile', 'tier' to confirm the gap, then record whether the per-agent {model,thinking} shape can absorb a preset layer without schema change.

root-107 · P1

• Members: fix(quiet-tools): preserve Pi's configured Bash shellPath on Windows #107 fix(quiet-tools): preserve Pi's configured Bash shellPath on Windows
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: high
• Causal hypothesis: All provided snippets are installer, package-verification, native-review-cli runtime, and contract-mirror scripts; none contain the quiet-tools VS Code extension source, TOOL_CREATORS, or createBashTool.
• Current-main evidence:
  • No retained current-main proof; see evidence gap.
• Rollback boundary: Components: extensions/quiet-tools
• Evidence gap: Zero snippets from extensions/quiet-tools.ts, the bash tool creation function, or any settings.json/shellPath resolution path. All 60+ fragments are from scripts/gentle-ai-installer.mjs, scripts/verify-package-files.mjs, scripts/measure-native-authority-slimming.mjs, scripts/install-gentle-ai.mjs, and runtime/native-review-cli.mjs.
• Exact next action: Read extensions/quiet-tools.ts (or the equivalent extension entry point) to verify whether createBashTool receives and forwards settings.json.shellPath from the Pi configuration, and inspect TOOL_CREATORS registration.

root-116 · P2

• Members: feat: Add OpenCode MCP Bridge with Antigravity IDE error handling #116 feat: Add OpenCode MCP Bridge with Antigravity IDE error handling
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: low
• Causal hypothesis: OpenCode environment lacks codegraph MCP state pre-population and error interception for the community-tool:codegraph unavailable pattern, especially when Antigravity IDE detection produces the reported error.
• Current-main evidence:
  • current_main_search terms: CODEGRAPH_ERROR_PATTERN, readOpenCodeConfig, detectCodeGraphServer, prePopulateCodeGraphState, interceptCodeGraphError, codegraph — The supplied bounded current-main search output only surfaced unrelated installer error handling matches and did not show exact proposed symbols or OpenCode/codegraph bridge handling. Because the search is truncated and bounded, this is not proof that the functionality is absent or present on current main.
• Rollback boundary: Components: OpenCode MCP bridge / codegraph integration
• Evidence gap: No reproduction on frozen current main, no confirmation that the reported OpenCode/codegraph error occurs there, and no proof of current-main implementation status for OpenCode MCP bridge codegraph handling.
• Exact next action: Reproduce the reported codegraph MCP unavailable error in OpenCode on frozen current main and inspect current main for existing OpenCode MCP bridge/codegraph handling before deciding whether implementation or documentation is needed.

root-125 · P2

• Members: My pi writes me spanish even I ask things with plain english #125 My pi writes me spanish even I ask things with plain english
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: high
• Causal hypothesis: All provided snippets are installer, package-verification, contract-mirror, and native-review-cli scripts; none contain language/locale handling, persona language rules, or prompt template code that could cause Spanish responses.
• Current-main evidence:
  • No retained current-main proof; see evidence gap.
• Rollback boundary: Components: persona/language selection
• Evidence gap: Zero snippets from persona configuration, system prompt templates, language detection/selection code, or any Spanish language rule definition. All 62 fragments are from scripts/gentle-ai-installer.mjs, scripts/verify-package-files.mjs, scripts/measure-native-authority-slimming.mjs, scripts/install-gentle-ai.mjs, scripts/check-provider-contract.mjs, scripts/mirror-provider-contract.mjs, and runtime/native-review-cli.mjs.
• Exact next action: Read the system prompt templates, persona configuration files, and any language/locale handling code in the pi-subagents or pi core to determine whether Rioplatense Spanish reply-language instructions are active on main.

root-160 · P1

• Members: fix(codegraph): execute npm shim safely on Windows #160 fix(codegraph): execute npm shim safely on Windows, bug(codegraph): MCP codegraph tool fails on Windows (execFile without shell cannot spawn .cmd shim) #500 bug(codegraph): MCP codegraph tool fails on Windows (execFile without shell cannot spawn .cmd shim)
• Active PRs: fix(codegraph): add injectable shell-free runner #163
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: high
• Causal hypothesis: Both issues trace to the same execFile("codegraph", args) call without shell: true in extensions/codegraph-tools.ts (line 181), which returns ENOENT/EINVAL on Windows where npm installs .cmd/.ps1 shims
• Current-main evidence:
  • extensions/codegraph-tools.ts:181-185 — The failing execFile("codegraph", args) call without shell option is the shared defect present on frozen main
  • extensions/codegraph-tools.ts:167-170 — ENoENT from that execFile maps to UNAVAILABLE status, producing identical user-visible behavior in both issues
  • extensions/codegraph-tools.ts:230-231 — Both issues affect the same registered tool entry point, confirming a single causal path to all codegraph operations
• Rollback boundary: Components: extensions/codegraph-tools.ts (MCP codegraph tool), extensions/codegraph-tools.ts (integrated CodeGraph tool / external process execution on Windows)
• Evidence gap: Issue bug(codegraph): MCP codegraph tool fails on Windows (execFile without shell cannot spawn .cmd shim) #500 provided no codegraph-tools.ts snippet in its data; adjudication relies on the supplied authoritative file matching fix(codegraph): execute npm shim safely on Windows #160's described defect pattern exactly
• Exact next action: Verify bug(codegraph): MCP codegraph tool fails on Windows (execFile without shell cannot spawn .cmd shim) #500's environment captures on frozen main reference the same codegraph-tools.ts file or equivalent MCP adapter to confirm exact line parity

root-185 · P2

• Members: fix(review): clean forecast-restored worktrees and cover the non-negotiated forecast path #185 fix(review): clean forecast-restored worktrees and cover the non-negotiated forecast path
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: Causal hypothesis centers on restoreForFinalizeFromNative, cleanupTerminal, correctionForecast, and non-negotiated status path; provided snippets are migration manifests and package-file hashes with zero relevant code
• Current-main evidence:
  • No retained current-main proof; see evidence gap.
• Rollback boundary: Components: native review candidate restore/worktree cleanup
• Evidence gap: No evidence from lib/review-integration-v2.ts, review reconciliation, forecast, worktree, or finalize paths
• Exact next action: Fetch and read lib/review-integration-v2.ts on main SHA searching for restoreForFinalizeFromNative, cleanupTerminal, correctionForecast, targetStatus, and worktree lifecycle to verify cleanup coverage

root-203 · P2

• Members: bug(review): NUL Git config breaks authority probes on Windows MSYS #203 bug(review): NUL Git config breaks authority probes on Windows MSYS
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: Causal hypothesis requires reviewGitEnvironment() assigning NUL to GIT_CONFIG_GLOBAL/GIT_CONFIG_SYSTEM on win32; all provided snippets come from installer, package-file verification, migration, and measurement scripts—zero code from the actual review Git environment or authority resolution path
• Current-main evidence:
  • No retained current-main proof; see evidence gap.
• Rollback boundary: Components: review-git-environment (Windows runtime parity)
• Evidence gap: No evidence from lib/review-repository.ts, lib/review-canonical.ts, or any reviewGitEnvironment/git rev-parse/MSYS NUL handling code
• Exact next action: Fetch and read lib/review-repository.ts on main SHA searching for reviewGitEnvironment, GIT_CONFIG_GLOBAL, GIT_CONFIG_SYSTEM, /dev/null, and NUL assignments to confirm whether the NUL sanitization still occurs on win32

root-213 · P2

• Members: bug(review): native status timeout leaves authority inventory unavailable #213 bug(review): native status timeout leaves authority inventory unavailable
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: Hypothesis: native status pre-flight inventory scan aborts on mixed-generation store (graph-v1 CURRENT missing, v2 LOCK persists) with retry_safe=false/next_action=stop. Snippets do not cover pre_native inventory logic, subcommand timeout handling, or the specific inventory-complete=false status path. Only decoder and contract docs are provided.
• Current-main evidence:
  • runtime/native-review-cli.mjs:1098-1104 — decodeNativeReviewStatus enforces 'incomplete inventory cannot be authoritative'
  • runtime/native-review-cli.mjs:1595-1601 — lock status includes OWNED, AMBIGUOUS, and unknown-stays-fail-closed
• Rollback boundary: Components: review authority / native review status adapter (graph-v1 CURRENT quorum + legacy v2 transaction lock inventory, Pi review controller disposition mapping)
• Evidence gap: No snippet showing the status pre_native flow, inventory scan timing, subcommand timeout handling, or how mixed-generation lock+missing-CURRENT is handled before status decoding.
• Exact next action: Run gentle-ai review status --cwd <repo> --next-transition with graph-v1 CURRENT absent and review-transactions/v2/LOCK present; capture the full status envelope including complete, authoritative, and locks fields.

root-215 · P1

• Members: bug(review): native snapshots drop gitlinks and create divergent review authority #215 bug(review): native snapshots drop gitlinks and create divergent review authority
• Active PRs: none
• Historical PR evidence: fix(review): preserve absent gitlinks in snapshots #217
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: Hypothesis: SnapshotBuilder.buildCurrentChanges runs git add -u -- . in a Pi metadata-only index, removing mode-160000 gitlink entries. No provided snippet shows buildCurrentChanges, workspace projection, index construction, or git add -u usage.
• Current-main evidence:
  • No retained current-main proof; see evidence gap.
• Rollback boundary: Components: native/snapshot-builder
• Evidence gap: No snippet of the buildCurrentChanges method, workspace projection, index snapshot, or the gitlink preservation fix.
• Exact next action: Show diff between HEAD and snapshot for a repo with clean gitlinks; verify that git diff-tree --no-commit-id -r --raw <base> <candidate> does not report mode 100644→0 as deletions, or read the buildCurrentChanges method at runtime/native-review-cli.mjs.

root-216 · P1

• Members: bug(sdd): sdd-init hardcodes unavailable provider model #216 bug(sdd): sdd-init hardcodes unavailable provider model
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: Hypothesis: sdd-init agent frontmatter in assets/agents/sdd-init.md pins openai-codex/gpt-5.3-codex, overriding model-profile resolution. The file exists (listed in verify-package-files.mjs) but its content is not included in the provided snippets.
• Current-main evidence:
  • No retained current-main proof; see evidence gap.
• Rollback boundary: Components: sdd-init package-managed agent asset / subagent model resolution / installer managed copy
• Evidence gap: No snippet of sdd-init.md content, model_profiles config, or the subagent model resolver.
• Exact next action: Read assets/agents/sdd-init.md frontmatter and check if the model field remains set to a provider-specific value; also read subagents.json and the model resolution code in the Pi host.

root-230 · P1

• Members: bug(review): corrected candidate drops staged initial tree and counts the full slice as the correction #230 bug(review): corrected candidate drops staged initial tree and counts the full slice as the correction
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: high
• Causal hypothesis: Hypothesis centers on review correction FINALIZE composing the corrected candidate incorrectly (base/candidate swap or missing staged+unstaged composition). Snippets are installer and package-verification scripts; none cover the review authority store, candidate diff computation, or FINALIZE logic.
• Current-main evidence:
  • scripts/gentle-ai-installer.mjs:74-80 — Snippet from gentle-ai-installer.mjs about prerelease asset mirrors; not review correction code
  • scripts/verify-package-files.mjs:19-25 — Snippet from verify-package-files.mjs listing package assets; not review correction code
  • scripts/verify-package-files.mjs:121-127 — Snippet from verify-package-files.mjs showing contract hash fixtures; not review correction code
• Rollback boundary: Components: review-finalize-correction-snapshot-accounting
• Evidence gap: Missing snippets from review authority / native review integration (e.g., review-integration-v2.ts, review-authority store code, native-review-cli.mjs FINALIZE handler, candidate/tree composition logic). No diff of the exact corrected candidate vs. staged + unstaged is present.
• Exact next action: Read the FINALIZE handler in the review integration codebase (lib/review-integration-v2.ts, runtime/native-review-cli.mjs, and the authority store write path); confirm whether the corrected candidate is composed from staged base + unstaged diff or taken as-is, and whether actual_correction_lines is computed from the diff or from the initial staged total.

root-234 · P2

• Members: feat(skill-registry): add opt-out for .atl gitignore mutation #234 feat(skill-registry): add opt-out for .atl gitignore mutation
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: The skill registry startup path appends .atl/ to .gitignore after checking only the local .gitignore contents, so a repository protected by a global Git ignore file still receives local mutation.
• Current-main evidence:
  • scripts/verify-package-files.mjs:54 — Bounded search evidence confirms extensions/skill-registry.ts is a packaged source file in the frozen main package list.
  • current_main_search terms: atl/, gitignore, GENTLE_PI_NO_SKILL_REGISTRY=1 — Search included the relevant behavior and opt-out terms, but no returned match showed the actual .atl/ .gitignore append path or the proposed GENTLE_PI_NO_ATL_GITIGNORE behavior.
• Rollback boundary: Components: skill-registry extension
• Evidence gap: Direct source proof is missing for the current .gitignore mutation path, and there is no reproduction evidence showing that a global ignore covering .atl/ still leads to repository-local .gitignore modification.
• Exact next action: Inspect the skill registry extension startup path for local .gitignore mutation and reproduce the reported scenario in a repository whose global Git ignore configuration already excludes .atl/.

root-238 · P2

• Members: pi-pretty never loads in compiled Pi binary (createRequire intercepted by jiti) #238 pi-pretty never loads in compiled Pi binary (createRequire intercepted by jiti)
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: high
• Causal hypothesis: Hypothesis is that extensions/pi-pretty.ts uses createRequire(packageJsonPath) with a bare specifier that jiti in the Bun-compiled Pi runtime cannot resolve via ~/.pi/agent/npm/node_modules. Snippets contain no reference to createRequire, require.resolve, or pi-pretty.
• Current-main evidence:
  • scripts/test-packed-runner.mjs:15-21 — npm path resolution in test-packed-runner.mjs; not pi-pretty loading code
  • runtime/native-review-cli.mjs:948-954 — native-review-cli.mjs transition operation set; unrelated to pi-pretty
  • tests/runtime-harness.mjs:6-12 — tests/runtime-harness.mjs imports via ES modules; no require/createRequire
• Rollback boundary: Components: extensions/pi-pretty loader and package resolution
• Evidence gap: Missing extensions/pi-pretty.ts content (the createRequire call if present), jiti/Bun resolution behavior evidence, and package.json showing the pi-pretty dependency version.
• Exact next action: Read extensions/pi-pretty.ts to find whether createRequire(createRequire(packageJsonPath)) or require.resolve is used, and whether the specifier is bare or scoped; verify jiti intercepts it under Bun; confirm the pi-pretty version pin in package.json.

root-277 · P2

• Members: feat(skill-registry): permitir excluir rutas de fuentes de skills #277 feat(skill-registry): permitir excluir rutas de fuentes de skills
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: Unproven as code-level cause. Reported hypothesis: the registry's fixed user-skill source list includes ~/.config/opencode/skills and the settings surface exposes only a global disable flag/env var, with no negative/path-filter option, so any additional client skill tree on the machine is unconditionally indexed. Needs frozen-main source inspection to confirm.
• Current-main evidence:
  • scripts/verify-package-files.mjs:54 — Frozen main's package file manifest lists "extensions/skill-registry.ts", confirming the file named in the report exists and is a shipped, verified part of the package.
  • skills/skill-registry/SKILL.md:18-20,36 — The registry contract documents writing .atl/skill-registry.md with "scanned sources" and only these reductions: skip sdd-*, _shared, skill-registry, and deduplicate by name preferring project over user. Nothing describes a per-source path exclusion, which is consistent with (but does not prove) the reported all-or-nothing controls.
  • skills/skill-registry/SKILL.md:19,37 — Registry persistence to Engram uses topic_key skill-registry with capture_prompt:false; the indexing contract is source-list driven, so a widened fixed source set would propagate directly into the agent-visible index.
  • bounded current_main_search terms: "config/opencode/skills", "excluir", "rutas", "skill-registry" (truncated: true) — The truncated search returned no hit for a hardcoded ~/.config/opencode/skills path and no hit for an exclusion option, and it did not surface the body of extensions/skill-registry.ts at all. Because the search was truncated and lacked the target file's contents, this is absence of evidence, not evidence of absence.
• Rollback boundary: Components: extensions/skill-registry.ts (skill registry generation + watcher, .atl/skill-registry.md indexing)
• Evidence gap: Missing direct reading of extensions/skill-registry.ts at frozen main SHA a6a73fa: (1) does userSkillDirs() actually enumerate ~/.config/opencode/skills; (2) is any per-source exclusion setting already present (settings.json key, GENTLE_PI_SKILL_REGISTRY_* env var, CLI flag) but undocumented; (3) which docs/README advertise only the two global kill switches. Also missing a reproduction that the generated .atl/skill-registry.md in a Pi-only project contains SKILL.md paths resolved under ~/.config/opencode/skills.
• Exact next action: Read extensions/skill-registry.ts at frozen main and dump userSkillDirs() plus every flag/env/setting it honors, and grep the repo and docs for the exclusion surface names (GENTLE_PI_SKILL_REGISTRY, no-skill-registry, exclude), then reproduce by running the registry generation in a project with ~/.config/opencode/skills populated and inspecting the "scanned sources" section of the produced .atl/skill-registry.md.

root-284 · P2

• Members: bug(review): Windows consent relay rejects equivalent cwd paths #284 bug(review): Windows consent relay rejects equivalent cwd paths
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: high
• Causal hypothesis: Hypothesis involves Windows path normalization failure in consent-invocation cwd comparison (C:\ vs /c/). Snippets cover review repository authority resolution, Git env, and installer constants; none cover the consent relay, invocation tokenizer, or cwd path comparison logic.
• Current-main evidence:
  • lib/review-repository.ts:47-53 — reviewGitEnvironment validates Git env; does not normalize paths for consent
  • lib/review-repository.ts:245-251 — probeLiveRepositoryV1 resolves common directory via Git; not the consent cwd check
  • scripts/gentle-ai-installer.mjs:30-36 — installer.mjs WINDOWS_SYSTEM_ROOT constant; not consent relay code
• Rollback boundary: Components: review-consent-relay
• Evidence gap: Missing snippets from the consent invocation relay (invocation tokenizer, cwd normalization, provider --cwd comparison), particularly the path-equivalence logic on Windows.
• Exact next action: Locate and read the consent relay/invocation builder that compares the provider --cwd against the request cwd; check whether it uses path.normalize, realpathSync, or raw string comparison, and whether backslashes in the invocation tokenizer are treated as escape characters.

root-298 · P2

• Members: fix(docs): prompts and assets reference pi-mono repository paths and URLs that do not exist in gentle-pi #298 fix(docs): prompts and assets reference pi-mono repository paths and URLs that do not exist in gentle-pi
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: low
• Causal hypothesis: The provided 60 snippets cover installer, package-verification, contract, and skill files but zero documentation or agent-asset files. No snippet touches prompts/gpr.md, prompts/gcl.md, or assets/agents/sdd-design.md where the issue claims stale pi-mono references persist.
• Current-main evidence:
  • scripts/verify-package-files.mjs:76-82 — verify-package-files.mjs enforces required paths but lists no prompts/ or assets/agents/ paths that would prove or disprove pi-mono residue
  • scripts/mirror-provider-contract.mjs:28-34 — mirror-provider-contract.mjs declares contract mirror paths, unrelated to doc/asset path residue
• Rollback boundary: Components: assets/agents (sdd-design.md) + shipped prompt/agent instruction assets
• Evidence gap: Must read prompts/gpr.md, prompts/gcl.md, and assets/agents/sdd-design.md on strict-v1-88113a3da8ce1ac3212facb363d91144184de689f094ec1ea8c249dca3045472 to verify whether stale pi-mono paths or packages/coding-agent instructions remain.
• Exact next action: Read prompts/gpr.md, prompts/gcl.md, and assets/agents/sdd-design.md on snapshot strict-v1-88113a3da8ce1ac3212facb363d91144184de689f094ec1ea8c249dca3045472; grep for earendil-works, pi-mono, packages/coding-agent.

root-299 · P2

• Members: fix(docs): README install guidance references npm:gentle-pi@0.14.0 and an RDD version policy superseded by the published 2.1.2 line #299 fix(docs): README install guidance references npm:gentle-pi@0.14.0 and an RDD version policy superseded by the published 2.1.2 line
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: Explicitly unproven for current main. Plausible cause: the README version-policy block was written during the 0.x RDD launch and never re-synced as the package moved to 1.x/2.x and the runtime pin became installer-authoritative.
• Current-main evidence:
  • scripts/verify-package-files.mjs:376, :336, :381 — Current main enforces that Gentle AI version pins cannot drift from the authoritative INSTALLER_VERSION and byte-pins contract artifacts for a v2.5.0-rc.3 runtime, i.e. the release/version story is now installer-driven rather than README-driven.
  • runtime/native-review-cli.mjs:765, :775, :784 — Capability rows now extend to 2.4.0 and 2.5.0-rc.3; the v2.2.3 reference in the codebase is a historical grounding comment, not a current pin, which undercuts the README row quoted in the report.
  • search term "README.md:67" / "v0.14.0" — No README.md line surfaced among the bounded current-main matches, so presence or absence of the 0.14.0 install command at the frozen SHA is not established by the supplied evidence.
• Rollback boundary: Components: README.md (install + version policy sections)
• Evidence gap: No line-level proof for README.md at a6a73fa (install command, RDD version-policy block, runtime row) and no re-check of live npm dist-tags; the report's evidence is 20 releases of drift old.
• Exact next action: At the frozen SHA, grep README.md for gentle-pi@0.14.0, v0.15.0, and the current runtime version, and compare against npm view gentle-pi dist-tags plus package.json version, recording whether each quoted line survives.

root-300 · P2

• Members: fix(docs): document the registered gentle:review-mode command and align the README command table with the quick-start list #300 fix(docs): document the registered gentle:review-mode command and align the README command table with the quick-start list
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: medium
• Causal hypothesis: Documentation inventory drift: commands were registered in the extension and mentioned in Quick start, but the reference Commands table is maintained by hand and was never regenerated from the registration list.
• Current-main evidence:
  • skills/_shared/review-ledger-contract.md:23 — "Before status/START, consult effective review mode. off creates no authority or authorization and yields organic disabled/unmanaged" — the review-mode switch is a live first-class control on current main, so its absence from the command table is a real discoverability gap, not a dead feature.
  • contracts/review-integration/v2/fixtures/consent.fixture.json:33 — Shipped v2 consent fixture note: "To turn reviews off for good, run 'gentle-ai review mode disable'." — confirms the user-facing sub-actions named in the report exist in the current contract surface.
  • search term "/gentle:review-mode" in README.md — Bounded current-main matches contain no README.md lines for /gentle:review-mode, /gentle:sdd-preflight, /gentle:commit-status or /gentle:commit-abort; search evidence is not proof of absence but nothing contradicts the report.
  • skills/branch-pr/SKILL.md:31, :79, :132 — status:approved gating (present on this issue) plus PR Validation requiring the linked issue approval means the docs-only fix is unblocked and can be landed as one PR.
• Rollback boundary: Components: README.md Commands table (documentation surface for extensions/gentle-ai.ts command registrations)
• Evidence gap: No line-level read of README.md's Commands table and Quick start at a6a73fa, and no evidence of the claimed contributor PR; the maintainer verification is dated 2026-08-15, before the claimed work.
• Exact next action: At the frozen SHA, diff the README Commands table rows against the pi.registerCommand names in extensions/gentle-ai.ts, then insert the four missing rows with descriptions taken from the registrations and align the Quick start list.

root-304 · P2

• Members: fix(skill-registry): skill-registry scan is not recursive into category subdirectories #304 fix(skill-registry): skill-registry scan is not recursive into category subdirectories
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: Scan loop, per the quoted code, does entries over one directory level, tests entry.isDirectory() and probes only join(root, entry.name, "SKILL.md"), so a directory without its own SKILL.md is skipped with no fallback; a category root therefore yields zero skills.
• Current-main evidence:
  • skills/skill-registry/SKILL.md:34 — Contract on current main still says "Scan all known user and project skill directories for */SKILL.md", i.e. single-level globbing is the documented behavior, so this is at minimum a design change and not merely a code slip.
  • skills/skill-improver/SKILL.md:36 — Second shipped contract repeats the flat */SKILL.md scan, confirming the current documented expectation.
  • tests/skill-registry.test.ts:119-120 — A test fixture pairs root/docs/SKILL.md with root/fixtures/nested/SKILL.md, so nested paths are exercised in current tests; the assertion direction (recursion supported vs deliberately skipped) is not visible in the bounded excerpt and is exactly what must be checked.
  • scripts/verify-package-files.mjs:54; tests/runtime-harness.mjs:19 — extensions/skill-registry.ts is a required packaged resource and harness-compiled module, so any scan-depth change is a shipped-surface change needing test coverage.
• Rollback boundary: Components: extensions/skill-registry.ts (scan) + skills/skill-registry contract and .atl/skill-registry.md generation
• Evidence gap: No current-main source lines for the scan loop and no reading of the nested-fixture test assertions; therefore it is unproven whether recursion is still one-level-only at the frozen SHA or whether the flat behavior is an intentional contract.
• Exact next action: At a6a73fa, read extensions/skill-registry.ts:90-130 and tests/skill-registry.test.ts:110-135, then reproduce with a two-level-deep SKILL.md under a configured root via /skill-registry:refresh to record whether the nested skill appears.

[barbatdev]

Immutable baseline survey shard

Snapshot strict-v1-88113a3da8ce1ac3212facb363d91144184de689f094ec1ea8c249dca3045472 · cutoff 2026-09-01T13:25:12.459361Z · main 6c391ad9e6fc2e4ca5931a7c22eaf08b97412162

root-314 · P2

• Members: test(devbinary): native-review-parity devtests still assert the pre-#302 synthetic confirmContext grant path #314 test(devbinary): native-review-parity devtests still assert the pre-fix(review): record consent latch after granted follow-up #302 synthetic confirmContext grant path
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: low
• Causal hypothesis: The issue describes devbinary native-review-parity tests failing with 3 of 6 tests: granted and declined consent journeys aborting pre-launch on consent-invocation-cwd-changed, and host-relay raising NativeBinaryUnavailableError. The provided 60 snippets contain installer, verification, contract, and skill files — zero test code, zero review consent validation, and zero reference to NativeReviewConsentBindingError or NativeBinaryUnavailableError.
• Current-main evidence:
  • scripts/gentle-ai-installer.mjs:513-519 — gentle-ai-installer.mjs declares installLockBlockedError and tombstone logic — installer-side, unrelated to review consent binding
  • scripts/test-packed-runner.mjs:63-69 — test-packed-runner.mjs checks decoded capabilities after install — E2E pack test, not devbinary parity test
• Rollback boundary: Components: tests/devbinary/native-review-parity.devtest.ts + dev-binary lane (test:dev-binary)
• Evidence gap: Must read tests/devbinary/native-review-parity.devtest.ts and any NativeReviewConsentBindingError / consent-invocation-cwd-changed / NativeBinaryUnavailableError definitions on the current main SHA to verify whether the consent-invocation-cwd-changed abort and devbinary override resolution failures still occur.
• Exact next action: Read tests/devbinary/native-review-parity.devtest.ts on SHA 6c391ad; grep for consent-invocation-cwd-changed, NativeBinaryUnavailableError, and devbinary override resolution; verify whether the three reported failures persist.

root-323 · P1

• Members: fix(review): candidate change within the consent TTL still dead-ends at candidate-target-projection-drift #323 fix(review): candidate change within the consent TTL still dead-ends at candidate-target-projection-drift
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: low
• Causal hypothesis: The issue claims the CandidateViewRegistry replay key {cwd, lineageId, input, inputPath} carries no candidate content identity, so content changes within the consent TTL reuse a stale view, causing assertNativeStartCandidateBinding to reject as candidate-target-projection-drift. The provided 60 snippets contain installer, package-verification, contract, and skill files — zero references to CandidateViewRegistry, replayKey, assertNativeStartCandidateBinding, or the review binding assertion layer.
• Current-main evidence:
  • scripts/verify-package-files.mjs:60-66 — verify-package-files.mjs lists runtime/review-integration-v2.mjs as a required file — confirms the file exists but does not prove or disprove the CandidateViewRegistry replay-key behavior
  • scripts/verify-package-files.mjs:83-89 — verify-package-files.mjs includes review-provider-contract-mirror bundle files — irrelevant to review replay key binding logic
• Rollback boundary: Components: review consent/candidate-view binding (CandidateViewRegistry.createOrReuse replay key, assertNativeStartCandidateBinding)
• Evidence gap: Must read lib/review-integration-v2.ts (or the file containing CandidateViewRegistry and assertNativeStartCandidateBinding) on the current main SHA to verify the replay key definition and whether it includes candidate content identity, and to check the binding assertion path for within-TTL content changes.
• Exact next action: Read lib/review-integration-v2.ts on SHA 6c391ad; grep for CandidateViewRegistry, replayKey, assertNativeStartCandidateBinding, candidate-target-projection-drift; verify whether the replay key includes candidate content identity and whether within-TTL content changes cause drift rejection.

root-327 · P1

• Members: bug(agents): builtin agent source is permanently empty — builtinAgentDirs probes agents/ directories the subagents packages never ship #327 bug(agents): builtin agent source is permanently empty — builtinAgentDirs probes agents/ directories the subagents packages never ship
• Active PRs: none
• Historical PR evidence: fix(background): probe live tool registry for subagent capability #326
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: high
• Causal hypothesis: The causal hypothesis asserts builtinAgentDirs() probes agents/ directories that pi-subagents packages never ship. The provided snippets contain installer scripts, file verification lists, and skill definitions; none contain the builtinAgentDirs function or the agent discovery code that would confirm the dead paths.
• Current-main evidence:
  • No retained current-main proof; see evidence gap.
• Rollback boundary: Components: extensions/gentle-ai.ts — agent discovery (builtinAgentDirs, listDiscoverableAgents)
• Evidence gap: Missing: builtinAgentDirs implementation, pi-subagents package manifest or directory listing showing absence of agents/.
• Exact next action: Read extensions/gentle-ai.ts around line 327 to inspect builtinAgentDirs and the comment documenting the expected agents/ path. Verify pi-subagents package contents for agents/ directory.

root-343 · P1

• Members: Candidate views leak 32 MB per session into the gentle-ai store and are never swept #343 Candidate views leak 32 MB per session into the gentle-ai store and are never swept
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: high
• Causal hypothesis: The causal hypothesis asserts candidate views leak because cleanup is wired only to session_shutdown. The provided snippets contain installer scripts, version pins, and provider contract mirrors; none contain review-candidate-view.ts or the session_shutdown cleanup handler.
• Current-main evidence:
  • No retained current-main proof; see evidence gap.
• Rollback boundary: Components: lib/review-candidate-view.ts — candidate view lifecycle and cleanup
• Evidence gap: Missing: lib/review-candidate-view.ts around line 596 showing cleanupAllPendingReviewConsents registration, and any startup sweep code.
• Exact next action: Read lib/review-candidate-view.ts around line 596 to inspect the session_shutdown listener registration and verify absence of a startup cleanup sweep for candidate views.

root-347 · P2

• Members: feat(pi): add an explicit session switch to a validated linked worktree #347 feat(pi): add an explicit session switch to a validated linked worktree
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: high
• Causal hypothesis: Not a defect. The current implementation targets worktrees via workspaceRoot routing (fix(review): bind gentle_review to an explicit workspace root #179) without moving the parent session. The feature requires wiring Pi's session switch primitives behind gentle-pi validation.
• Current-main evidence:
  • issue body references #179 for workspaceRoot routing — Prior work added validated workspaceRoot routing without session move
  • issue body references #169 for cwd/worktree mismatch — The mismatch problem is acknowledged and tracked
• Rollback boundary: Components: worktree management, session lifecycle (Pi primitives: forkFrom, switchSession)
• Evidence gap: No implementation exists on main. The issue is approved but awaits the public command name decision and the actual fork+switch integration.
• Exact next action: Resolve the one open decision (public command name), then implement the fork-and-switch integration with validation, tests, and safety checks per acceptance criteria.

root-348 · P1

• Members: fix(orchestrator): enforce inline skill output contracts #348 fix(orchestrator): enforce inline skill output contracts
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: high
• Causal hypothesis: The causal hypothesis asserts inline skill output contracts are not consistently applied. The provided snippets contain migration asset manifests, SDD chain definitions, and 4R review chain definitions; none contain the orchestrator inline execution path or skill loading/contract enforcement code.
• Current-main evidence:
  • No retained current-main proof; see evidence gap.
• Rollback boundary: Components: assets/orchestrator-skills.md — inline skill execution path, Output Contract application
• Evidence gap: Missing: orchestrator code that loads skills by exact path and applies Output Contract enforcement logic.
• Exact next action: Identify and read the orchestrator module responsible for inline skill execution and output contract enforcement. Verify whether loaded skill contracts are applied consistently.

root-354 · P2

• Members: assets: non-SDD agents are installed and refreshed by install-sdd, which is what let SDD vocabulary bleed into the delegation worker #354 assets: non-SDD agents are installed and refreshed by install-sdd, which is what let SDD vocabulary bleed into the delegation worker
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: high
• Causal hypothesis: All 22 agents (10 non-SDD, 12 SDD) co-located under assets/agents/ and swept by install-sdd
• Current-main evidence:
  • scripts/verify-package-files.mjs:13-19 — verify-package-files.mjs lists all agents (gentle-ai-explore, gentle-ai-verify, gentle-ai-worker, jd-fix-agent, etc.) as required co-packaged files under assets/agents/
  • scripts/verify-package-files.mjs:25-31 — assets/agents/gentle-ai-worker.md (non-SDD) shipped alongside assets/agents/jd-fix-agent.md, jd-judge-a.md, jd-judge-b.md (SDD-related)
  • scripts/verify-package-files.mjs:27-33 — assets/agents/review-readability.md (non-SDD) packaged with all SDD agents
  • scripts/verify-package-files.mjs:31-37 — assets/agents/sdd-onboard.md through assets/agents/sdd-verify.md listed as required co-packaged files
  • scripts/verify-package-files.mjs:33-39 — assets/agents/sdd-explore.md bundled with non-SDD agents in the same required list
• Rollback boundary: Components: assets/agents/ — asset subsystem ownership, sddGlobalAssetDriftCount(), install-sdd command
• Evidence gap: Installer entry-point and drift-checker scope not directly visible in provided snippets; rely on verify-package-files co-packaging evidence
• Exact next action: Inspect sddGlobalAssetDriftCount() and /gentle:install-sdd source to confirm single-directory sweep; add directory-level ownership separation before patching

root-360 · P1

• Members: bug(installer): Windows EBUSY during staging cleanup leaves runtime absent #360 bug(installer): Windows EBUSY during staging cleanup leaves runtime absent
• Active PRs: none
• Historical PR evidence: fix(install): build trusted Windows runtime #244
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: low
• Causal hypothesis: EBUSY on Windows during safeRemoveDirectory() staging cleanup — root cause unidentified; Go toolchain or OS file handle suspected
• Current-main evidence:
  • scripts/gentle-ai-installer.mjs:143-149 — isWindowsGoSumdbSourceTarget confirms Windows x64/arm64 uses Go SumDB source build path
  • scripts/gentle-ai-installer.mjs:151-157 — Windows source install path includes staging directory and cleanup of stale staging bundles
  • scripts/gentle-ai-installer.mjs:318-324 — sealedGoEnvironment sets GOENV=off, GOWORK=off, GOTOOLCHAIN=local, GOSUMDB=sum.golang.org — package-local Go environment
  • scripts/gentle-ai-installer.mjs:320-326 — sealedGoEnvironment includes GOPATH, GOCACHE under the staging directory (implied by buildDirectory); Go may retain file handles on build artifacts
  • scripts/gentle-ai-installer.mjs:32-38 — INSTALLER_VERSION=2.5.0; GENTLE_AI_WINDOWS_SOURCE_MODULE=github.com/gentleman-programming/gentle-ai/v2 pinned
  • scripts/gentle-ai-installer.mjs:23-29 — GO_COMMAND_TIMEOUT_MS=120s extended to GO_INSTALL_TIMEOUT_MS=15min due to bug(installer): Windows source build exceeds the postinstall timeout #400 cold-build timing
• Rollback boundary: Components: scripts/gentle-ai-installer.mjs — Windows source-build staging cleanup (safeRemoveDirectory, cleanupStaleStagingBundles)
• Evidence gap: No EBUSY-specific code path visible; safeRemoveDirectory implementation not in snippets; no Windows test fixtures provided
• Exact next action: Add EBUSY retry with exponential backoff to safeRemoveDirectory (or isolate Go build artifacts to a separate subtree removed before Go cleanup); reproduce on Windows with process-hold detection (Handle.exe / Sysinternals)

root-369 · P1

• Members: fix(skill-registry): mirror Pi-resolved loaded skills #369 fix(skill-registry): mirror Pi-resolved loaded skills
• Active PRs: feat(skill-registry): index npm-package skills (node_modules/*/skills) #281
• Historical PR evidence: feat(skill-registry): index pi.skills project-declared skill dirs #285
• One-PR verdict: no
• Chain risk: true
• Confidence: high
• Causal hypothesis: Registry generation scans fixed loose skill roots; does not consume Pi-resolved loaded-skill records (before_agent_start.systemPromptOptions.skills)
• Current-main evidence:
  • scripts/verify-package-files.mjs:7-13 — requiredPaths begins with assets/orchestrator.md, assets/orchestrator-delegation.md, assets/orchestrator-memory.md — all fixed loose paths
  • scripts/verify-package-files.mjs:102-108 — requiredPaths ends with skills/work-unit-commits/SKILL.md — all entries are explicit filesystem paths under assets/ and skills/ directories
  • scripts/verify-package-files.mjs:195-201 — requiredPaths only lists explicit paths and contract hashes — no dynamic discovery, no package-root glob, no pi.skills consumption
  • scripts/verify-package-files.mjs:104-110 — contractHashes pins exact SHA-256 values for specific files — static manifest only, no runtime package resolution
• Rollback boundary: Components: skill-registry generation (extensions/skill-registry.ts + skills/skill-registry contract)
• Evidence gap: Registry generation entry-point (atl/skill-registry.md or skill-registry.ts implementation) not visible in provided snippets; rely on verify-package-files static-list evidence
• Exact next action: Locate registry generation entry-point (skill-registry.ts or atl/skill-registry.md); confirm it reads fixed loose roots instead of before_agent_start.systemPromptOptions.skills; patch to consume Pi-resolved loaded-skill records

root-370 · P1

• Members: review: a linked worktree loses the project-local .pi configuration, so review fails claiming pi is an unsupported runtime #370 review: a linked worktree loses the project-local .pi configuration, so review fails claiming pi is an unsupported runtime
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: Two unproven candidate causes (extension-set handshake vs capture-result relay failure); provided snippets contain only installer download code, verify-package-files asset lists, and contract schema hashes—none show relay/handshake logic or the runtime behavior on linked worktrees.
• Current-main evidence:
  • scripts/gentle-ai-installer.mjs:192-198 — installer.mjs download implementation confirmed in snapshot, but download code is unrelated to review relay/handshake logic
  • scripts/verify-package-files.mjs:138-144 — lib/review-relay-contract.ts listed in package-files manifest, proving the file exists; does not prove the relay handshake behavior on worktrees
  • scripts/verify-package-files.mjs:56-62 — lib/review-host-relay.ts and lib/review-integration-v2.ts listed in package-files; file existence does not confirm or refute the relay/handshake mechanism
• Rollback boundary: Components: Pi host review relay / worktree workspace configuration resolution (runtime/native-review-cli.ts, runtime/review-relay-contract.mjs)
• Evidence gap: No snippets from the actual review relay, handshake, or .pi configuration resolution paths; only contract schema hashes, fixture references, and package-file lists.
• Exact next action: Diff the relay/handshake code between the bug-triggering version and current main: lib/review-host-relay.ts, lib/review-integration-v2.ts, lib/native-review-cli.ts, and the .pi workspace configuration loader.

root-372 · P2

• Members: bug(codegraph): query requests repeatedly run init and hang Pi #372 bug(codegraph): query requests repeatedly run init and hang Pi
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: Reported trigger (init over query) narrowed to provider-relay-matrix offering an interactive clack.select under WSL /mnt with no Git sync hooks; provided snippets contain no CodeGraph code, no init/query routing, and no offerWatchFallback or clack.select references.
• Current-main evidence:
  • scripts/maintainer/provider-relay-matrix.mjs:46-52 — scripts/maintainer/provider-relay-matrix.mjs ProviderRoleVectorError defined; this is relay matrix error handling, not CodeGraph init/query code
  • scripts/gentle-ai-installer.mjs:168-174 — scripts/gentle-ai-installer.mjs download code present in snapshot; unrelated to CodeGraph hang
  • scripts/verify-package-files.mjs:13-19 — verify-package-files.mjs asset lists present; file existence does not confirm or refute the hang
• Rollback boundary: Components: extensions/codegraph-tools.ts (CodeGraph tool integration and subprocess lifecycle)
• Evidence gap: No snippets from CodeGraph integration code, init/query operation handlers, clack.select calls, or WSL /mnt path handling. Provider-relay-matrix.mjs error types are not the hang code.
• Exact next action: Diff CodeGraph integration code between the bug version and current main: search for offerWatchFallback, clack.select, init/query routing, and WSL path handling in the extension and provider-adapter layers.

root-373 · P2

• Members: fix(skill-registry): degrade truthfully when the project registry is unwritable #373 fix(skill-registry): degrade truthfully when the project registry is unwritable
• Active PRs: fix: EACCES on Android external storage during skill registry refresh (#1291) #195
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: Reported issue is about skill-registry refresh collapsing EACCES/ENOTDIR write failures into generic warnings plus fix: EACCES on Android external storage during skill registry refresh (#1291) #195 over-suppression risk (regenerated:true without durable write); provided snippets contain no skill-registry refresh code, no write-path logic, and no .gitignore update code.
• Current-main evidence:
  • skills/skill-registry/SKILL.md:13-19 — skills/skill-registry/SKILL.md lists 'Add .atl/ to .gitignore when possible' in its rules; confirms the file exists but does not show the write/refresh implementation
  • runtime/native-review-cli.mjs:2031-2037 — runtime/native-review-cli.mjs is in snapshot; it handles review errors (EACCES, ENOENT), not skill-registry file writes
  • skills/skill-registry/SKILL.md:1-5 — skills/skill-registry/SKILL.md present in snapshot; skill rules do not prove the refresh/write bug behavior
• Rollback boundary: Components: skill-registry refresh result contract and filesystem error handling (extensions/skill-registry.ts)
• Evidence gap: No snippets from the skill-registry refresh implementation, the .gitignore update path, or the registry cache write functions. The PR fix: EACCES on Android external storage during skill registry refresh (#1291) #195 implementation (the specific over-suppression risk) is not present in any snippet.
• Exact next action: Find and diff the skill-registry refresh implementation (likely in extensions/skill-registry.ts or a skill-registry provider/adapter) between the bug version and current main; also check for the .gitignore regeneration logic.

root-375 · P2

• Members: feat(startup): compact the HUD stats panel with a stable column grid #375 feat(startup): compact the HUD stats panel with a stable column grid
• Active PRs: fix(startup): reorganize HUD stats panel layout #36
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: medium
• Causal hypothesis: Stated design cause, unverified in code: wide-mode rows are centered independently with different reserved widths, so column boundaries shift and horizontal space becomes padding; the fix is one shared column grid across every row plus a narrow-mode priority order. No layout source was evidenced, so this remains the issue's hypothesis rather than a confirmed current-main mechanism.
• Current-main evidence:
  • search terms: compact, column, panel, narrow, grid, stats, startup (current_main_search, truncated) — No startup HUD layout module appeared in the evidence; all version matches are Gentle AI installer pins (e.g. scripts/gentle-ai-installer.mjs:39), so the 122-column threshold, per-row reserved widths and the 46-character path capacity named in the acceptance criteria are unevidenced.
  • skills/issue-creation/SKILL.md:113 — Repository practice is to state expected behavior and current limitation explicitly, which is what the acceptance criteria do here; no contradicting main evidence was found.
• Rollback boundary: Components: startup HUD stats panel layout (wide/narrow grid rendering)
• Evidence gap: The current-wide HUD layout constants and the project path capacity at the frozen SHA are unknown from this record, as is the narrow-mode row order, so there is nothing to assert the 122-column boundary, the 46-to-20 regression risk, or the no-wrap behavior against.
• Exact next action: Locate the wide/narrow HUD layout module on a6a73fa and record its current per-row reserved widths, column threshold, and project path capacity before reworking fix(startup): reorganize HUD stats panel layout #36.

root-376 · P2

• Members: feat(agents): execute managed tasks in linked worktrees #376 feat(agents): execute managed tasks in linked worktrees
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: Not a defect — an unresolved authority and dependency decision. The record asserts fix(review): detect a linked worktree missing its project-local .pi config #371's background-subagent work is not on main, fix(orchestrator): prevent unauthorized execution-surface escalation #379 must land first, and main currently detects only a ready|absent background-subagent capability with no versioned adapter registry or explicit runtime-authority selector; none of those prerequisite states are verified in the supplied evidence, and no first supported runtime, minimum adapter version, or selection mechanism is named.
• Current-main evidence:
  • runtime/native-review-cli.mjs:764 — The only versioned capability table visible in the evidence is per-Gentle-AI-release review columns (start, finalize, validate, status, and related), not a background-subagent runtime adapter registry.
  • contracts/review-integration/v2/fixtures/capabilities.fixture.json:85 — target_selector_variants belong to the review provider capability surface, confirming the selector concept in evidence is review-side and not task/runtime-side.
  • scripts/measure-native-authority-slimming.mjs:20 — Worktree handling present in the evidence is maintainer measurement against a WORKTREE snapshot, not task execution in a linked worktree.
  • search terms: worktree, task, select, delegated (current_main_search, truncated) — No task mode or worktree-selection implementation appeared, consistent with the reviewer's prerequisite claim but not proving it.
• Rollback boundary: Components: agents / managed task execution and worktree selection
• Evidence gap: No verified inventory of current-main state for the named prerequisites (feat(agents): add same-cwd managed execution #419 task mode, fix(orchestrator): prevent unauthorized execution-surface escalation #379 worktree safety check, fix(review): detect a linked worktree missing its project-local .pi config #371 background-subagent foundation), no evidence of how capability detection is exposed today, and no maintainer decision naming the first supported runtime, minimum adapter version and selection mechanism, cwd validation rules, or capability-specific persistence and ancestry behavior.
• Exact next action: Inventory current main for the prerequisites — confirm the background-subagent capability surface reports only ready|absent and that no task mode or worktree preflight exists — then require the issue to name the first supported runtime, minimum adapter version and selection mechanism before status:approved.

root-379 · P1

• Members: fix(orchestrator): prevent unauthorized execution-surface escalation #379 fix(orchestrator): prevent unauthorized execution-surface escalation
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: high
• Causal hypothesis: Supplied snippets are build/install scripts only; no code path for openProjectPaneIfMissing or tool_call hook classification is present.
• Current-main evidence:
  • scripts/build-runtime-modules.mjs:12-18 — Snippet I379-S001 shows build-runtime-modules.mjs source list but contains no openProjectPaneIfMissing or tool_call logic.
  • scripts/gentle-ai-installer.mjs:123-129 — Snippet I379-S009 shows installer asset form; irrelevant to execution-surface guard.
• Rollback boundary: Components: extensions/orchestrator (tool_call hook)
• Evidence gap: Missing extensions/gentle-ai.ts tool_call hook, orchestrator prompt injection, and any code that checks openProjectPaneIfMissing as an own property.
• Exact next action: Read extensions/gentle-ai.ts at snapshot 6c391ad and locate the tool_call handler; verify whether openProjectPaneIfMissing is checked before UI/tool execution and whether gentle-ai.execution-surface.external-fallback-prohibited is returned.

root-381 · P2

• Members: feat(models): expose a machine-readable routing contract #381 feat(models): expose a machine-readable routing contract
• Active PRs: none
• Historical PR evidence: refactor(models): fail closed on invalid saved routing #408
• One-PR verdict: no
• Chain risk: true
• Confidence: high
• Causal hypothesis: This is a planning artifact, not a defect. It establishes the shared root for the multi-PR chain. No implementation code is expected at the tracker level.
• Current-main evidence:
  • current_main_search term: gentle-pi.model-routing/v1 — Zero matches; the contract schema is not yet implemented on current main.
• Rollback boundary: Components: model-routing authority
• Evidence gap: N/A for a tracker. Verify the staged dependency status by checking whether refactor(models): extract saved-routing authority #398 merged (delivers refactor(models): isolate saved-routing read authority #389) and PR refactor(models): fail closed on invalid saved routing #408 (delivers refactor(models): isolate fail-closed saved-routing apply #396) is open.
• Exact next action: No direct implementation; ensure each child issue remains within its approved boundary and update this tracker's status table when a child completes.

root-382 · P2

• Members: refactor(models): share safe routing persistence authority #382 refactor(models): share safe routing persistence authority
• Active PRs: none
• Historical PR evidence: refactor(models): fail closed on invalid saved routing #408
• One-PR verdict: no
• Chain risk: true
• Confidence: high
• Causal hypothesis: The first single-PR attempt conflated too many causal boundaries into one candidate, exceeding the 400 A+D limit after corrections and introducing unsafe merge semantics. A new causal split is required.
• Current-main evidence:
  • issue comment 2026-08-21T19:24:23Z by decode2 — Design hold: PR refactor(models): centralize routing persistence authority #385 draft at 2d7f19e; correction reached 637 A+D uncommitted and unverified; status:approved removed.
  • labels: enhancement (no status:approved) — Confirms the approval was withdrawn on current main at snapshot time.
• Rollback boundary: Components: model-routing persistence authority
• Evidence gap: The new safe child boundary has not been designed or approved. Need the maintainer to publish the corrected split before dependent work resumes.
• Exact next action: Design and approve the replacement child boundary that freezes parser-compatible agent names, base-preserving semantics, file/directory durability, real production callers, and a hard stop above 400 A+D.

root-383 · P3

• Members: feat(models): add headless inspect and validate contract #383 feat(models): add headless inspect and validate contract
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: high
• Causal hypothesis: The original three-PR topology was superseded by the staged dependency chain in feat(models): expose a machine-readable routing contract #381/refactor(models): share safe routing persistence authority #382. This issue's scope will change once the foundation is delivered.
• Current-main evidence:
  • labels: enhancement, status:needs-review — Hold label confirms the issue is not actionable in its current form.
  • issue comment 2026-08-24T18:33:43Z — Do not implement from existing body; re-freeze only after complete refactor(models): share safe routing persistence authority #382 foundation.
• Rollback boundary: Components: model-routing external contract (inspect/validate)
• Evidence gap: The re-frozen scope and acceptance criteria do not exist yet.
• Exact next action: No implementation. After refactor(models): share safe routing persistence authority #382 foundation completes, rewrite the body with the corrected contract scope and re-approve.

root-384 · P3

• Members: feat(models): add apply CLI and package executable #384 feat(models): add apply CLI and package executable
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: high
• Causal hypothesis: Same topology supersession as feat(models): add headless inspect and validate contract #383. This was PR 3 of the obsolete three-PR chain and now sits at the end of the corrected chain.
• Current-main evidence:
  • labels: enhancement, status:needs-review — Hold label confirms non-actionable status.
  • issue comment 2026-08-24T18:34:02Z — Do not implement from existing body; depends on re-frozen refactor(models): share safe routing persistence authority #382 AND feat(models): add headless inspect and validate contract #383.
• Rollback boundary: Components: model-routing external contract (apply/executable/packaging)
• Evidence gap: The re-frozen scope and acceptance criteria do not exist yet.
• Exact next action: No implementation. After feat(models): add headless inspect and validate contract #383 is re-frozen, rewrite this body with corrected apply/executable/packaging scope and re-approve.

root-391 · P2

• Members: refactor(models): enforce canonical routing names and target authority #391 refactor(models): enforce canonical routing names and target authority
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: medium
• Causal hypothesis: This is a correctly scoped, approved child whose implementation is gated by upstream merges. The causal boundary (canonical names + target authority) was explicitly separated from durable-write and materialization to respect the 400 A+D limit.
• Current-main evidence:
  • current_main_search term: pnpm@11.1.1 — Zero matches in search results; environment prerequisite not yet validated on the base this issue targets.
  • issue body dependency diagram — Explicitly blocked: main → refactor(models): isolate saved-routing read authority #389 → refactor(models): isolate fail-closed saved-routing apply #396 → refactor(models): migrate routing UI and lifecycle consumers #395 → 📍 refactor(models): enforce canonical routing names and target authority #391. No comments on the issue.
• Rollback boundary: Components: model-routing canonical name authority / target resolver
• Evidence gap: Cannot verify whether refactor(models): isolate saved-routing read authority #389 (delivered through refactor(models): extract saved-routing authority #398), refactor(models): isolate fail-closed saved-routing apply #396 (PR refactor(models): fail closed on invalid saved routing #408), and refactor(models): migrate routing UI and lifecycle consumers #395 have merged to the frozen main SHA a6a73fa. Need git log or PR merge-status check at the snapshot.
• Exact next action: Confirm refactor(models): extract saved-routing authority #398 merge, PR refactor(models): fail closed on invalid saved routing #408 merge, and refactor(models): migrate routing UI and lifecycle consumers #395 delivery on the target main base; then implement the shared canonical validator, strict saved-config boundary, target resolver with alias denial, and minimal save-validation bridge per the TDD plan.

root-395 · P2

• Members: refactor(models): migrate routing UI and lifecycle consumers #395 refactor(models): migrate routing UI and lifecycle consumers
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: Specific and self-consistent per the record: consumers reimplement authority behavior instead of consuming the shared status-carrying authority; the migration is not startable because its base (refactor(models): isolate saved-routing read authority #389 then refactor(models): isolate fail-closed saved-routing apply #396 on main) is unverified as merged in this snapshot.
• Current-main evidence:
  • current_main_search term exportSavedModelConfig — No returned line contains exportSavedModelConfig; visible hits are generic collisions on main/null/inherit. truncated=true, so presence on frozen main a6a73fa is neither confirmed nor refuted.
  • current_main_search terms /gentle:models, /gentle:status, root/agents, pnpm@11.1.1 — None appear in the returned subset; the searched-and-returned files are scripts/, skills/, assets/chains and runtime/native-review-cli.mjs, so no consumer or model-routing-authority file surfaced.
  • scripts/verify-package-files.mjs:52 and :55 — extensions/gentle-ai.ts and lib/gentle-ai-binary.ts are packaged paths, confirming the surfaces that would host these consumers exist in main, without proving migration state.
• Rollback boundary: Components: model-routing saved-config consumers (export, /gentle:models, session start, /gentle:status)
• Evidence gap: No frozen-main proof that refactor(models): isolate saved-routing read authority #389 and refactor(models): isolate fail-closed saved-routing apply #396 merged, that model-routing-authority exists as an importable authority, or that tests/model-routing-authority.test.ts is present; also no proof of the current call-site shape of exportSavedModelConfig. Search was truncated and term collisions dominate the hits.
• Exact next action: On main at the snapshot SHA run a targeted check: confirm refactor(models): isolate saved-routing read authority #389 and refactor(models): isolate fail-closed saved-routing apply #396 merge state, grep for exportSavedModelConfig, model-routing-authority and the /gentle:models plus /gentle:status handlers, and list tests/model-routing-authority.test.ts; record the dependency hydration state (pnpm@11.1.1, @earendil-works/pi-tui) before treating this slice as startable.

root-405 · P2

• Members: feat(safety): harden high-risk model-initiated Pi operations #405 feat(safety): harden high-risk model-initiated Pi operations
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: Design gap, explicitly reasoned rather than observed as a defect: regex-only matching cannot establish shell/argv structure, exact package specifications, manager capability evidence or SQL lexical context, so the stated protection cannot be achieved by extending existing matchers.
• Current-main evidence:
  • current_main_search terms psql, mysql, mariadb, sqlite3, DROP TABLE, 11.16.0, 11.21.0 — No returned line contains any of these; the hit set is installer, package-verification, provider-contract and review-cli files. truncated=true, so the absence of existing package/SQL policy on main is not proven.
  • scripts/verify-package-files.mjs:52 — extensions/gentle-ai.ts is a required packaged resource, confirming the single integration surface named in the issue ships in main.
  • runtime/native-review-cli.mjs:954 — Main already carries a fail-closed precedent: an unsupported transition operation is refused rather than having an invocation synthesized for it, matching the issue's fail-closed posture.
  • scripts/maintainer/provider-relay-matrix.mjs:64 and :236 — Main already demonstrates bounded, explicitly armed subprocess execution with stream byte caps and typed blocked reasons, a reusable pattern for the proposed fixed read-only probes.
• Rollback boundary: Components: extensions/gentle-ai.ts ordered Bash tool_call policy path
• Evidence gap: No changed-line/review-workload forecast artifact, no frozen answers to the five Open Decisions (always-on vs non-weakening config, SQL in first slice vs child, work-unit split, diagnostics composition, chain strategy), and no current-main mapping of the existing dangerous-command and sensitive-path handlers that must be composed rather than duplicated.
• Exact next action: Map the existing Bash tool_call path in extensions/gentle-ai.ts to the current sensitive-path and dangerous-command handlers, publish the changed-line forecast plus proposed child issue decomposition, and put the five Open Decisions to maintainer vote before any implementation branch exists.

root-414 · P2

• Members: feat(parity): consume upstream SDD harness-attempt correction #414 feat(parity): consume upstream SDD harness-attempt correction
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: high
• Causal hypothesis: Supplied snippets contain installer, contract mirror, and agent prompts but no sdd-attempt ledger code, harness fault logs, or max-attempts=2 budget enforcement logic.
• Current-main evidence:
  • scripts/verify-package-files.mjs:12-18 — Snippet I414-S014 lists sdd-orchestrator-workflow.md as a required package file but does not contain ledger or attempt logic.
  • assets/agents/review-resilience.md:9-15 — Snippet I414-S031 is a review resilience agent prompt; irrelevant to sdd-attempt ledger behavior.
• Rollback boundary: Components: SDD runtime-attempt orchestration boundary (mirrored assets plus native sdd-attempt ledger consumption)
• Evidence gap: Missing gentle-ai sdd-attempt authority code (native-sdd-attempt-authority.test.ts or Go ledger), harness fault reproduction steps, and the canonical upstream prevention fix.
• Exact next action: Obtain and inspect the canonical upstream prevention correction (successor to #2588/#2959 versus #3152/#2968/#3563) and reproduce the max-attempts=2 exhaustion with a Pi SDD actor; confirm whether the ledger charges at begin and whether the harness fault (TypeORM migration probe, ts-node/register cwd) is the root cause.

root-419 · P2

• Members: feat(agents): add same-cwd managed execution #419 feat(agents): add same-cwd managed execution
• Active PRs: feat(runtime): add authoritative managed-task registry #441
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: high
• Causal hypothesis: Specific design cause: no single authoritative task registry exists in main, so state, cancellation and final result are handled per execution shape rather than through one linearized admission-and-terminal-commit authority; the remaining behavior is undetermined rather than misimplemented.
• Current-main evidence:
  • issue labels status:needs-review, type:feature — The frozen record itself shows the tracker is not implementation-ready after the maintainer audit and the narrowed parent-level rewrite.
  • current_main_search terms final, managed, result, start, same-cwd, execution, tasks, visible — Returned hits are installer, package-verification and native-review-cli lines matching generic words; no managed-task registry, admission barrier or final-result module surfaced. truncated=true, so absence is unproven.
  • scripts/verify-package-files.mjs:37 and :43-45 — assets/agents/sdd-tasks.md and the managed-assets migration manifests ship in main, confirming the packaged agent-asset surface that any managed mode must extend.
• Rollback boundary: Components: agent/subagent managed task identity, lifecycle and final-result authority
• Evidence gap: No frozen enable/disable cutover policy, barrier owner or acknowledgement semantics; no closed-code vocabulary, cancellation cut-off, shutdown completion bound or concurrency limit; no proof of which capability sequence step is actually delivered (roadmap records 447 linked natively while 448 and 449 reference the parent only textually, an acknowledged reconciliation gap).
• Exact next action: Freeze the four approval-blocking decisions in the normative body (closed-code vocabulary, cancellation cut-off, shutdown bound, concurrency limit) plus one enable/disable barrier policy, then reconcile the native sub-issue hierarchy for 447, 448 and 449.

root-430 · P2

• Members: feat(ui): add an OpenCode-like agent/subagent status bar #430 feat(ui): add an OpenCode-like agent/subagent status bar
• Active PRs: fix(startup): reorganize HUD stats panel layout #36
• Historical PR evidence: feat(quiet-tools): add bounded cards and Gentle AI redaction #425
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: Dependency gap, explicitly stated: no Gentle-owned authoritative projection exists yet, so any status bar built now would have to infer lifecycle and would become a competing truth authority; the UI itself is not the missing cause.
• Current-main evidence:
  • current_main_search terms gentle-pi.agent-activity/v1, ctx.ui.setStatus, agent_settled, ctx.mode, ctx.thinkingLevel, onTerminalInput — None appear in the returned subset; only generic and review-pipeline files surfaced. truncated=true, so the projection API and widget usage state on main are unproven.
  • lib/review-candidate-view.ts:1726 and :74 — Main validates and mutates the actual Pi subagent_run tool input against a closed key set (agent, agents, task, context, mode), showing the only current main-side subagent surface is review-side input handling, not task state.
  • tests/runtime-harness.mjs:2027 — Main emits and asserts kind gentle-pi.agent_model_routing in tests, evidencing routing export exists as a typed artifact but nothing resembling an activity projection.
  • openspec/changes/orchestrator-lazy-diet/proposal.md:5 — Delegation behavior today is prompt text injected from assets/orchestrator.md via getOrchestratorPrompt (extensions/gentle-ai.ts:123-133, appended at :2208), matching the issue's claim that lifecycle behavior is a prompt contract rather than enforced state.
• Rollback boundary: Components: extension UI status/widget consumer (read-only projection surface)
• Evidence gap: No proof that feat(agents): define a Gentle-owned read-only activity projection #432/feat(agents): add same-cwd managed execution #419 projection (module, event names, snapshot schema) exists on frozen main; no recorded check of Pi 0.74.0 documentation and declarations for setStatus, setWidget/custom, registerCommand/registerShortcut; the three maintainer decisions (readiness, shortcut registration, idle line wording) are unfrozen.
• Exact next action: Verify on main whether the gentle-pi:agent-activity:v1:request/changed seam shipped and record the exact Pi 0.74.0 docs and declarations for the status and widget surfaces; until both hold, keep this issue blocked and decide only the shortcut and idle-line questions.

root-432 · P2

• Members: feat(agents): define a Gentle-owned read-only activity projection #432 feat(agents): define a Gentle-owned read-only activity projection
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: Specific structural cause: main has no Gentle-owned producer/consumer boundary for task activity, so consumers would have to infer state; this issue exists to remove that inference pressure by fixing the envelope, ordering, caps and unavailable semantics.
• Current-main evidence:
  • current_main_search terms gentle-pi.agent-activity/v1, source_generation, revision, availability, total_count — The returned subset contains only INSTALLER_VERSION, version-pin and package-verification lines matching the generic word version; no projection module, schema or event constant surfaced. truncated=true, so absence on main is unproven.
  • scripts/verify-package-files.mjs:37 and :79 — Packaged-resource verification is the mechanism any new contract module must be added to, matching the issue's package-verification acceptance criterion.
  • runtime/review-relay-contract.mjs:17 — Main already ships a named cross-boundary contract constant pattern (gentle-pi.review-relay/v1), showing the convention this v1 activity contract follows.
• Rollback boundary: Components: agent-activity contract and cross-extension event seam
• Evidence gap: No frozen-main proof that the contract module, validator or event seam is absent versus already partially shipped, and no runtime producer exists to validate the envelope against; the closed producer-side decisions still owned by feat(agents): add same-cwd managed execution #419 (closed codes, concurrency limit, terminal retention) are outside this contract.
• Exact next action: Search frozen main for the agent-activity module, its event constants and any package-verification entry, and re-confirm Pi's documented ExtensionAPI.events.on/emit signature for the target Pi version; if absent, this is implementation-ready exactly as specified.

root-436 · P2

• Members: refactor(agents): isolate delegation admission behind a Gentle Pi seam #436 refactor(agents): isolate delegation admission behind a Gentle Pi seam
• Active PRs: refactor(agents): isolate delegation admission seam #439
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: medium
• Causal hypothesis: Design rationale, not a defect cause: current-main admission logic inline in the tool_call handler blocks clean execution cutover. The claimed current architecture (policy resolution, capability probe, review-actor blocking, CandidateViewRegistry injection) is asserted by the issue and not verified line-by-line in the supplied bounded search.
• Current-main evidence:
  • search term 'delegation-admission.ts' / 'lib/agent-runtime/' — Both terms were searched; no file under lib/agent-runtime/ appears in the (truncated) matches, consistent with the seam not yet existing on main.
  • scripts/verify-package-files.mjs:52 — 'extensions/gentle-ai.ts' is a required packaged path, confirming the large extension handler the issue targets is on current main.
  • package.json:15 — 'subagents' ships as a package keyword; delegation is a shipped surface, so the seam touches real production wiring.
• Rollback boundary: Components: subagent delegation / extensions/gentle-ai.ts pre-dispatch hook
• Evidence gap: Truncated bounded search cannot show extensions/gentle-ai.ts hook internals, so the equivalence baseline (exact pass-through, review block reasons, candidate injection) is unverified at frozen SHA a6a73fa; no PR diff confirms whether refactor(agents): isolate delegation admission seam #439 has landed.
• Exact next action: Read the subagent_run tool_call handler in extensions/gentle-ai.ts at a6a73fa and snapshot current pass-through/block/inject behavior as the characterization-test baseline, then implement the seam if PR refactor(agents): isolate delegation admission seam #439 is unmerged.

root-438 · P2

• Members: Adopt Biome linting (+ evaluate oxlint, fallow) before next release #438 Adopt Biome linting (+ evaluate oxlint, fallow) before next release
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: Absent lint gate (no lint script or linter dependency reported in package.json v2.2.0) permits accumulation of low-level findings; whether any of the 70 reported errors are real behavioral bugs is unconfirmed — the counts come from an older commit with default rules and no repo config.
• Current-main evidence:
  • search terms 'lint', 'package.json' — Search matched package.json script lines (e.g. package.json:46 'test:maintainer') but no lint script or Biome/oxlint/ESLint dependency appears in the (truncated) matches, consistent with the claimed absence of lint tooling.
  • issue body: 'fresh biome lint (v2.5.1, default recommended rules) on main at 6451f33c' — Concrete reproducible evidence exists, but at a commit earlier than the frozen main SHA a6a73fa, so counts do not bind to current main.
• Rollback boundary: Components: repo tooling / lint & CI quality gates
• Evidence gap: No scan against frozen SHA a6a73fa; no separation of authored sources vs generated runtime/fixtures vs rule-policy disagreements; no confirmation that any reported error is an actual product bug.
• Exact next action: Run npx @biomejs/biome@2.5.1 lint on a clean checkout of a6a73fa, split findings into authored-source vs generated vs policy-disagreement buckets, and record per-rule counts as the candidate baseline before approving any CI gate.

[barbatdev]

Immutable baseline survey shard

Snapshot strict-v1-88113a3da8ce1ac3212facb363d91144184de689f094ec1ea8c249dca3045472 · cutoff 2026-09-01T13:25:12.459361Z · main 6c391ad9e6fc2e4ca5931a7c22eaf08b97412162

root-440 · P2

• Members: feat(runtime): add authoritative managed-task registry #440 feat(runtime): add authoritative managed-task registry
• Active PRs: feat(runtime): add authoritative managed-task registry #441
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: medium
• Causal hypothesis: Design rationale, not a defect cause: the absence of a single lifecycle authority before executor construction is the risk the module removes. The issue's evidence that no registry exists is consistent with, but not proven by, the bounded search.
• Current-main evidence:
  • search term 'lib/agent-runtime/' — Term searched; no lib/agent-runtime registry module appears in the (truncated) matches, consistent with the registry not existing on main.
  • issue body: '#436 / PR #439 … intentionally creates no task or runtime state' — The preceding admission seam is documented as state-free, so no current module rivals the proposed registry's authority.
  • scripts/verify-package-files.mjs:258-285 — Three-way runtime reconciliation exists, so any new module must be registered in sources/runtime/requiredPaths — implementation-bound verification is already in place.
• Rollback boundary: Components: lib/agent-runtime task lifecycle registry
• Evidence gap: No PR feat(runtime): add authoritative managed-task registry #441 diff or search evidence confirming the module is unmerged or merged at a6a73fa; transition-graph parity with feat(agents): define a Gentle-owned read-only activity projection #432's frozen states is asserted from the issue text, not from the feat(agents): define a Gentle-owned read-only activity projection #432/feat(agents): add owned activity projection contract #435 artifact.
• Exact next action: Check merge state of PR feat(runtime): add authoritative managed-task registry #441; if unmerged, implement the registry exactly against feat(agents): define a Gentle-owned read-only activity projection #432/PR feat(agents): add owned activity projection contract #435's frozen state vocabulary and add it to the runtime-module/package verification triad.

root-442 · P2

• Members: feat(peers): add local authenticated peer messaging #442 feat(peers): add local authenticated peer messaging
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: Explicitly unproven: no baseline evidence of how sessions currently behave on the same machine and user is supplied, and no design for the trust/transport layer exists in the record, so no causal account of implementability can be given.
• Current-main evidence:
  • search terms 'gentle_list_agents', 'gentle_inspect_project', 'gentle_send_message' — All three tool names were searched and none appear in the (truncated) matches, consistent with the feature being entirely absent on current main.
  • issue label 'status:needs-review' — The issue is not approved for implementation; the record contains no design artifact or PoC.
• Rollback boundary: Components: peer discovery and local messaging subsystem
• Evidence gap: No design or proof-of-concept for the local authenticated transport (POSIX socket/lock mechanics, single-active-user Windows, bilateral trust store, delivery/queue semantics); no verified baseline of two concurrent Pi sessions on the frozen main.
• Exact next action: Produce a scoped design artifact for the trust and local-transport mechanism, or a throwaway two-session prototype on a6a73fa proving the message accept/reject/duplicate/full outcomes, before any implementation PR.

root-443 · P2

• Members: Pi cannot complete RDD receipt gate or disable RDD only for Pi #443 Pi cannot complete RDD receipt gate or disable RDD only for Pi
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: medium
• Causal hypothesis: The reported dead end (immutable_review_transport_unsupported for --agent pi) is caused by missing GENTLE_PI_REVIEW_RELAY_CONTRACT after the controller failed due to the wrong gentle-ai binary on PATH (bug(review): Pi controller misreports gentle-ai binary availability #462). Main 584c495 confirms the bare --agent pi STATUS refuses without the handshake variable by design, and gentle-pi has exported it since v2.2.0.
• Current-main evidence:
  • scripts/gentle-ai-installer.mjs:35-41 — Snippet I443-S001 shows INSTALLER_VERSION 2.5.0 and release URL derivation, confirming the runtime pin is active.
  • scripts/gentle-ai-installer.mjs:101-107 — Snippet I443-S009 shows the installer error path that mentions dev-binary.json registration, confirming the binary-on-PATH failure mode is known and handled.
  • scripts/gentle-ai-installer.mjs:183-189 — Snippet I443-S020 shows download retry logic with maxAttempts, confirming the installer's bounded attempt design.
• Rollback boundary: Components: review integration / RDD delivery gates (Pi controller path)
• Evidence gap: No direct snippet of GENTLE_PI_REVIEW_RELAY_CONTRACT export or --agent pi STATUS refusal, but the installer and contract mirror snippets confirm the relay-contract infrastructure is present and the controller path is the intended gate.
• Exact next action: Verify GENTLE_PI_REVIEW_RELAY_CONTRACT is exported from lib/review-relay-contract.ts and that --agent pi STATUS refuses without it on current main; confirm the controller path answers review.start when the variable is present.

root-444 · P2

• Members: feat(runtime): add isolated Pi AgentSession adapter #444 feat(runtime): add isolated Pi AgentSession adapter
• Active PRs: feat(runtime): add authoritative managed-task registry #441
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: medium
• Causal hypothesis: Design rationale, not a defect cause: combining construction isolation, lifecycle classification, registry transitions, cleanup, and cutover would exceed the 400-line review limit and obscure failures, and Pi 0.74's two event surfaces make lifecycle wiring a separate child. The API facts are claimed against the repository-local 0.74 SDK and were not independently verified here.
• Current-main evidence:
  • search terms 'createAgentSession' / 'ResourceLoader' / 'SessionManager.inMemory' — Terms were searched; matches show related but distinct code (e.g. lib/sdd-preflight.ts:626-630 reading getActiveTools, tests/runtime-harness.mjs tool registry) and no adapter module under lib/agent-runtime/, consistent with it being unimplemented on main.
  • issue body: '#440 / PR #441' and 'Pi 0.74' — The issue evidences its delivery position (next child after the registry) and pins the SDK version against which the contract must be checked, distinguishing it from the newer 0.84.3 global docs.
  • scripts/build-runtime-modules.mjs:34-45 — Runtime-module generation/check mode exists, so the new TS module must flow through the sources/runtime/requiredPaths reconciliation when implemented.
• Rollback boundary: Components: lib/agent-runtime Pi AgentSession adapter
• Evidence gap: No verification of the repository-local @earendil-works/pi-coding-agent 0.74 declarations (createAgentSession, inMemory managers, ResourceLoader shape, dispose semantics) was included in the supplied evidence, and no PR record shows whether the adapter has landed at a6a73fa.
• Exact next action: Diff the issue's frozen API list against the repository-local 0.74 public declarations (node_modules/@earendil-works/pi-coding-agent), then implement the adapter only if its contract holds verbatim.

root-446 · P2

• Members: feat(runtime): execute one prepared foreground Pi task through the authoritative registry #446 feat(runtime): execute one prepared foreground Pi task through the authoritative registry
• Active PRs: feat(runtime): add authoritative managed-task registry #441, feat(runtime): add isolated Pi AgentSession adapter #445
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: medium
• Causal hypothesis: Not a defect. Blocking cause is dependency ordering: PR feat(runtime): add isolated Pi AgentSession adapter #445 (preparePiAgentSession), PR feat(runtime): add authoritative managed-task registry #441/feat(runtime): atomically settle managed tasks with one closed final result #448 (registry settle + final result) and the separate classifier child (feat(runtime): classify one direct Pi foreground run outcome #449) must merge before feat(runtime): execute one prepared foreground Pi task through the authoritative registry #446 can be implemented or approved; feat(runtime): atomically settle managed tasks with one closed final result #448 additionally needed an amendment removing its narrative exclusion of failure-to-enter-running. The feat(runtime): atomically settle managed tasks with one closed final result #448 body text already contains that queued/failed-on-enter-running rule, so the stated amendment precondition appears satisfied in the record even though feat(runtime): execute one prepared foreground Pi task through the authoritative registry #446's prose still describes it as pending.
• Current-main evidence:
  • current_main_search term: preparePiAgentSession — Returned matches are installer, review-lifecycle, and skill/doc files only; no session-preparation boundary or agent-runtime executor module surfaced. Truncated result, so this is absence of evidence, not proof of absence.
  • issue record labels — status:approved + type:feature with updated_at 2026-08-27T03:54:16Z, i.e. after feat(runtime): classify one direct Pi foreground run outcome #449 was created (03:49:26Z) and after feat(runtime): atomically settle managed tasks with one closed final result #448 was updated (03:22:28Z).
  • issue #446 body: Dependencies, delivery — Explicitly declares parent feat(agents): add same-cwd managed execution #419 and required merged foundations PR feat(runtime): add authoritative managed-task registry #441, PR feat(runtime): add isolated Pi AgentSession adapter #445, the future feat(runtime): atomically settle managed tasks with one closed final result #448 implementation PR, plus the separately approved classifier child; states feat(runtime): execute one prepared foreground Pi task through the authoritative registry #446 'remains needs-review' until the feat(runtime): atomically settle managed tasks with one closed final result #448 amendment and classifier are approved.
  • issue #448 body: Transition, atomicity, and idempotence — 'After create succeeds, a caller may settle directly from queued only as failed/failed or failed/cleanup-failed, including recovery after a failed attempt to enter running' — the exact amendment feat(runtime): execute one prepared foreground Pi task through the authoritative registry #446 said feat(runtime): atomically settle managed tasks with one closed final result #448 still needed.
• Rollback boundary: Components: lib/agent-runtime — foreground Pi executor
• Evidence gap: No frozen-main file/symbol evidence that preparePiAgentSession (feat(runtime): add isolated Pi AgentSession adapter #445), registry.settle (feat(runtime): atomically settle managed tasks with one closed final result #448), or Pi 0.74 agent_end observation exists at a6a73fa; search was truncated. Also unresolved record drift: body prose says feat(runtime): execute one prepared foreground Pi task through the authoritative registry #446 is needs-review while the label says approved, and the 'new separately approved bounded child' for classification is described as not yet existing even though feat(runtime): classify one direct Pi foreground run outcome #449 covers it.
• Exact next action: Reconcile feat(runtime): execute one prepared foreground Pi task through the authoritative registry #446's body prose with its status:approved label and with feat(runtime): atomically settle managed tasks with one closed final result #448/feat(runtime): classify one direct Pi foreground run outcome #449 (point the classifier child at feat(runtime): classify one direct Pi foreground run outcome #449 and drop the already-applied feat(runtime): atomically settle managed tasks with one closed final result #448 amendment precondition), then verify on frozen main that PR feat(runtime): add isolated Pi AgentSession adapter #445 and the feat(runtime): atomically settle managed tasks with one closed final result #448 implementation have merged before any feat(runtime): execute one prepared foreground Pi task through the authoritative registry #446 implementation is authorized.

root-447 · P2

• Members: feat(agents): add an optional managed mode for delegated work #447 feat(agents): add an optional managed mode for delegated work
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: medium
• Causal hypothesis: Not a defect. Deliverability is gated on fix(orchestrator): prevent unauthorized execution-surface escalation #379 (unsupported execution must not start alternate work) and on the feat(runtime): execute one prepared foreground Pi task through the authoritative registry #446/feat(runtime): atomically settle managed tasks with one closed final result #448/feat(runtime): classify one direct Pi foreground run outcome #449 managed runtime chain; the slice is deliberately non-functional (selection point only), so it can be reviewed independently but cannot demonstrate end-to-end value until the chain exists.
• Current-main evidence:
  • current_main_search terms: managed, delegated, selection point, execution choice — All matches are installer confinement logic, provider-relay matrix, skill docs, and package-file lists; no delegated-work mode-selection implementation or eligibility test surfaced (truncated).
  • issue record labels — status:needs-review + type:feature; no assignees; created and last updated 2026-08-27T00:50:14Z, i.e. never revised after submission.
• Rollback boundary: Components: agents / delegation routing — execution-mode selection boundary
• Evidence gap: No frozen-main evidence for the fix(orchestrator): prevent unauthorized execution-surface escalation #379 fail-closed precondition, for what makes same-cwd delegated work 'eligible', or for where the single selection point lives; the acceptance criteria name a compatibility test but no test file or module is identified.
• Exact next action: Ask for feat(agents): add an optional managed mode for delegated work #447 to be amended with the concrete interface of the selection point (where the choice is recorded, which call site reads it) and the exact fix(orchestrator): prevent unauthorized execution-surface escalation #379 fail-closed error path, or resolve it as subsumed by the feat(runtime): execute one prepared foreground Pi task through the authoritative registry #446/feat(runtime): atomically settle managed tasks with one closed final result #448 chain if the managed-mode cutover is now owned elsewhere.

root-448 · P2

• Members: feat(runtime): atomically settle managed tasks with one closed final result #448 feat(runtime): atomically settle managed tasks with one closed final result
• Active PRs: feat(runtime): add authoritative managed-task registry #441
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: high
• Causal hypothesis: Not a defect but a closed authority gap in the merged registry contract: feat(runtime): add authoritative managed-task registry #440's transition(id, terminalState) can commit a terminal state with no result because result storage was intentionally out of scope there and in feat(runtime): execute one prepared foreground Pi task through the authoritative registry #446. The fix is one indivisible mutation inside the existing registry plus a terminal-transition guard; the risk is that the per-row source-state matrix, identity checks (result === task.result === snapshot record === listener snapshot), and observable exact-record validation are the hard parts, not the API shape.
• Current-main evidence:
  • current_main_search terms: createAgentTaskRegistry, transition(id, terminalState, atomic settle, task result storage, registry result — Only scripts/verify-package-files.mjs contract-hash entries and review-CLI/runtime result code matched; no agent-runtime registry module or settle/transition surface surfaced (truncated).
  • issue #448 body: Evidence and decision — Anchors the gap to a specific commit, feat(runtime): add authoritative managed-task registry #440/PR feat(runtime): add authoritative managed-task registry #441 at fd2d7f5, and states its bare terminal transition 'can create result-less terminal records; that is the sole authority gap this child closes.'
  • issue #448 body: Transition, atomicity, and idempotence — Contains the direct-queued settlement rule for a failed attempt to enter running plus queued rejection of failed/subscription-failed and failed/prompt-failed — the content feat(runtime): execute one prepared foreground Pi task through the authoritative registry #446 demanded as a feat(runtime): atomically settle managed tasks with one closed final result #448 amendment; feat(runtime): atomically settle managed tasks with one closed final result #448 updated_at 2026-08-27T03:22:28Z postdates feat(runtime): execute one prepared foreground Pi task through the authoritative registry #446's creation.
  • scripts/verify-package-files.mjs:138-178 (search output) — Frozen main pins contracts/review-integration v1/v2 fixture and schema digests including consent schemas, which is consistent with a repository norm of hash-pinned closed contract surfaces that this registry union (invalid-result, conflicting-result) must also respect.
• Rollback boundary: Components: lib/agent-runtime — managed task registry authority
• Evidence gap: No frozen-main read of the registry module itself: cannot confirm PR feat(runtime): add authoritative managed-task registry #441 is merged at a6a73fa, cannot confirm the current failure-code union lacks invalid-result/conflicting-result, and cannot confirm whether fd2d7f5 is an ancestor of the snapshot. The 260–350 line forecast is also unvalidated against the real registry and test file sizes.
• Exact next action: Read lib/agent-runtime/agent-task-registry.ts (or its merged equivalent) and its focused test at a6a73fa, confirm whether settle and the terminal-transition guard already exist there, and record the current closed failure-code union so the delta and line forecast are measured rather than estimated.

root-449 · P2

• Members: feat(runtime): classify one direct Pi foreground run outcome #449 feat(runtime): classify one direct Pi foreground run outcome
• Active PRs: feat(runtime): add authoritative managed-task registry #441
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: medium
• Causal hypothesis: Not a defect. Implementation is gated on feat(runtime): atomically settle managed tasks with one closed final result #448 merging first because the helper type-only imports AgentTaskFinalResultInput; until then the type specifier cannot resolve and the narrowed three-pair return union cannot be expressed. The subtle correctness driver is Pi 0.74's loop shape (assistant messages appended, then tool results, with early shouldStopAfterTurn able to emit a trailing-tool-result agent_end), which is why latest-assistant selection must coexist with whole-array validation.
• Current-main evidence:
  • current_main_search terms: lib/agent-runtime/pi-foreground-run-outcome.ts, tests/pi-foreground-run-outcome.test.ts — Neither target path appears in the returned (truncated) matches, so the helper and its focused suite are not evidenced on frozen main.
  • current_main_search terms: agent_end.messages, stopReason, state/code pair literals — Matches are installer error handling and package-file hashes only; no Pi agent-event or stopReason consumption site surfaced in the package sources.
  • issue #449 body: Pi 0.74 evidence — Cites repository-local public package declarations: @earendil-works/pi-agent-core AgentEvent agent_end carrying messages: AgentMessage[], @earendil-works/pi-ai StopReason as exactly stop|length|toolUse|error|aborted, and Agent.subscribe() awaiting listener promises so agent_end participates in run settlement.
• Rollback boundary: Components: lib/agent-runtime/pi-foreground-run-outcome.ts — pure classification
• Evidence gap: The pinned Pi version in frozen main (package.json/lockfile) is not in the supplied evidence, so 'Pi 0.74' is asserted by the issue rather than verified against the snapshot; likewise the feat(runtime): atomically settle managed tasks with one closed final result #448 exported type name and path, and the exact module specifier, are unverified. The static no-forbidden-dependency checks in the required tests are also not yet exercised.
• Exact next action: Verify the pinned Pi version and the feat(runtime): atomically settle managed tasks with one closed final result #448 merged export path/symbol at a6a73fa, then confirm whether feat(runtime): classify one direct Pi foreground run outcome #449's type-only import specifier resolves; if not, mark the issue as blocked-on-merge rather than implementation-ready.

root-455 · P1

• Members: fix(review): preserve consent bindings across active Pi sessions #455 fix(review): preserve consent bindings across active Pi sessions
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: high
• Causal hypothesis: All 20 snippets are asset manifest/migration JSON files (S001–S014), installer error-handlers (S006–S008), build config (S015–S017), and package-verify lists (S018–S060). None contain consent binding, answerConsentBinding, or session-local map code.
• Current-main evidence:
  • No retained current-main proof; see evidence gap.
• Rollback boundary: Components: review integration consent binding — active session registry (extensions/lib review-integration)
• Evidence gap: No snippets from review-integration v2 consent path, session binding maps, or answerConsentBinding.
• Exact next action: Read lib/review-integration-v2.ts and runtime/review-integration-v2.mjs for consent binding storage and answerConsentBinding.

root-460 · P2

• Members: bug(guard): confirmation prompt hides the guarded action #460 bug(guard): confirmation prompt hides the guarded action
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: high
• Causal hypothesis: All 59 snippets are installer scripts (S001–S021), chain definitions (S022–S037), skills templates (S038–S043), provider-relay-matrix (S044–S053), comment-writer skill (S054–S056), and review-resilience (S057–S060). Zero snippets contain guard dialog, confirmation title, or preview truncation code.
• Current-main evidence:
  • No retained current-main proof; see evidence gap.
• Rollback boundary: Components: guarded Bash command confirmation UI (TUI dialog title + bounded preview)
• Evidence gap: No snippets from guard dialog UI, confirmation title builder, or preview truncation logic.
• Exact next action: Read guard/confirmation dialog source and classification-to-preview code paths.

root-463 · P2

• Members: bug(orchestrator): command presence forces subagent verification for bounded checks #463 bug(orchestrator): command presence forces subagent verification for bounded checks
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: medium
• Causal hypothesis: By design, not a code fault: assets/orchestrator.md and assets/orchestrator-delegation.md define command-running non-SDD verification as gentle-ai-verify work and treat command presence as an unconditional trigger (introduced via feat(agents): add generic non-SDD exploration and verification roles #87/PR feat(agents): add non-SDD exploration and verification roles #89), and tests/package-manifest.test.ts plus tests/sdd-agent-tools.test.ts pin that text. Runtime code does not independently classify small tasks; the parent only follows the rendered prompt.
• Current-main evidence:
  • issue triage comment https://github.com/Gentleman-Programming/gentle-pi/issues/463#issuecomment-5478926831 (main 584c495c, 2026-08-31) — Maintainer read assets/orchestrator.md and assets/orchestrator-delegation.md on main and confirmed every command-executing verification still routes to gentle-ai-verify, with the routing pinned by tests/package-manifest.test.ts and tests/sdd-agent-tools.test.ts; also confirmed gentle-ai's own synced orchestrators list tests/builds/installs as 'allowed as a bounded action'.
  • issue body 'Ownership evidence' section — Names the two prompt files as the policy source and states the extension renders this policy into the parent prompt with no independent runtime classification.
  • current_main_search terms gentle-ai-verify / assets/orchestrator.md / inline — Bounded search returned only unrelated scripts/gentle-ai-installer.mjs matches (truncated); it did NOT surface the verification-routing text, so it adds no confirmation. The confirmatory evidence is the maintainer triage read, not the search output.
• Rollback boundary: Components: orchestrator delegation routing prompts (assets/orchestrator.md, assets/orchestrator-delegation.md) and routing contract tests
• Evidence gap: Triage was performed at main 584c495; the frozen snapshot is a6a73fa, so the exact current wording of the command-presence trigger has not been read at this SHA. There is also no measured cost/latency baseline for the claimed overhead, and no artifact showing gentle-pi's synced orchestrator text diverging from gentle-ai's 'bounded action' wording.
• Exact next action: Read the verification-routing sections of assets/orchestrator.md and assets/orchestrator-delegation.md at frozen main a6a73fa, diff them against the triage'd 584c495 text, and record the maintainer's design decision on the exact inline fast-path predicate before any prompt or test edit.

root-464 · P1

• Members: bug(release): latest v2.2.0 still strands corrected lineages before last-event migration #464 bug(release): latest v2.2.0 still strands corrected lineages before last-event migration
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: high
• Causal hypothesis: All 60 snippets are release/install scripts (S001–S026), contract mirror (S027–S040), packed-test (S041–S047), trademark (S048–S054), build config (S055–S056), and architecture docs (S057–S060). None contain the Go binary review_facade.go code or correction preflight validation that would confirm or refute the reported rejection.
• Current-main evidence:
  • No retained current-main proof; see evidence gap.
• Rollback boundary: Components: release/version boundary: published npm latest and GitHub release vs main; review correction closure lifecycle (retired FINALIZE sequence)
• Evidence gap: No snippets from package-local Go binary (review_facade.go or equivalent), correction preflight, or FINALIZE path.
• Exact next action: Inspect package-local .gentle-ai Go binary diff between shipped v2.4.0 and main for review_facade.go correction preflight and FINALIZE.

root-466 · P1

• Members: Pi compact facade cannot relay maintainer authorization binding for review abandon #466 Pi compact facade cannot relay maintainer authorization binding for review abandon
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: high
• Causal hypothesis: The provided snippets do not contain the facade's abandon input schema or the argument serialization path. Snippet I466-S011 shows a test fixture with a correction_plan input, not abandon. Snippets I466-S001-S002 are package verification scripts. No code proving the field is omitted or mangled is present.
• Current-main evidence:
  • tests/runtime-harness.mjs:250-256 — Test fixture shows a correction_plan input schema, not abandon, so it does not prove the abandon schema is correct or incorrect.
• Rollback boundary: Components: compact review facade transport boundary: gentle_review MCP input schema to native CLI argv for the ABANDON operation
• Evidence gap: Missing: lib/review-host-relay.ts abandon handler, lib/native-review-cli.ts abandon argument builder, lib/review-integration-v2.ts abandon schema definition.
• Exact next action: Read lib/review-host-relay.ts and lib/native-review-cli.ts to verify if abandon() validates and forwards maintainerAuthorization.

root-469 · P2

• Members: bug(orchestrator): preserve authorization across resumable subagent interactions #469 bug(orchestrator): preserve authorization across resumable subagent interactions
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: high
• Causal hypothesis: The provided snippets do not contain the orchestrator's interaction relay logic, subagent lifecycle management, or authorization binding code. Snippets I469-S001-S021 are installer scripts. Snippets I469-S022-S034 are package verification. Snippets I469-S035-S060 are provider relay matrix and contract mirror scripts. None address background task terminal state or authorization identity binding.
• Current-main evidence:
  • scripts/maintainer/provider-relay-matrix.mjs:13-19 — Provider relay matrix imports review-host-relay but does not implement orchestrator interaction relay or authorization scoping.
• Rollback boundary: Components: orchestrator interaction/authorization continuity: lossless blocking-prompt relay contract plus background subagent task lifecycle and authorization identity
• Evidence gap: Missing: orchestrator interaction relay code, subagent lifecycle state machine, authorization binding resolution logic.
• Exact next action: Read the orchestrator's subagent interaction relay and authorization binding code to verify background task terminal handling and candidate-identity binding.

root-471 · P2

• Members: feat(sdd): enable source-backed research through pi-web-access #471 feat(sdd): enable source-backed research through pi-web-access
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: Not a defect: an intentional fail-closed grant posture that is now potentially obsolete. The stated precondition is upstream — the exact Pi evidence grants must be defined and approved in Gentle AI issue #3846 — before Pi can replace the unconditional empty-grant/blocking prose with exact-tool admission. The enabling technical fact (exact-name extension tools propagate into lean child sessions) is reporter-provided from a scratch-project probe, not repo evidence.
• Current-main evidence:
  • scripts/verify-package-files.mjs:33 — Frozen main packages 'assets/agents/sdd-research.md' as a required resource, confirming the agent definition whose grants are in question exists and is release-gated.
  • current_main_search terms 'sdd-research', 'documentation=[]', 'open-web=[]', 'web_search', 'source_check', 'fetch_content', 'get_search_content' — The bounded search matched only verify-package-files packaging and unrelated provider-relay/status code; it did NOT surface the agent file's grant declarations or any of the four tool names. The current grant state on frozen main is therefore unverified.
  • assets/sdd-orchestrator-workflow.md:67 and :185 — Main's workflow text governs selector/deferral behavior and no-scope-drift validation for planning phases; nothing evidences a research-phase web grant.
• Rollback boundary: Components: sdd-research executor capability contract (assets/agents/sdd-research.md tool allowlist) plus packaging/runtime/allowlist tests
• Evidence gap: assets/agents/sdd-research.md has not been read at frozen main a6a73fa, so the claimed documentation=[]; open-web=[] and unconditional blocked persistence are unconfirmed. Also unconfirmed: the probe results (which ran in a scratch project, not this repo), the exact tool set and grant semantics upstream #3846 will approve, and whether a lean child-session integration probe can be made deterministic in CI.
• Exact next action: Read assets/agents/sdd-research.md at frozen main a6a73fa to confirm the empty-grant declaration and the unconditional blocked path, and record Gentle AI #3846's current decision state before drafting any Pi allowlist or test change.

root-476 · P2

• Members: feat(agents): allow constrained CodeGraph context in generic worker and verifier #476 feat(agents): allow constrained CodeGraph context in generic worker and verifier
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: medium
• Causal hypothesis: Design asymmetry, not a fault: CodeGraph support was added narrowly to the explorer in feat(agents): allow constrained CodeGraph exploration #106/PR feat(review): harden runtime contracts and CodeGraph exploration #108 after the generic roles were introduced in feat(agents): add generic non-SDD exploration and verification roles #87/PR feat(agents): add non-SDD exploration and verification roles #89, and the worker/verify allowlists were never widened. The proposed change is additive to read-side capability while leaving write, shell, delegation and delivery authority untouched; the open question is whether normal index freshness (project-root MCP watcher, incremental sync, connect-time catch-up) is sufficient to admit the tool without index-lifecycle machinery.
• Current-main evidence:
  • scripts/verify-package-files.mjs:16-18 — Frozen main packages 'assets/agents/gentle-ai-explore.md', 'assets/agents/gentle-ai-verify.md' and 'assets/agents/gentle-ai-worker.md' as required resources, confirming the three role definitions at issue exist and are release-verified.
  • current_main_search terms 'codegraph', 'gentle-ai-explore', 'grep', 'find', 'constrained' — The bounded search returned no CodeGraph allowlist text from any agent file; it matched installer and verify-package-files lines only. Neither the explore-only grant nor the worker/verify absence of codegraph is confirmed on frozen main.
  • assets/orchestrator-delegation.md:158 — Main already establishes that the bounded writer refuses to write outside parent-provided allowed edit surfaces and stops with status: interaction_required when they are missing, which is the pre-existing invariant the issue asks to preserve while adding read capability.
  • assets/agents/gentle-ai-worker.md:72 — Worker prompt treats the allowed edit surface as a derived, parent/human-approved candidate set, evidencing that read-capability expansion need not touch write surfaces.
• Rollback boundary: Components: generic agent role tool allowlists (assets/agents/gentle-ai-worker.md, assets/agents/gentle-ai-verify.md) with role-boundary prompts and packaging/runtime-harness tests
• Evidence gap: The actual tool allowlists in assets/agents/gentle-ai-worker.md and assets/agents/gentle-ai-verify.md were not read at frozen main a6a73fa, so 'codegraph granted to explore only' is unconfirmed. No evidence in this record demonstrates lean child-session behavior of the constrained codegraph tool under worker/verify prompts, its cwd confinement under those roles, or whether stale/pending index reporting is observable enough for the proposed direct-read fallback.
• Exact next action: Inspect the tool allowlists in assets/agents/gentle-ai-worker.md and assets/agents/gentle-ai-verify.md at frozen main a6a73fa to confirm codegraph is absent while present in gentle-ai-explore.md, then run one read-only lean child-session probe of the constrained tool under each generic role before any grant or test change.

root-479 · P2

• Members: feat(orchestrator): make parallel subagent execution policy configurable #479 feat(orchestrator): make parallel subagent execution policy configurable
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: medium
• Causal hypothesis: The parent orchestration contract encodes a constant background ceiling (reported as two) and has no validated config field or deciding-source in status; specific defect: exact next action is to locate the embedded ceiling before designing. Partially unproven: the bounded search for 'max_concurrency', 'default_mode: background', 'subagent_continue', 'total_timeout', 'stall_timeout' returned no orchestrator-policy hit at frozen main.
• Current-main evidence:
  • skills/gentle-ai/SKILL.md:19-20 — Orchestration is explicitly parent-owned with delegation triggers, but no concurrency ceiling or config field appears in the matched policy text.
  • search term 'max_concurrency' / 'default_mode: background' — Bounded current-main search matches contain no orchestrator policy source defining a configurable or numeric background ceiling, so the claimed hard-coded 'two' is not located in shown evidence.
  • assets/agents/gentle-ai-worker.md:16 — Packaged worker contract reinforces parent-only delegation (children must not delegate), matching the issue's non-goal 'Allowing child agents to delegate'.
• Rollback boundary: Components: orchestrator policy / packaged skills + subagent run-mode contract
• Evidence gap: The exact current-main location of the fixed two-task parent ceiling (prompt asset vs code) is not surfaced by the truncated search, so the premise is asserted, not shown; no evidence that a validation or status-source for such a field exists or is absent by design.
• Exact next action: Grep the frozen main orchestrator prompt/contract assets for the embedded background ceiling text ('two', 'max_concurrency', 'background') to confirm the current default and its deciding source, then scope the policy-schema + status + contract-wording slice below 400 lines as proposed.

root-497 · P2

• Members: Notify Moshi when gentle-pi is waiting for user input #497 Notify Moshi when gentle-pi is waiting for user input
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: medium
• Causal hypothesis: Pi has no universal question lifecycle (interactions are extension-defined), so Moshi's generated Pi hook cannot know blocking semantics; gentle-pi emits herdr:blocked but performs no side-channel notification, leaving attention delivery to the terminal.
• Current-main evidence:
  • search terms 'herdr:blocked', 'moshi-hook', 'Moshi' — Bounded current-main search returned installer/path matches only; no matched line in the supplied evidence shows an existing Moshi adapter or moshi-hook invocation, consistent with an unimplemented integration.
• Rollback boundary: Components: interactive blocker lifecycle / notification adapter
• Evidence gap: The current blocker lifecycle implementation (where herdr:blocked is emitted and how inactive-to-active transitions are deduplicated) is not present in the shown matches, so the exact adapter insertion point and existing effective-blocker rules are unverified.
• Exact next action: Read the gentle-pi blocker-state owner in current main (the herdr:blocked emitter and its transition dedup) and confirm a single replaceable adapter seam with a fake-process contract test that passes with no moshi-hook on PATH.

root-498 · P1

• Members: Autonomous Pi subagent panes remain open after completion #498 Autonomous Pi subagent panes remain open after completion
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: high
• Causal hypothesis: The provided snippets do not contain the pi-herdr-subagents pane lifecycle code or the auto-exit frontmatter handling logic. Snippets I498-S001-S028 are installer scripts. Snippets I498-S029-S060 are native review CLI and contract code. None address pane closing or auto-exit.
• Current-main evidence:
  • scripts/gentle-ai-installer.mjs:35-41 — Installer script does not contain pi-herdr-subagents pane lifecycle or auto-exit logic.
• Rollback boundary: Components: packaged agent definitions (assets/agents/) consumed by pi-herdr-subagents
• Evidence gap: Missing: pi-herdr-subagents pane closing logic, auto-exit frontmatter parser, agent definition schema.
• Exact next action: Read pi-herdr-subagents pane lifecycle code and agent frontmatter schema to verify auto-exit handling.

root-503 · P2

• Members: bug(review): correction-required capture disappears from compact-v2 status #503 bug(review): correction-required capture disappears from compact-v2 status
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: high
• Causal hypothesis: Causal hypothesis (compact-v2 store fails to reconstruct lineage or status mapping discards valid result) cannot be verified; supplied snippets contain installer constants, migration manifests, verify-package-files, and relay matrix code but no compact-v2 store or review status mapping logic.
• Current-main evidence:
  • scripts/gentle-ai-installer.mjs:35-41 — INSTALLER_VERSION constant
  • scripts/gentle-ai-installer.mjs:64-70 — GENTLE_AI_PENDING_DIGEST sentinel
• Rollback boundary: Components: compact-v2 native review authority store (Git common directory) surfaced through runtime/native-review-cli.mjs
• Evidence gap: Missing: compact-v2 store implementation, review.status mapping logic, correction_required state handling, and lineage enumeration code.
• Exact next action: Read the compact-v2 store and review status mapping source files on main SHA 6c391ad to verify correction_required -> unrelated transition.

root-509 · P2

• Members: gentle_review_capture rejects ordinary collect bindings — forecast ack resubmit and even first no-ack capture fail (capture-binding-rejected) #509 gentle_review_capture rejects ordinary collect bindings — forecast ack resubmit and even first no-ack capture fail (capture-binding-rejected)
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: medium
• Causal hypothesis: Issue reports capture-binding-rejected due to wrapper schema decoder rejecting newer provider contract revisions (start/v4, status v5, consent v2). Prior disposition states fixed by forward-acceptance patches fix(review): accept status v5 receipt #487, fix(review): validate consent choice agent bindings #496, fix(review): name stale consent bindings, project renames like Go, and surface admission refusals #527.
• Current-main evidence:
  • skills/_shared/review-ledger-contract.md:1-6 — review-ledger-contract describes native Go owns validation, canonicalization, and legal lifecycle transitions
  • runtime/native-review-cli.mjs:74-80 — native-review-cli.mjs defines error codes including schema-incompatible
  • runtime/native-review-cli.mjs:75-81 — schema-incompatible error code present in native-review-cli
• Rollback boundary: Components: runtime/review-integration-v2
• Evidence gap: No direct snippet showing the forward-acceptance patch code itself, but prior disposition is explicit and snippets are consistent with the fix being in place.
• Exact next action: No action required; issue is resolved per prior disposition.

root-510 · P2

• Members: bug(models): "Set all agents" persists template placeholder names as routing entries #510 bug(models): "Set all agents" persists template placeholder names as routing entries
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: high
• Causal hypothesis: Hypothesis cites validation asymmetry between discovery/parse and applyModelConfig with specific line numbers (1362-1375, 1488-1503, 1280, 1723-1726), but none of these files or line ranges appear in the provided snippets.
• Current-main evidence:
  • No retained current-main proof; see evidence gap.
• Rollback boundary: Components: extensions/gentle-ai.ts — /gentle:models agent discovery + ~/.pi/gentle-ai/models.json write/apply path
• Evidence gap: Missing: extensions/gentle-ai.ts (especially lines 1280, 1362-1375, 1488-1503, 1723-1726), models.json content, parseAgentName and applyModelConfig implementations.
• Exact next action: Fetch extensions/gentle-ai.ts at snapshot strict-v1-88113a3da8ce1ac3212facb363d91144184de689f094ec1ea8c249dca3045472 and verify whether parseAgentName accepts arbitrary frontmatter name values and whether applyModelConfig silently skips keys failing /^[A-Za-z0-9._:@/+%-]+$/.

root-511 · P2

• Members: fix(orchestrator): enforce parent hot memory persistence on inline direct tasks #511 fix(orchestrator): enforce parent hot memory persistence on inline direct tasks
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: high
• Causal hypothesis: Prompt-spec omission, not a code defect: the Memory Contract sentence assigns persistence to subagents only, and the Inline Direct route description carries no persistence step, so the parent model treats small tasks as purely conversational.
• Current-main evidence:
  • assets/orchestrator.md:76-78 — § Memory Contract reads 'the parent selects context and subagents save significant discoveries before returning' — subagent saving only; no parent/inline obligation on main.
  • assets/orchestrator.md:42 — Route 1 'Inline Direct' is defined as small/mechanical parent work with 'No SDD ceremony' and lists no persistence action.
  • scripts/verify-package-files.mjs:11-13 — assets/orchestrator.md, orchestrator-delegation.md, orchestrator-memory.md and orchestrator-skills.md are required package paths, so the proposed second file exists and is shipped.
  • openspec/changes/orchestrator-lazy-diet/review-ledger.md:28 — Evidence that the orchestrator prompt is byte-budgeted and that at realistic install-path lengths the substituted prompt measured 10,174 B against a 10,240 B budget — which contradicts this issue's stated 'canonical 8,192-byte budget' and must be reconciled before wording is added.
• Rollback boundary: Components: assets/orchestrator.md § Memory Contract + assets/orchestrator-memory.md (always-on parent prompt, byte-budgeted)
• Evidence gap: No supplied evidence of the actual current budget constant in tests/orchestrator-budget.test.ts, so the 8,192 vs 10,240 byte figures conflict; also no measured headroom for the added text.
• Exact next action: Read tests/orchestrator-budget.test.ts to fix the real byte budget and current headroom, then draft the parent-inline mem_save sentence for assets/orchestrator.md § Memory Contract plus a '### Direct Inline & General Tasks' rule in assets/orchestrator-memory.md and re-measure the substituted prompt at a realistic install path.

root-525 · P2

• Members: bug(quiet-tools): filtering quiet-tools disables rose lifecycle renderer #525 bug(quiet-tools): filtering quiet-tools disables rose lifecycle renderer
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: high
• Causal hypothesis: Hypothesis claims extensions/quiet-tools.ts default export registers all TOOL_CREATORS including bash, causing filter-side drops. The fix direction is to export a dedicated Bash registration. Neither file is in the provided snippets.
• Current-main evidence:
  • No retained current-main proof; see evidence gap.
• Rollback boundary: Components: extensions/quiet-tools.ts — tool-owner registration surface
• Evidence gap: Missing: extensions/quiet-tools.ts, extensions/gentle-ai-bash.ts, rose lifecycle renderer code, TOOL_CREATORS registration.
• Exact next action: Fetch extensions/quiet-tools.ts at the target snapshot and verify whether the default export registers bash TOOL_CREATORS and whether a package filter on this file would drop the Bash rose renderer.

root-531 · P2

• Members: Windows: CodeGraph Pi tool reports ENOENT because execFile("codegraph") does not resolve npm/Scoop shims #531 Windows: CodeGraph Pi tool reports ENOENT because execFile("codegraph") does not resolve npm/Scoop shims
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: high
• Causal hypothesis: Hypothesis claims execFile without shell on Windows does not resolve .bat/.cmd shims for codegraph. The causal mechanism is documented Node behavior but the specific file (extensions/codegraph-tools.ts) is not in the snippets.
• Current-main evidence:
  • No retained current-main proof; see evidence gap.
• Rollback boundary: Components: extensions/codegraph-tools.ts — CodeGraph process invocation on win32
• Evidence gap: Missing: extensions/codegraph-tools.ts (or equivalent codegraph spawn/execFile code), any Windows-specific spawn configuration for codegraph.
• Exact next action: Fetch extensions/codegraph-tools.ts at the target snapshot and verify whether codegraph is spawned via execFile/spawn without shell, confirming that bare command names on win32 would fail to resolve .cmd/.bat shims.

root-533 · P2

• Members: feat(review): launch 4R lens capture subprocesses concurrently in the host relay #533 feat(review): launch 4R lens capture subprocesses concurrently in the host relay
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: reproduce-first
• Chain risk: false
• Confidence: low
• Causal hypothesis: Not a defect. Premise is that the gentle-pi relay satisfies one collect input at a time while native Go already permits concurrency (shared CapturePhaseRevision, lock-owned idempotent CAS merge). Neither the sequential relay behavior nor the Go-side concurrency claim is verifiable from the supplied evidence.
• Current-main evidence:
  • scripts/verify-package-files.mjs:58-65 — Confirms lib/review-host-relay.ts, lib/review-integration-v2.ts and lib/review-relay-contract.ts plus their generated runtime/ mirrors exist as required package paths — the change surface is real.
  • scripts/verify-package-files.mjs:94 — skills/_shared/review-ledger-contract.md is a required package path, so the requested prose-parity target exists on main.
  • current_main_search terms: review.capture-result, collect.inputs, CapturePhaseRevision, --expected-revision — Truncated search returned no relay source lines for those identifiers, so the sequential-capture premise is unconfirmed at a6a73fa.
• Rollback boundary: Components: lib/review-host-relay.ts / lib/review-relay-contract.ts — host relay capture transaction runner; skills/_shared/review-ledger-contract.md prose parity
• Evidence gap: Read the relay capture loop to confirm serial execution and that no shared revision or binding is invalidated by overlap; confirm the upstream concurrent-reviewer-group contract has actually landed (cross-repo #3989); measure before/after wall clock and prove idempotent same-slot replay under concurrent submission.
• Exact next action: Read the capture-result collection loop in lib/review-host-relay.ts at SHA a6a73fa and record whether slots are awaited serially, then verify the upstream concurrent reviewer group contract is published before implementing parallel launch.

root-543 · P2

• Members: review happy-path MCP traffic exhausts agent context: full collectBindings inlined in status and re-pasted twice per capture #543 review happy-path MCP traffic exhausts agent context: full collectBindings inlined in status and re-pasted twice per capture
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: high
• Causal hypothesis: Design consequence of in-process host relay: the parent orchestrator is the only session, so all binding payloads (full changed-path manifest per lens) must pass through it serially. No handle-materialize indirection exists.
• Current-main evidence:
  • lib/review-candidate-view.ts:1710 — candidate context embeds full gzip+base64url manifest with SHA-256 inline; consumers must echo it back through the binding.
  • skills/_shared/review-ledger-contract.md:25 — Capture contract requires supplying the exact negotiated collect input per slot, confirming full-payload re-submission is current design.
• Rollback boundary: Components: runtime/review-integration-v2
• Evidence gap: No benchmark or trace quantifying actual token cost in Pi parent context; 200KB is reported volume but no measured compaction threshold data supplied.
• Exact next action: Design an opaque handle indirection for collectBindings: STATUS emits handle IDs, capture resolves server-side, acknowledged re-submission references the same handle. Spec the wire contract change.

[barbatdev]

Immutable baseline survey shard

Snapshot strict-v1-88113a3da8ce1ac3212facb363d91144184de689f094ec1ea8c249dca3045472 · cutoff 2026-09-01T13:25:12.459361Z · main 6c391ad9e6fc2e4ca5931a7c22eaf08b97412162

root-544 · P2

• Members: feat(review): run reviewer roles as Pi subagents (fresh constrained executor per lens) #544 feat(review): run reviewer roles as Pi subagents (fresh constrained executor per lens)
• Active PRs: none
• Historical PR evidence: none
• One-PR verdict: no
• Chain risk: true
• Confidence: medium
• Causal hypothesis: The negotiated route executes reviewer roles through the in-process pi host relay by design; there is no subagent dispatch surface for review roles. Managed-execution groundwork in feat(agents): add same-cwd managed execution #419 does not yet map review roles onto it.
• Current-main evidence:
  • skills/_shared/review-ledger-contract.md:25 — Contract describes the gentle-pi host relay launching a fresh locked-down print-mode pi subprocess in an empty scratch directory for lens captures — subagent execution already exists for lens but is described as host-internal, not exposed as parent-delegatable.
  • lib/review-host-relay.ts:62 — PI_EMPTY_OUTPUT and relay failure codes exist for the subprocess path, confirming the host-relay spawns a child process but does not surface it as a delegatable slot.
• Rollback boundary: Components: runtime/review-host-relay
• Evidence gap: No explicit design doc showing why the existing host-relay subprocess cannot be exposed as a parent-launchable subagent slot; the gap may be architectural (relay owns the process lifecycle) rather than missing code.
• Exact next action: Write an ADR evaluating whether the existing review-host-relay subprocess can be promoted to a parent-launchable constrained subagent slot, preserving native Go authority for admission, or whether a new execution surface is required.

PR #36 · rebase-resolve

• Title: fix(startup): reorganize HUD stats panel layout
• Draft: false
• Linked issues: feat(startup): compact the HUD stats panel with a stable column grid #375
• Affected roots: root-375
• Status: MERGEABLE/CLEAN but CHANGES_REQUESTED by maintainer barbatdev (2026-08-18); CI verify SUCCESS; CodeRabbit present
• Risks: Maintainer documented that LayoutBuilder.center causes 16-column horizontal shift between wide rows at 122-col boundary, dropping PATH capacity from 46 to 20 chars; No regression tests for narrow mode, 122-col threshold, long paths/branches/versions; Base is stale relative to current main (frozen SHA 6c391ad vs review baseline 2e2ca3)
• Rationale: Maintainer review explicitly requests rework: rebase onto current main, fix column grid geometry, add Closes feat(startup): compact the HUD stats panel with a stable column grid #375, add regression tests. CHANGES_REQUESTED is a hard gate. The layout direction is approved but the implementation geometry is broken per reviewer evidence.
• Uncertainty: Whether the author (aporcelli) is still active; last update 2026-08-18; no response to review yet
• Exact next action: Author rebases onto main, reworks wide-row geometry to a single consistent column grid, adds regression tests for 122-col and larger terminals, adds Closes feat(startup): compact the HUD stats panel with a stable column grid #375, requests re-review from barbatdev

PR #51 · rebase-resolve

• Title: fix(harness): refresh orchestrator prompt from disk on each access
• Draft: false
• Linked issues: feat(harness): strengthen harness persistence using Pi's context and compaction events #39
• Affected roots: root-39
• Status: CONFLICTING/DIRTY merge state; no CI check results in evidence; CodeRabbit posted 2 review rounds with unresolved suggestions (cross-platform path, redundant ENOENT branch); no human review decision
• Risks: Merge conflicts mean the branch is behind main; cannot merge without rebase; Unresolved CodeRabbit nitpicks (hardcoded Unix path in test, dead ENOENT branch) signal incomplete review iteration; No maintainer approval visible; review_decision is empty; Part of a 4-PR stacked chain; merging this alone requires the stack to be rebased together
• Rationale: CONFLICTING merge state is a hard blocker. Additionally needs author to address CodeRabbit cross-platform path fix and dead-code simplification, then rebase the entire 4-PR stack onto current main. Original fix(harness): refresh orchestrator prompt from disk on each access #40 was approved but this re-land has not yet received human review.
• Uncertainty: Whether feat(harness): strengthen harness persistence using Pi's context and compaction events #39 is still the active tracking issue or has been superseded by post-v0.6.0 harness architecture changes; no CI status data available
• Exact next action: Author rebases stack onto current main, applies tmpdir()-based test path fix and collapses ENOENT branch per CodeRabbit, then requests maintainer review

PR #52 · rebase-resolve

• Title: feat(harness): re-anchor discipline via context event on every LLM call
• Draft: false
• Linked issues: feat(harness): strengthen harness persistence using Pi's context and compaction events #39
• Affected roots: root-39
• Status: CONFLICTING/DIRTY merge state; no CI check results; CodeRabbit flagged self-referential ReturnType annotation; no human review decision; stacked on fix(harness): refresh orchestrator prompt from disk on each access #51 which is also CONFLICTING
• Risks: Blocked by its parent fix(harness): refresh orchestrator prompt from disk on each access #51 which is also CONFLICTING; cannot merge independently; Self-referential return type in buildContextReminder will cause TypeScript error on rebase; No maintainer review yet; review_decision empty; Injects ~90 tokens per call cost; no human assessment of acceptable token overhead
• Rationale: CONFLICTING merge state and dependency on conflicting parent fix(harness): refresh orchestrator prompt from disk on each access #51. Must fix the ReturnType annotation per CodeRabbit, rebase entire stack onto current main, and obtain human review.
• Uncertainty: Whether the context event API surface is stable across pi 0.74.0+ versions the project may now run
• Exact next action: Author removes explicit ReturnType annotation (let TS infer), rebases onto main after fix(harness): refresh orchestrator prompt from disk on each access #51 is resolved, then requests maintainer review of the full stack

PR #53 · rebase-resolve

• Title: feat(harness): re-anchor after session compaction
• Draft: false
• Linked issues: feat(harness): strengthen harness persistence using Pi's context and compaction events #39
• Affected roots: root-39
• Status: CONFLICTING/DIRTY merge state; no CI check results; CodeRabbit flagged module-global compactionPending race condition across concurrent sessions and hardcoded test path; no human review decision; stacked on feat(harness): re-anchor discipline via context event on every LLM call #52
• Risks: Blocked by parent chain (fix(harness): refresh orchestrator prompt from disk on each access #51→feat(harness): re-anchor discipline via context event on every LLM call #52 both CONFLICTING); Global mutable compactionPending is a genuine race condition for multi-session; CodeRabbit finding is architecturally valid; No human maintainer review on any of the four PRs; Test uses non-portable hardcoded path same as fix(harness): refresh orchestrator prompt from disk on each access #51
• Rationale: CONFLICTING state, unresolved valid race-condition finding from CodeRabbit, and stacked dependency on two conflicting parents. Author must scope compactionPending per session (Map keyed by session ID), rebase stack, fix test path.
• Uncertainty: Whether Pi's context event carries a stable session identifier usable as map key across all supported versions
• Exact next action: Author refactors compactionPending into a session-scoped Map, replaces hardcoded path with tmpdir-derived path, rebases onto main, requests maintainer review

PR #54 · rebase-resolve

• Title: feat(harness): apply slim rules layer to named and SDD agents
• Draft: false
• Linked issues: feat(harness): strengthen harness persistence using Pi's context and compaction events #39
• Affected roots: root-39
• Status: CONFLICTING/DIRTY merge state; NO CI checks (empty checks array); CodeRabbit flagged same compactionPending race condition (inherited from stack) and missing negative SDD-preflight assertions; no human review decision
• Risks: No CI run recorded at all (checks: []); no test/build evidence; Blocked by entire conflicting parent chain (fix(harness): refresh orchestrator prompt from disk on each access #51→feat(harness): re-anchor discipline via context event on every LLM call #52→feat(harness): re-anchor after session compaction #53); Revisits a deliberate design choice noted in feat(harness): strengthen harness persistence using Pi's context and compaction events #39 — requires explicit maintainer product decision on whether subagents get any rules at all; Runtime harness test changes (tests/runtime-harness.mjs) modify assertion logic for all subagent prompts; regression risk if assertions are too loose
• Rationale: CONFLICTING state, zero CI evidence, inherited race-condition finding, and this is the PR that the author themselves flags as revisiting a design decision requiring maintainer sign-off. Must rebase stack, ensure CI passes, and obtain explicit maintainer review on the design call.
• Uncertainty: Whether maintainer has decided on the slim-rules-vs-fully-clean subagent design in feat(harness): strengthen harness persistence using Pi's context and compaction events #39; no human response visible in evidence
• Exact next action: Author rebases full stack onto main, triggers CI, then requests maintainer review specifically on the slim-rules design decision noted in PR body

PR #68 · rebase-resolve

• Title: perf(prompt): improve cache-friendly prompt ordering
• Draft: false
• Linked issues: Improve prompt cache ordering for el Gentleman harness #67
• Affected roots: root-67
• Status: MERGEABLE/CLEAN; review_decision empty; maintainer verified technically sound but explicitly requested branch refresh for CI before formal approval; CodeRabbit 1 actionable comment (pre-existing, non-blocking per maintainer); related PR feat(persona): single canonical identity channel for Pi sessions #74 touches same file/area.
• Risks: Related PR feat(persona): single canonical identity channel for Pi sessions #74 modifies same extensions/gentle-ai.ts systemPrompt composition — potential conflict or redundancy if both merge independently; No formal approval recorded despite maintainer comment; Branch is stale relative to current main; CI not yet run against updated codebase
• Rationale: Maintainer comment on 2026-08-16 explicitly states 'Before formal approval and merge, please refresh the branch onto current main so the current project CI runs against the updated codebase.' This is the sole remaining blocker; the change itself is validated and the PR is structurally MERGEABLE/CLEAN.
• Uncertainty: Interaction with PR feat(persona): single canonical identity channel for Pi sessions #74 (same file, same systemPrompt ordering area) is unverified; one of them may supersede the other after rebase.
• Exact next action: Rebase perf/prompt-cache-ordering onto current main, resolve any overlap with feat(persona): single canonical identity channel for Pi sessions #74 if already merged, push to trigger project CI.

PR #163 · rebase-resolve

• Title: fix(codegraph): add injectable shell-free runner
• Draft: false
• Linked issues: fix(codegraph): execute npm shim safely on Windows #160
• Affected roots: root-160
• Status: APPROVED by maintainer 2026-08-20 with full CI reproduction (1295 passed, 0 failed); however mergeable is CONFLICTING/DIRTY — main advanced after approval; dependent Windows slice (PR 2) reopened as fork-local draft fix(codegraph): resolve safe Windows npm entry MarsSall/gentle-pi#1.
• Risks: CONFLICTING status despite explicit approval means it cannot merge as-is; Dependent PR 2 (Windows npm entry resolution) remains incomplete; merging PR 1 without closing plan for PR 2 leaves fix(codegraph): execute npm shim safely on Windows #160 partially addressed; Branch divergence from main may grow if further merges land
• Rationale: merge-ready-after-policy was incorrect: PR shows mergeable=CONFLICTING and mergeStateStatus=DIRTY due to main advancing since barbatdev's 2026-08-20 approval. Clean mergeability is a hard precondition for merge-ready-after-policy. Root coverage is complete and narrow, but conflict resolution is mechanically required before any merge decision.
• Uncertainty: mergeable is CONFLICTING; mergeStateStatus is DIRTY; main (6c391ad) diverged from PR head (1b6aea1) after maintainer approval; no evidence of rebase attempt or conflict hunk resolution
• Exact next action: git rebase fix/codegraph-runner-foundation onto current main, resolve any conflict hunks (structural additions from main), push, confirm CI passes, then merge

PR #195 · feature-review

• Title: fix: EACCES on Android external storage during skill registry refresh (#1291)
• Draft: false
• Linked issues: none explicit
• Affected roots: none proven
• Status: MERGEABLE/CLEAN; review_decision CHANGES_REQUESTED (maintainer 2026-08-18); body references gentle-ai#1291 not a local issue; maintainer confirmed local tracking issue fix(skill-registry): degrade truthfully when the project registry is unwritable #373 created; CodeRabbit CHANGES_REQUESTED state.
• Risks: Implementation suppresses ALL filesystem errors, not only permission — proven by maintainer's negative fixture returning false success {regenerated:true, skillCount:62} with no files written; Promised reason:'permission-denied' and user notification not implemented in diff; Closing issue #1291 is in gentle-ai (provider repo), not gentle-pi; local tracking is fix(skill-registry): degrade truthfully when the project registry is unwritable #373 which PR does not reference; No regression tests added
• Rationale: Maintainer issued CHANGES_REQUESTED 2026-08-18 with 7 concrete rework items (typed error handling, truthful result, test coverage, correct issue linkage). The implementation has a verified correctness defect (false success on non-permission errors). PR remains in active review cycle awaiting contributor rework.
• Uncertainty: Contributor engagement unknown — no push or comment activity since maintainer review (2026-08-18), 14 days before snapshot cutoff. Risk of staleness/abandonment.
• Exact next action: Await contributor push addressing the 7 rework items. If no activity within 14 more days, close with pointer to fix(skill-registry): degrade truthfully when the project registry is unwritable #373 and invite a new PR.

PR #202 · feature-review

• Title: fix(review): surface pre-pr binding failures
• Draft: false
• Linked issues: fix(review): surface pre-pr lifecycle binding failures #201
• Affected roots: none proven
• Status: CONFLICTING/DIRTY; review_decision empty (maintainer COMMENTED with rework instructions 2026-08-17); CodeRabbit passed 5/5 pre-merge checks; correction comment confirms issue persists on current main 49833c8.
• Risks: Broad error instanceof Error fallback misclassifies unrelated probe/config/remote failures as lifecycle-command-binding-failed; CONFLICTING — nativeOperationFailure evolved in main since PR branch was cut; Contributor lacks repo admin rights to fix type:bug label; CodeRabbit flagged 3 possibly-related PRs (fix(review): unify worktree and native authority #132, fix(review): preserva diagnósticos ante módulos duplicados #151, feat(review): adopt review integration v1 end to end #164) touching same extensions/gentle-ai.ts area
• Rationale: Maintainer review 2026-08-17 explicitly states 'Rework is required before approval' with 5 specific items: rebase, remove broad fallback, route only typed CandidateViewError path, retain tests, add negative regression. The conflict is secondary to the required source changes; once reworked, rebase resolves naturally.
• Uncertainty: No contributor activity after maintainer rework request (2026-08-17), 15 days before snapshot. Engagement status unknown. Multiple related PRs (fix(review): unify worktree and native authority #132, fix(review): preserva diagnósticos ante módulos duplicados #151, feat(review): adopt review integration v1 end to end #164) may have partially addressed the same diagnostic area already.
• Exact next action: Await contributor rework addressing the 5 items. Verify whether fix(review): unify worktree and native authority #132/ fix(review): preserva diagnósticos ante módulos duplicados #151/feat(review): adopt review integration v1 end to end #164 already delivered overlapping diagnostic routing before investing review effort on this PR.

PR #281 · feature-review

• Title: feat(skill-registry): index npm-package skills (node_modules/*/skills)
• Draft: false
• Linked issues: none explicit
• Affected roots: none proven
• Status: MERGEABLE/UNSTABLE; review_decision CHANGES_REQUESTED (maintainer 2026-08-18); CodeRabbit had multiple review rounds (rate-limited, scoped-package gaps, watcher reconciliation); maintainer requires complete architectural rework; local tracking issue fix(skill-registry): mirror Pi-resolved loaded skills #369 referenced by maintainer but not in closing_issues.
• Risks: Architectural mismatch: filesystem scan duplicates Pi's own resource loader without matching its resolution logic (glob, exclusions, filters, disabled resources, project-local installs); Scoped npm packages (@scope/pkg/skills) not handled per CodeRabbit Major finding; Watcher reconciliation gap: newly discovered package skill dirs never added to activeWatchers; npm root watcher absent at startup is never retried; No regression tests; UNSTABLE merge state despite MERGEABLE flag; Maintainer directed absorbing PR feat(skill-registry): index pi.skills project-declared skill dirs #285 into this consolidation — scope expansion implied; Body identifies separate shouldSkipDuplicateExtensionLoad path bug (flagged as out of scope)
• Rationale: Maintainer CHANGES_REQUESTED 2026-08-18 with complete rework direction: replace direct filesystem scanning with Pi's resolved before_agent_start.systemPromptOptions.skills as the authority, absorb feat(skill-registry): index pi.skills project-declared skill dirs #285, add comprehensive acceptance coverage per fix(skill-registry): mirror Pi-resolved loaded skills #369. This is a fundamental design change, not a mechanical fix. PR remains in active review cycle.
• Uncertainty: Contributor engagement unknown; last activity 2026-08-06 (CodeRabbit), maintainer review 2026-08-18, 14 days before snapshot. Rework scope is substantially larger than original patch. Whether contributor will adopt the consolidation approach or abandon is unclear. Relationship with feat(skill-registry): index pi.skills project-declared skill dirs #285 and fix(skill-registry): mirror Pi-resolved loaded skills #369 requires cross-PR coordination.
• Exact next action: Await contributor rework per maintainer's architectural direction. If no activity within 21 days, close with link to fix(skill-registry): mirror Pi-resolved loaded skills #369 and a detailed comment capturing the valid discovery-gap analysis and the consolidation requirements so a new contributor or maintainer can implement from scratch.

PR #351 · retain-and-verify

• Title: test(agents): assert per-JD-agent model profiles can diverge
• Draft: false
• Linked issues: Release gentle-pi 2.5.0 pinned to gentle-ai v2.7.0 #703
• Affected roots: none proven
• Status: mergeable MERGEABLE but merge_state_status UNSTABLE; only a CodeRabbit check is listed, no CI verify run in evidence; review_decision empty (only a Copilot COMMENTED entry that failed on quota); author reports branch merged up to main ce0db0f and local suite green; CodeRabbit pre-merge 'Linked Issues check' is a warning, not a pass.
• Risks: No CI verify evidence on the current head; UNSTABLE merge state may hide a failing/pending check (body also documents a pre-existing tests/runtime-harness.mjs failure on main).; Closure claim is config-surface only — CodeRabbit states the test does not verify an end-to-end JD run or runtime delegate calls, so gentle-ai#703's runtime diversity guarantee is not proven solved.; Cross-repo closing reference: this PR cannot by itself close gentle-ai#703; the routing itself landed elsewhere.; Branch drift versus frozen main 6c391ad after the Aug 29 merge-up (head cc9c72e).
• Rationale: Single-file test addition with no production behavior change and minimal stated merge risk, so nothing supports closing or splitting; but the only available checks are CodeRabbit-side, merge_state_status is UNSTABLE, no human review decision exists, and the reviewer explicitly flags that the test does not cover the runtime path the linked issue is about. That combination is a verify condition, not a merge condition.
• Uncertainty: Whether UNSTABLE comes from a pending required check or a failing one; whether maintainers accept apply-pass divergence as sufficient closure evidence for gentle-ai#703; the state of the linked issue and of the routing work claimed to have landed earlier.
• Exact next action: Re-run/status-check CI verify on head cc9c72e, then have a maintainer decide in gentle-ai#703 whether the config-surface divergence test closes it or whether the referenced E2E judge-model test must still be written.

PR #355 · feature-review

• Title: fix(startup-banner): eliminate flicker, freeze, and reload-damage intro
• Draft: false
• Linked issues: none explicit
• Affected roots: none proven
• Status: mergeable MERGEABLE, merge_state_status UNSTABLE; only a CodeRabbit check listed (no CI verify in evidence); review_decision empty — CodeRabbit COMMENTED with 1 unresolved actionable comment; CodeRabbit rates Merge Risk Moderate and says the cleanup gap should be addressed before merging.
• Risks: Unresolved actionable finding: render-error catch paths and the ctx.ui.custom() rejection path can leave the full-width overlay and TUI cursor model in an unfinished/stuck state — the exact class of defect this PR claims to fix.; Interactive terminal-overlay behavior is not covered by any test in the diff, and no CI verify run is visible for head a2e2eb9.; No linked issue, so the 'originally reported and verified' symptom set cannot be traced to a durable root record; closure is self-asserted from local reproduction on 2026-08-15.; Fixes are described as a local copy recovered from a worktree and squashed from two commits, with no human maintainer review yet.; Branch drift versus frozen main 6c391ad (review base 3031136).
• Rationale: This changes user-visible TUI lifecycle behavior (overlay show/teardown, reload gating) with a still-open reviewer finding on the error/cleanup paths and a Moderate merge risk. It needs a substantive human functional review of the teardown contract before any merge-readiness judgement; there is no conflict or linkage ambiguity that would make rebase-resolve or deduplicate-select-authority the primary action.
• Uncertainty: Whether the underlying Ghostty-reported symptoms have a tracked issue anywhere (none is referenced or auto-linked), and whether the reported 204/204 local suite result corresponds to any CI run — no verify check appears in the evidence.
• Exact next action: Have a maintainer review the overlay teardown contract in extensions/startup-banner.ts and resolve the CodeRabbit inline comment by making every render-error catch path and the ctx.ui.custom() rejection path call cleanup() then done(undefined).

PR #371 · retain-and-verify

• Title: fix(review): detect a linked worktree missing its project-local .pi config
• Draft: false
• Linked issues: review: a linked worktree loses the project-local .pi configuration, so review fails claiming pi is an unsupported runtime #370
• Affected roots: root-370
• Status: CI verify SUCCESS, merge_state_status CLEAN, mergeable MERGEABLE, label type:bug; review_decision empty (no approval). CodeRabbit posted 1 actionable inline comment (async, time-bounded git probe) and rates Merge Risk Moderate; a human reviewer (PwnLabsmx, 2026-08-21) raised two unresolved boundaries plus a 400-line budget objection. Body states the bug(review): hardcoded 600s pi host-relay reviewer timeout strands large candidates in a FINALIZE retry loop #367 confirming run did not isolate which of provisioning vs. handshake export was load-bearing.
• Risks: inspectWorktreeWorkspaceConfig() runs synchronous, unbounded Git probes from both session_start and /gentle:status — a stalled git process can freeze user-facing paths (CodeRabbit and human reviewer agree).; Closure claim is not testable as written: tests instantiate createGentleAiExtension() directly, so they cannot prove the project-local-install case where the extension never loads in the worktree at all.; +549/-1 exceeds the repository's stated 400-line review budget with no size:exception label present.; Detection can only announce the gap; it cannot make the pi review runtime work, so review: a linked worktree loses the project-local .pi configuration, so review fails claiming pi is an unsupported runtime #370's user-visible failure persists after merge — the root is mitigated, not solved.; Related refusal-message work is explicitly left to gentle-ai#3440 / gentle-ai#3443, so scope boundaries between repos are still open.
• Rationale: Green CI and a CLEAN mergeable state would support merge if not for an unresolved human review that names concrete correctness and evidence gaps, an empty review_decision (no maintainer has approved), and an unmet size budget. The correct posture is to keep it open, verify the closure claim is actually demonstrable, and bound the startup probe — not to merge, split, or close.
• Uncertainty: Whether review: a linked worktree loses the project-local .pi configuration, so review fails claiming pi is an unsupported runtime #370 remains open in the evidence (no issue state supplied); whether the 400-line budget will be met by splitting tests into a follow-up or by an explicit size:exception; whether the async-probe refactor was landed after 2d92255 — the head shown is still the reviewed commit and the CodeRabbit inline comment remains unresolved.
• Exact next action: Convert the Git probes to an asynchronous, time-bounded call with an injectable test seam (timeout/spawn failure resolving to undetermined), narrow the guarantee to sessions where gentle-pi is already loaded or add a real runtime-startup test, then reduce the diff under 400 lines or request size:exception.

PR #387 · rebase-resolve

• Title: fix(orchestrator): contain execution-surface escalation
• Draft: false
• Linked issues: fix(orchestrator): prevent unauthorized execution-surface escalation #379
• Affected roots: root-379
• Status: DIRTY / mergeable CONFLICTING against frozen main 6c391ad (body itself reviewed against main 514f9e5). CI verify SUCCESS on head 3e0f8cf but that head is now in conflict. review_decision CHANGES_REQUESTED (Alan-TheGentleman, 2026-08-29). Labels status:needs-review, type:bug. CodeRabbit has open actionable comments including a dependency-graph contradiction.
• Risks: Blocking changes-requested: the guard applies managed-runtime containment before any managed runtime has been explicitly selected, so it changes default-off/unmanaged behavior.; Verification policy is contradictory and whole-surface tests can mask per-file gaps; the documented restoration path is not executable today, so the claimed guarantee is not demonstrable.; Merge conflicts (CONFLICTING) plus a 9-file diff mixing runtime guard, orchestrator/SDD policy prose, README migration wording, and three contract-test suites — the reviewer explicitly asks to keep the stale-companion deletion separate from the authority mechanism.; Unresolved CodeRabbit findings: delegated-direct instruction still routes to the platform's native bounded worker, and assets/sdd-orchestrator-workflow.md defines design as branching from proposal, which can start design before spec and conflicts with README's phase order.; Diff is +278/-51 (329 changed lines), inside the 400-line budget, but the mixed policy/runtime scope still weakens reviewability; feat(agents): execute managed tasks in linked worktrees #376 positive target-cwd support is deferred, so an unsupported cwd yields a stop rather than a capability.
• Rationale: The PR is structurally blocked (merge_state_status DIRTY, CONFLICTING) and carries an explicit CHANGES_REQUESTED decision naming the required redesign — gate containment on a real selected-authority seam, semantic per-file guards, default-off negative controls, one runnable restoration path, and separation of the companion-deletion change. Both the mechanical conflict and the requested design changes must be resolved on the branch before any re-review; nothing in the evidence supports merging or closing it as superseded.
• Uncertainty: Whether fix(orchestrator): prevent unauthorized execution-surface escalation #379 remains open and whether any other branch/PR now owns the selected-authority seam Alan-TheGentleman is waiting on (his review says 'wait for a real selected-authority seam', which could later justify supersede-close, but no successor is identified in this evidence). Whether feat(agents): execute managed tasks in linked worktrees #376 has landed the positive target-cwd capability the PR defers to.
• Exact next action: Update the branch onto frozen main 6c391ad resolving the conflicts, then re-scope per the changes-requested review: gate containment on an explicit selected-authority seam, add per-file semantic guards with default-off negative controls, make one restoration path runnable, and split the README stale-companion deletion into its own change.

PR #420 · rebase-resolve

• Title: fix(installer): give the Windows source build its own timeout and report timeouts precisely
• Draft: false
• Linked issues: bug(installer): Windows source build exceeds the postinstall timeout #400
• Affected roots: none proven
• Status: DIRTY / mergeable CONFLICTING; no CI verify run in the evidence (only a CodeRabbit check entry). review_decision empty. CodeRabbit recent review on 7682958..a35733b posted no actionable comments; an earlier inline asking for Windows process-tree termination before staging cleanup is not shown as resolved; Merge Risk rated Low with named follow-ups.
• Risks: Merge conflicts against frozen main 6c391ad — cannot merge as-is.; No CI verify evidence for head a35733b; the only published verification is the author's own Windows x64 run (130.7s end-to-end) plus 4 acknowledged pre-existing POSIX-specific suite failures on Windows.; Open process-lifecycle concern from review: a timeout kill may not terminate the whole Windows process tree before safeRemoveDirectory(stagingDirectory) runs, risking leftover compiler processes or staging files.; Closure depends on a timing threshold measured on two machines (232s vs 119.3s); the 600s bound is headroom-based, not a contract validated across the Windows install matrix.; Timeout-message wording was required to become phase-neutral (not assert compilation or checksum/SumDB status); current head wording is not confirmed in this evidence.
• Rationale: The binding blocker is mechanical and evidenced: merge_state_status DIRTY with mergeable CONFLICTING on a main that has moved past this branch's base. The substantive review points (process-tree cleanup, Windows timeout-contract validation, phase-neutral message) are follow-ups CodeRabbit itself rates Low merge risk, so the correct next step is to bring the branch current and resolve the conflicts while confirming CI and the open cleanup concern, not to close, split, or declare it merge-ready.
• Uncertainty: Whether bug(installer): Windows source build exceeds the postinstall timeout #400 is still open and whether another installer PR has already changed the same timeout path, which would make deduplicate-select-authority the right call instead; whether the process-tree termination request was addressed by the later commits, since those reviews are truncated to empty bodies; no CI check status is supplied at all for the current head.
• Exact next action: Rebase/merge current main 6c391ad into fix/400-windows-source-build-timeout, resolve the conflicts in scripts/gentle-ai-installer.mjs, push to obtain a CI verify run on the new head, and confirm whether Windows process-tree termination before staging removal still needs work.

PR #426 · rebase-resolve

• Title: fix(quiet-tools): preserve Pi's configured Bash shellPath on Windows
• Draft: false
• Linked issues: fix(quiet-tools): preserve Pi's configured Bash shellPath on Windows #107
• Affected roots: root-107
• Status: CI verify SUCCESS; CodeRabbit linked-issue check passed and reported minimal merge risk; mergeability is CONFLICTING/DIRTY.
• Risks: Branch conflicts with main and cannot currently merge.; No human review or approval is present in the evidence.; Open PR status does not prove issue fix(quiet-tools): preserve Pi's configured Bash shellPath on Windows #107 root is solved after conflict resolution.
• Rationale: CI and automated review evidence are favorable, but GitHub reports the PR as CONFLICTING/DIRTY, so the blocker is mergeability rather than missing scope evidence.
• Uncertainty: Cannot verify whether the conflict resolution will preserve the reported test results or close the root cause in fix(quiet-tools): preserve Pi's configured Bash shellPath on Windows #107.
• Exact next action: Rebase fix/quiet-tools-bash-shellpath onto main, resolve conflicts, rerun pnpm test and the quiet-tool tests, then request review.

PR #427 · retain-and-verify

• Title: fix(sdd): remove provider-specific model pin from sdd-init
• Draft: false
• Linked issues: bug(sdd): sdd-init hardcodes unavailable provider model #216
• Affected roots: root-216
• Status: CI verify SUCCESS; CodeRabbit review was rate limited and did not produce a completed actionable review; mergeability is CLEAN/MERGEABLE; no human reviews are present.
• Risks: Automated review evidence is incomplete.; No human approval is recorded.; Migration and hash-preservation behavior is not independently verified by the supplied evidence.
• Rationale: The PR is mergeable and CI passed, but the required strong check/review evidence is incomplete because CodeRabbit was rate limited and no human review is present.
• Uncertainty: Cannot confirm whether repository policy requires completed CodeRabbit review or human approval before merging this migration-related fix.
• Exact next action: Obtain a completed CodeRabbit or human review against bug(sdd): sdd-init hardcodes unavailable provider model #216 and verify the migration tests before considering the PR policy-ready.

PR #428 · split-scope

• Title: docs(readme): document review-mode and sdd-preflight commands
• Draft: false
• Linked issues: fix(docs): document the registered gentle:review-mode command and align the README command table with the quick-start list #300
• Affected roots: root-300
• Status: No successful CI evidence is present; mergeability is UNSTABLE/MERGEABLE; CodeRabbit posted actionable terminology-consistency comments; no human approval is present; required type label is missing.
• Risks: The PR claims documentation-only changes but also modifies source code.; The latest CodeRabbit actionable comment about canonical review-mode terminology is not shown as resolved.; No successful CI evidence is present and merge state is UNSTABLE.; Missing type:docs can block PR validation according to the PR body and comments.
• Rationale: Evidence shows a docs issue fix(docs): document the registered gentle:review-mode command and align the README command table with the quick-start list #300 linked to a PR that also changes extensions/gentle-ai.ts, while the PR still describes itself as documentation-only; the safe inventory action is to separate or re-authorize the code-description change.
• Uncertainty: The supplied fix(docs): document the registered gentle:review-mode command and align the README command table with the quick-start list #300 scope text is not included, so it is unclear whether the one-line source description update is within the approved issue.
• Exact next action: Separate the README command-table documentation from the extensions/gentle-ai.ts terminology update, or obtain explicit issue scope approval for the source change.

PR #429 · retain-and-verify

• Title: test(agents): pin YAML tool parsing contract for packaged agents
• Draft: false
• Linked issues: bug(agents): YAML tools frontmatter leaves Gentle subagents without filesystem tools #62
• Affected roots: root-62
• Status: No successful CI evidence is present; mergeability is UNSTABLE/MERGEABLE; CodeRabbit review flagged runtime compatibility and left actionable comments; the PR has a slop label instead of a required type label.
• Risks: The tests validate a vendored parser contract but may not prove runtime parser compatibility.; No CI pass is present in the supplied checks.; Missing required type label can block PR validation.; Closure of bug(agents): YAML tools frontmatter leaves Gentle subagents without filesystem tools #62 depends on whether a regression-only guard satisfies the approved issue closure condition.
• Rationale: The PR is a test-only guard with a clear issue reference, but the evidence lacks CI verification, has unresolved review concerns about runtime compatibility, and cannot prove the root cause in bug(agents): YAML tools frontmatter leaves Gentle subagents without filesystem tools #62 is fully handled.
• Uncertainty: Cannot confirm from the supplied evidence whether the bug(agents): YAML tools frontmatter leaves Gentle subagents without filesystem tools #62 closure condition explicitly accepts a regression test without runtime parser enforcement.
• Exact next action: Run CI, resolve the CodeRabbit runtime-compatibility comment, and obtain maintainer confirmation that bug(agents): YAML tools frontmatter leaves Gentle subagents without filesystem tools #62 can be closed by this regression test alone.

PR #435 · rebase-resolve

• Title: feat(agents): add owned activity projection contract
• Draft: false
• Linked issues: feat(agents): define a Gentle-owned read-only activity projection #432
• Affected roots: root-432
• Status: CI verify SUCCESS; CodeRabbit recent review reports no actionable comments and minimal merge risk; mergeability is CONFLICTING/DIRTY; no human reviews are present.
• Risks: Branch conflicts with main and cannot currently merge.; No human review or approval is recorded for a complex new contract.; PR explicitly defers runtime producer and UI integration, so broader activity feature roots remain unproven.; CodeRabbit reported a docstring coverage warning.
• Rationale: The supplied evidence shows favorable CI and no actionable recent CodeRabbit comments, but the immediate blocker is CONFLICTING/DIRTY mergeability, which requires rebase and revalidation before any merge decision.
• Uncertainty: Cannot verify whether the conflict resolution will preserve test results or whether feat(agents): define a Gentle-owned read-only activity projection #432 closure requires the later producer/UI work described in the PR.
• Exact next action: Rebase feat/432-agent-activity-projection onto main, resolve conflicts, rerun the full test suite and CI, then request maintainer review.

PR #439 · rebase-resolve

• Title: refactor(agents): isolate delegation admission seam
• Draft: false
• Linked issues: refactor(agents): isolate delegation admission behind a Gentle Pi seam #436
• Affected roots: root-436
• Status: CI verify SUCCESS on head 1b45911; CodeRabbit context present with no actionable comments and self-reported Minimal merge risk; no human review requests, review_decision empty; merge_state_status DIRTY / mergeable CONFLICTING against frozen main 6c391ad.
• Risks: Merge state is CONFLICTING, so the verified head is not the deliverable candidate and CI evidence does not apply to a conflict-free tree.; extensions/gentle-ai.ts is a hot file also modified by PR fix(review): preserve consent binding continuity #457, so the conflict likely needs deliberate resolution rather than a blind rebase.; No human review decision recorded; behavior-preservation claims (payload identity, public block reasons) are author- and bot-asserted only.; Open PR is not proof that refactor(agents): isolate delegation admission behind a Gentle Pi seam #436 is solved; the seam is characterization-only and lands nothing downstream.
• Rationale: The only hard blocker in the evidence is mergeability: merge_state_status DIRTY with mergeable CONFLICTING while base_ref is main. Checks themselves pass and scope is narrow (3 files, 150 changed lines of 400), so no split or linkage clarification is indicated; the branch must be rebased onto current main and conflicts in extensions/gentle-ai.ts resolved, after which the seam tests should be re-run on the new head.
• Uncertainty: Cannot see the actual conflicting hunks or whether the conflict is with PR fix(review): preserve consent binding continuity #457's consent changes already in main; cannot confirm the rebase will keep the 8 characterization tests green.
• Exact next action: Rebase refactor/436-delegation-admission onto 6c391ad, resolve conflicts in extensions/gentle-ai.ts, force-push, and re-run pnpm test plus the focused delegation-admission suite on the new head before requesting human review.

PR #441 · retain-and-verify

• Title: feat(runtime): add authoritative managed-task registry
• Draft: false
• Linked issues: feat(runtime): add authoritative managed-task registry #440
• Affected roots: root-440
• Status: CI verify SUCCESS on head fd2d7f5; merge_state_status CLEAN / mergeable MERGEABLE; CodeRabbit context with no actionable comments and Minimal merge risk; review_decision empty and reviews array empty; 399 changed lines against a stated 400 budget; docstring-coverage warning only.
• Risks: No human review at all on a component self-described as the authoritative identity/lifecycle registry; correctness claims (81 state/transition combinations, hostile-Proxy containment, locale-free ordering) are unverified by a second party.; Line count sits at 399/400, leaving zero headroom, and the file is additive-only so any post-review fix risks exceeding the review budget.; Registry is not yet wired to anything, so it neither proves feat(runtime): add authoritative managed-task registry #440 solved in production terms nor can its semantics be validated by integration evidence.; Being MERGEABLE and green is not sufficient policy evidence: no approving review, no maintainer assignment, and it is one of three sibling PRs (refactor(agents): isolate delegation admission seam #439, feat(runtime): add isolated Pi AgentSession adapter #445) in the same feat(agents): add same-cwd managed execution #419 stack whose landing order is unstated.
• Rationale: Scope is single-purpose, cleanly linked to feat(runtime): add authoritative managed-task registry #440, mergeable and green, so rebase, split, deduplication and closure actions are unsupported. But merge-ready-after-policy requires strong check AND review evidence, and reviews is empty with review_decision blank; for a component declared to be the lifecycle authority this gap is material. The conservative, evidence-backed move is to keep the PR open and obtain independent verification of the transition/bounds/ordering semantics before merge.
• Uncertainty: Cannot confirm from the evidence whether repository branch-protection actually requires a human approval, nor whether feat(runtime): add authoritative managed-task registry #441 is intended to merge before, after, or independently of refactor(agents): isolate delegation admission seam #439 and feat(runtime): add isolated Pi AgentSession adapter #445 despite the body's 'lands independently and imports neither sibling' claim.
• Exact next action: Request a human code review on head fd2d7f5 focused on the transition graph, revision bounds, subscription reentrancy and terminal-ID ordering, and record the intended merge order within the feat(agents): add same-cwd managed execution #419 stack on the PR before merging.

PR #445 · retain-and-verify

• Title: feat(runtime): add isolated Pi AgentSession adapter
• Draft: false
• Linked issues: feat(runtime): add isolated Pi AgentSession adapter #444
• Affected roots: root-444
• Status: Mergeable MERGEABLE / CLEAN on head 56014aa; three CI 'verify' SUCCESS entries are reported but the author states GitHub Actions never created a pull_request run for the current SHA after synchronize/reopened retries; CodeRabbit review state COMMENTED with 1 actionable inline typing finding on the pre-fix range; review_decision empty.
• Risks: CI provenance for the reviewed head is unconfirmed: the SUCCESS entries are tied to the earlier head 3857511 and possibly reruns, not demonstrably to 56014aa, so 'checks pass' cannot be treated as current-head evidence.; An open CodeRabbit actionable comment on Specification field typing (model, modelRegistry, thinkingLevel vs CreateAgentSessionOptions) has no recorded human disposition; review_decision is empty.; Version-sensitive contract: the adapter intentionally excludes newer Pi 0.74 ResourceLoader methods while the dev install is 0.84.3, so drift risk is real and unverified by any independent check.; Landed capability is not proof feat(runtime): add isolated Pi AgentSession adapter #444 is solved: it is unwired, so its behavior is only established by its own 5 focused tests.; 372/400 review workload leaves little room for a post-review corrective commit.
• Rationale: The PR is mergeable and cleanly linked, so rebase-resolve is not required, and its scope is coherent so split-scope and deduplication do not apply. What blocks merge-readiness is verification integrity, not code shape: the author explicitly records that a maintainer still needs to approve or trigger CI for 56014aa and that no pull_request run exists for it. Without current-head CI plus disposition of the inline typing finding, strong check/review evidence is absent and the conservative action is to keep the PR open and re-establish evidence.
• Uncertainty: Cannot determine from the supplied evidence whether any of the three SUCCESS verify entries actually bound to 56014aa, nor whether the CodeRabbit inline typing comment was marked resolved by the reviewer rather than only by the author.
• Exact next action: Have a maintainer trigger or approve the 'verify' workflow for head 56014aa, confirm the bound check conclusion, and explicitly resolve or reject the CodeRabbit Specification typing comment before merge.

PR #457 · rebase-resolve

• Title: fix(review): preserve consent binding continuity
• Draft: false
• Linked issues: fix(review): preserve consent bindings across active Pi sessions #455
• Affected roots: root-455
• Status: CI verify SUCCESS and CodeRabbit context reported on head abf1889; merge_state_status DIRTY / mergeable CONFLICTING against frozen main; reviews contains one CodeRabbit COMMENTED review with 1 actionable inline finding and no human approval; review_decision empty; 23/50 added lines in extensions/gentle-ai.ts.
• Risks: CONFLICTING merge state means the approved-candidate evidence (including the body's claim of a candidate-scoped RDD review after rebasing onto origin/main) no longer describes a mergeable candidate.; CodeRabbit rates Merge Risk 🟡 Moderate and identifies a concrete failure mode: the binding is consumed before validation completes, so a pre-provider rejection can leave candidate state retained without its cleanup path while the originating session loses the binding.; That unresolved inline finding is a lifecycle/leak-class defect in a consent authority path, the highest-consequence area of the RDD flow, and there is no human disposition recorded.; Overlaps extensions/gentle-ai.ts with PR refactor(agents): isolate delegation admission seam #439, so conflict resolution is non-mechanical and cross-PR ordering matters.; Test delta is a parity suite only; 12 deleted lines indicate behavior replacement whose downstream callers are not evidenced.
• Rationale: The hard, unambiguous blocker in the evidence is merge_state_status DIRTY / mergeable CONFLICTING on base main, and the change is small, single-purpose and cleanly closes fix(review): preserve consent bindings across active Pi sessions #455, so split-scope, clarify-linkage and deduplication are unsupported. Rebase is the first required step; the unresolved Moderate-risk cleanup finding must then be addressed and re-reviewed on the conflict-free head, because merging on the verified-but-conflicting head is impossible and any silent conflict resolution could itself change the claim/cleanup ordering.
• Uncertainty: Cannot determine which main-side change caused the conflict (PR refactor(agents): isolate delegation admission seam #439's routing edits are a plausible candidate but unproven), nor whether the author added cleanup guarantees after CodeRabbit's 2026-08-27T19:19:21Z review, since no later head or reply is present in the evidence.
• Exact next action: Rebase fix/455-consent-binding-continuity onto 6c391ad, resolve the extensions/gentle-ai.ts conflict, then guarantee pending.cleanupCandidate() on every pre-provider failure path per the CodeRabbit inline finding and re-run tests/native-review-parity.test.ts on the new head.

PR #481 · split-scope

• Title: docs: proponer un modelo resiliente de supervisión watchdog y recuperación
• Draft: false
• Linked issues: none explicit
• Affected roots: none proven
• Status: No repository CI check evidence at all beyond the CodeRabbit context; merge_state_status UNSTABLE (mergeable MERGEABLE); no labels; no linked or closing issue; two CodeRabbit COMMENTED reviews reporting 7 then 1 actionable findings (provenance metadata, source-path mismatch vs SOURCE-SHA256.txt, progress-based invalidation semantics, continuation lineage, confidence-label consistency) plus a docstring-coverage warning; a later CodeRabbit run hit the review rate limit.
• Risks: No CI/verify signal exists for this head, so nothing in the evidence establishes that the 13.5k-line reference file is even syntactically checked; UNSTABLE merge state means required contexts are missing or pending.; The PR mixes a documentation proposal with a large vendored reference implementation and an author-authored design decision it explicitly defers to maintainers, which exceeds a reviewable unit and bypasses the project's small-PR convention the author themselves proposes.; No approved issue, no type:* label, and a Spanish-language body and titles conflict with the repository's conventional-commit/label and English-artifact norms, so provenance of approval cannot be established.; Reproducibility of the claimed validation evidence is explicitly unsupported: CodeRabbit found runs lacking run IDs, revisions, exact commands and raw-log references, and confidence labels inconsistent with their own definitions.; Provenance and licensing exposure: a single-file 13,516-line reference drop with a README path that mismatches SOURCE-SHA256.txt, so integrity of the sanitization/licensing claims is unverified.; Open PR is not evidence any supervision or recovery root is solved; content is self-described as proposal and reference.
• Rationale: The evidence supports restructuring rather than rebase or merge: the author states the PR 'contiene deliberadamente documentación y una implementación de referencia' and offers to 'dividir el código en PRs más pequeños' upstream, while the diff shows 23 files / 14,725 additions dominated by one 13,516-line reference file. Combined with zero CI check evidence, UNSTABLE merge state, no linked issue or label, and 8 unresolved CodeRabbit findings spanning provenance, licensing and validation reproducibility, no smaller action can make the unit reviewable. Splitting into a docs-only proposal (issue-backed) plus separately-scoped implementation changes, each with provenance metadata, is the proportionate response.
• Uncertainty: Cannot determine from the evidence whether an accepted issue or maintainer endorsement exists for this contribution, whether any project CI gate would ever run on a docs-only diff, or whether the 13,516-line reference file is original author code versus third-party derived (LICENSE-NOTE.md is 8 additions and its content is not supplied).
• Exact next action: Ask the author to close or restack this PR into separate units — an issue-backed docs-only PR for the supervision proposal, a PR per subsystem with the CodeRabbit-required provenance/run-metadata and corrected source-path/SHA256 alignment, and any future implementation PR for phase-router behavior — and to add a type:docs label and an English body before review resumes.

[barbatdev]

Immutable baseline survey shard

Snapshot strict-v1-88113a3da8ce1ac3212facb363d91144184de689f094ec1ea8c249dca3045472 · cutoff 2026-09-01T13:25:12.459361Z · main 6c391ad9e6fc2e4ca5931a7c22eaf08b97412162

PR #501 · retain-and-verify

• Title: fix(agents): remove pi-mono path residue
• Draft: false
• Linked issues: fix(docs): prompts and assets reference pi-mono repository paths and URLs that do not exist in gentle-pi #298
• Affected roots: root-298
• Status: CI verify: SUCCESS; mergeable CLEAN; CodeRabbit COMMENTED only (no approve); 1 unresolved actionable finding on scanned-paths coverage; platform-path risk flagged (backslash separators).
• Risks: Linked issue fix(docs): prompts and assets reference pi-mono repository paths and URLs that do not exist in gentle-pi #298 may require removal of residue from prompts/gpr.md and prompts/gcl.md, but PR only touches assets/agents/sdd-design.md — CodeRabbit flagged this as a failed Linked-Issues check.; Regression assertions may fail on Windows due to hard-coded forward-slash path expectations.
• Rationale: CI passed but no human approval exists, scope-vs-issue gap remains open (per CodeRabbit Linked Issues warning), and platform-path concern is unresolved. Cannot confirm the issue's full requirements are met without verification.
• Uncertainty: Cannot inspect issue fix(docs): prompts and assets reference pi-mono repository paths and URLs that do not exist in gentle-pi #298 body from supplied evidence; whether prompts/gpr.md and prompts/gcl.md are truly in-scope depends on the issue's wording.
• Exact next action: Verify issue fix(docs): prompts and assets reference pi-mono repository paths and URLs that do not exist in gentle-pi #298 acceptance criteria against the two changed files; if residue exists in prompts/gpr.md or prompts/gcl.md, expand scope or close explicitly; fix test path-joining to be platform-native.

PR #504 · retain-and-verify

• Title: fix: clarify guarded command confirmations
• Draft: false
• Linked issues: bug(guard): confirmation prompt hides the guarded action #460
• Affected roots: root-460
• Status: CI verify: SUCCESS; mergeable CLEAN; CodeRabbit COMMENTED (no approve); 1 Major finding: duplicate published declaration at line 747 of tests/gentle-ai.test.ts 'prevents the test module from parsing'; no native RDD receipt due to adapter consent loop.
• Risks: CodeRabbit reports the current head fails to parse the test module, contradicting the CI SUCCESS result — either CI ran before the introducing commit or the test runner tolerated the error.; No independent verification (native RDD) exists because of the documented consent-prompt loop bug.
• Rationale: The CI/review contradiction on test-file parseability must be resolved before trusting the green check. CodeRabbit explicitly says 'not merge-ready until that test error is removed.'
• Uncertainty: Cannot confirm from evidence alone whether head d401f actually contains the duplicate declaration or whether CodeRabbit's line-747 observation is a diff-context misplacement.
• Exact next action: Re-run node --experimental-strip-types --test tests/gentle-ai.test.ts on head d401f to confirm parseability; remove or fix the duplicate declaration if real; then request a human review pass.

PR #505 · retain-and-verify

• Title: fix(review): match the native eight-line abandon authorization binding
• Draft: false
• Linked issues: Pi compact facade cannot relay maintainer authorization binding for review abandon #466
• Affected roots: root-466
• Status: No CI check entries in evidence; mergeable UNSTABLE; CodeRabbit 'No actionable comments' + Merge Risk Minimal; Copilot review failed due to quota; no human approval.
• Risks: UNSTABLE merge state means at least one required check is missing or failing — cannot confirm green without CI evidence.; The contributor noted a known environment-dependent test failure unrelated to this change; verification still pending.; Issue Pi compact facade cannot relay maintainer authorization binding for review abandon #466 has status:approved label per contributor checklist, but no maintainer code-review approval exists.
• Rationale: Evidence contains no passing CI entry and the merge state is UNSTABLE; CodeRabbit minimal-risk assessment cannot substitute for the required check. Awaiting check completion is the conservative move.
• Uncertainty: Cannot determine from evidence which check is producing the UNSTABLE status.
• Exact next action: Confirm the CI workflow reaches SUCCESS on head d5100; then request maintainer review (alan-thegentleman per CodeRabbit suggestion).

PR #512 · feature-review

• Title: fix(orchestrator): enforce parent hot memory persistence on inline direct tasks
• Draft: false
• Linked issues: fix(orchestrator): enforce parent hot memory persistence on inline direct tasks #511
• Affected roots: root-511
• Status: No CI workflow entry; mergeable UNSTABLE; CodeRabbit 3 actionable unresolved findings + Merge Risk HIGH; no human approval.
• Risks: HIGH merge risk flagged: contract mandates persistence but does not guard against saving secrets, credentials, personal data, raw untrusted content, or speculation — creates durable contamination vector.; Scope wording may leave new rules unenforced for non-SDD work.; No CI verification evidence.
• Rationale: CodeRabbit's HIGH risk assessment identifies a real semantic defect in the instruction design (unbounded mem_save directive). This is not mechanical; it needs architectural review of the safety guard before the change can be accepted.
• Uncertainty: Cannot confirm whether the existing memory contract already implicitly excludes unsafe data; the PR diff is only 7+1 lines.
• Exact next action: Maintainer must review whether the memory contract requires an explicit data-hygiene guard (reject secrets, PII, raw untrusted content) and whether the scope statement covers all non-SDD work.

PR #532 · rebase-resolve

• Title: fix(review): support pre-start untracked selection
• Draft: false
• Linked issues: gentle_review_capture cannot consume pre-start collect slots — intended_untracked_selection_required gate is unanswerable through the wrapper #502
• Affected roots: none proven
• Status: No CI workflow entry; mergeable CONFLICTING (DIRTY); CodeRabbit Merge Risk HIGH (security finding); zero reviews; no approval.
• Risks: CONFLICTING merge state: the 389-line change requires a rebase onto current main before any further evaluation is meaningful.; HIGH security risk: no repository-ownership check; could target unrelated local worktrees.; Pending selection may be lost or preserved unexpectedly on failed/competing starts.; Complex change (4/5 review-effort) with zero human review.
• Rationale: GitHub reports CONFLICTING/DIRTY; no review or evaluation can proceed until the branch merges cleanly with main. The security findings will be re-examined post-rebase but the blocker today is mechanical.
• Uncertainty: Cannot determine conflict surface area without inspecting main vs head tree diff.
• Exact next action: Rebase head 3fe5a onto current main (6c391ad), resolve conflicts, push, and re-request CodeRabbit review to confirm the ownership-check gap after the rebase.

PR #539 · rebase-resolve

• Title: test/ci: harden Windows test baseline and enable Windows CI
• Draft: false
• Linked issues: none explicit
• Affected roots: none proven
• Status: open non-draft; mergeability CONFLICTING/DIRTY; review_decision empty; only CodeRabbit automated review evidence; no GitHub Actions result supplied.
• Risks: Base branch conflicts must be resolved before merge.; CodeRabbit reported 9 actionable comments and moderate merge risk, including mutable setup-go action, checkout credentials, Windows test paths, and sandbox Git environment gaps.; Author evidence is local-only and states CI is waiting for workflow approval.; PR alone reports one test failure owned by companion PR fix(skills/review): dedup drive-less URLs and sanitize scoped git config #540.; No linked issue evidence ties this to a tracked candidate root.
• Rationale: The supplied PR metadata gives a concrete mergeability blocker (CONFLICTING/DIRTY) on the frozen main, and the change is not merge-ready because automated review findings and missing CI evidence remain. Rebase/conflict resolution is the minimal next step before verification can be interpreted.
• Uncertainty: No CI job states or logs are included, so the actual GitHub Actions outcome for Ubuntu/Windows and whether CodeRabbit actionable findings have been fixed since the submitted review are unverified.
• Exact next action: Rebase feat/harness-audit-test-infra onto main at 6c391ad, resolve conflicts, address the CodeRabbit actionable findings, then request/re-run CI and collect Ubuntu plus Windows verification results.

PR #540 · retain-and-verify

• Title: fix(skills/review): dedup drive-less URLs and sanitize scoped git config
• Draft: true
• Linked issues: none explicit
• Affected roots: none proven
• Status: draft; mergeability CONFLICTING/DIRTY; CodeRabbit skipped review because draft; review_decision empty; no CI result supplied; stacked on test/ci: harden Windows test baseline and enable Windows CI #539.
• Risks: Draft status means no automated review coverage and should not be treated as ready.; Merge conflicts remain unresolved.; The PR contains test/ci: harden Windows test baseline and enable Windows CI #539 commits and therefore inherits test/ci: harden Windows test baseline and enable Windows CI #539's pending verification.; Changed-file list includes many malformed OpenSpec archive paths, indicating possible accidental generated files or packaging defect.; Scope mixes product fixes, docs, package hygiene, and archival metadata.; No linked issue evidence ties the change to an external tracked candidate root.
• Rationale: The safest evidence-backed disposition is to retain the PR as draft and verify its contents: CodeRabbit explicitly did not review it, it is CONFLICTING, it is stacked on unmerged test/ci: harden Windows test baseline and enable Windows CI #539, and its file list includes anomalous archive paths that require manual inspection before any merge or split decision.
• Uncertainty: It is unclear whether the malformed OpenSpec paths are intentional repository content, a tooling artifact, or evidence that the diff generation script failed; the actual post-test/ci: harden Windows test baseline and enable Windows CI #539 diff size is also unknown.
• Exact next action: Keep fix(skills/review): dedup drive-less URLs and sanitize scoped git config #540 draft and audit its 196 changed paths for malformed OpenSpec archive entries before deciding whether to remove them, rebase, or split the scope.

D-001 · unresolved

• Question: Does the root cause of background-subagent capability require the reported prerequisites (first supported runtime, minimum adapter version, selection mechanism, task mode, worktree preflight), or is the current ready|absent capability sufficient for initial deployment?
• Roots: root-376
• PRs: none
• Options: Approve prerequisites as reported and require maintainer-specified first runtime and adapter version; Reduce prerequisites and allow deployment with current ready|absent capability; Close as deferred until prerequisites are defined elsewhere
• Recommendation: Reduce prerequisites until a concrete runtime boundary is identified
• Rationale: The evidence shows current main only reports ready|absent with no versioned adapter registry; blocking on unverified prerequisites delays deliverable work without proven benefit.
• Exact next action: Maintainer specifies the minimal required runtime and adapter version, or approves current ready|absent as sufficient for initial deployment

D-002 · unresolved

• Question: Should the inline fast-path classification for non-SDD verification commands use command presence as an unconditional trigger (as assets/orchestrator.md and assets/orchestrator-delegation.md define), or require a more discriminating predicate to avoid misclassifying small tasks?
• Roots: root-463
• PRs: none
• Options: Retain unconditional command-presence trigger as documented; Replace with a discriminating predicate (e.g., task-size or scope heuristic); Document the tradeoff and leave current behavior as-is
• Recommendation: Replace with a discriminating predicate that respects the 400-line budget and avoids unnecessary delegation
• Rationale: Unconditional command presence causes unnecessary delegation and budget consumption for small tasks; a size or scope-based predicate would preserve intent while reducing overhead.
• Exact next action: Maintainer approves the replacement predicate criteria before any orchestrator or test edit

D-003 · unresolved

• Question: What are the four approval-blocking decisions and the one enable/disable barrier policy that must be frozen in root-419 before the task-activity subsystem can proceed (closed-code vocabulary, cancellation cut-off, shutdown bound, concurrency limit, and the enable/disable policy)?
• Roots: root-419
• PRs: none
• Options: Approve specific values for each of the four design decisions plus one barrier policy; Defer all four plus the policy to a separate design phase; Close the issue and defer to a later design effort
• Recommendation: Defer to a separate design phase that produces a frozen specification before any implementation
• Rationale: The root cause explicitly states that state, cancellation, and final result are handled per-execution-shape rather than through a single authority; implementing without frozen decisions risks rework.
• Exact next action: Maintainer provides or approves the four decision values and the barrier policy, or authorizes a separate design phase

D-004 · unresolved

• Question: Should PR 39's investigation be abandoned due to staleness, or does post-v0.6.0 harness architecture still make its findings relevant?
• Roots: root-39
• PRs: feat(harness): strengthen harness persistence using Pi's context and compaction events #39, fix(harness): refresh orchestrator prompt from disk on each access #51, feat(harness): re-anchor discipline via context event on every LLM call #52, feat(harness): re-anchor after session compaction #53, feat(harness): apply slim rules layer to named and SDD agents #54
• Options: Abandon PR 39 and its follow-up PRs as superseded by post-v0.6.0 architecture; Continue and merge pending PRs if findings remain valid; Close PRs and open a new investigation
• Recommendation: Continue only if a maintainer confirms the findings remain relevant to current architecture
• Rationale: The investigation targets pre-v0.6.0 mechanisms that may have been replaced; merging stale findings introduces technical debt.
• Exact next action: Maintainer confirms or rejects the relevance of PR 39's findings to current main architecture

D-005 · unresolved

• Question: Should stale PRs 195, 281, 539 be closed due to lack of contributor engagement (14+ days since last activity), or should they be pinged and preserved?
• Roots: none
• PRs: fix: EACCES on Android external storage during skill registry refresh (#1291) #195, feat(skill-registry): index npm-package skills (node_modules/*/skills) #281, test/ci: harden Windows test baseline and enable Windows CI #539
• Options: Close as stale/abandoned; Ping contributors and preserve for response; Close and offer to reopen if contributor responds
• Recommendation: Close as stale/abandoned with a comment explaining the inactivity window
• Rationale: All three PRs show no contributor activity after maintainer feedback or last update, and the code context may have diverged significantly.
• Exact next action: Maintainer closes PRs 195, 281, and 539 with a stale/abandoned comment, or explicitly preserves them

D-006 · unresolved

• Question: Should the root-370 fix (relay/handshake bug) be implemented as a single PR, or split given the suspected 400-line budget risk and the complexity of diffing relay/handshake code between versions?
• Roots: root-370
• PRs: fix(review): detect a linked worktree missing its project-local .pi config #371
• Options: Implement as a single PR with an explicit size:exception; Split into separate PRs by component (relay, handshake, config); Close and defer to a later design phase
• Recommendation: Split into separate PRs by component to stay within the 400-line budget
• Rationale: Root-370 requires inspecting multiple files (review-host-relay.ts, review-integration-v2.ts, native-review-cli.ts, config loader) and diffing between versions, which is likely to exceed the budget in a single candidate.
• Exact next action: Maintainer approves the split approach or grants size:exception for a single-PR implementation

D-007 · unresolved

• Question: For PRs 420, 426, 435, 439, 457: should maintainers rebase and attempt resolution of each conflicted PR, or close them and require fresh implementations given the divergence?
• Roots: root-379, root-107, root-432, root-436, root-455
• PRs: fix(installer): give the Windows source build its own timeout and report timeouts precisely #420, fix(quiet-tools): preserve Pi's configured Bash shellPath on Windows #426, feat(agents): add owned activity projection contract #435, refactor(agents): isolate delegation admission seam #439, fix(review): preserve consent binding continuity #457
• Options: Rebase and resolve conflicts for all five PRs; Close the five PRs and require fresh implementations; Close conflicted PRs and manually port the accepted fixes
• Recommendation: Close conflicted PRs and require fresh implementations
• Rationale: All five PRs show unverified conflict states and diverged heads; rebasing without confirmed CI status risks introducing regressions from the intervening commits.
• Exact next action: Maintainer chooses between the three options and confirms CI strategy for any fresh implementation

D-008 · unresolved

• Question: Should PR 387 be merged (deferring to fix(orchestrator): prevent unauthorized execution-surface escalation #379's selected-authority seam) or closed, given that fix(orchestrator): prevent unauthorized execution-surface escalation #379's state and the selected-authority seam ownership are unknown?
• Roots: root-379
• PRs: fix(orchestrator): contain execution-surface escalation #387
• Options: Merge PR 387 after confirming fix(orchestrator): prevent unauthorized execution-surface escalation #379 still exists; Close PR 387 as superseded by another branch/PR; Merge with a placeholder and complete the seam later
• Recommendation: Close PR 387 as superseded unless fix(orchestrator): prevent unauthorized execution-surface escalation #379's seam is explicitly confirmed as still required
• Rationale: The PR defers to a seam whose ownership and existence are unverified; merging unverified dependencies creates integration risk.
• Exact next action: Maintainer confirms fix(orchestrator): prevent unauthorized execution-surface escalation #379's selected-authority seam still exists and is owned by this PR, or closes PR 387

D-009 · unresolved

• Question: Should PR 504's duplicate declaration be confirmed as a real bug before merging, or is it a diff-context misplacement that should be resolved during rebase?
• Roots: root-460
• PRs: fix: clarify guarded command confirmations #504
• Options: Merge after confirming the duplicate is a real bug; Rebase and resolve the potential duplicate as a non-issue; Close PR 504 if the duplicate is confirmed as a misplacement
• Recommendation: Rebase and resolve as a non-issue unless CodeRabbit's observation is confirmed on the rebased head
• Rationale: Without inspecting the conflict surface, it is unclear whether the duplicate is real or a diff-context artifact; resolution should happen during rebase verification.
• Exact next action: Maintainer rebases PR 504, verifies the duplicate declaration, and merges or closes accordingly

D-010 · unresolved

• Question: Should the five Open Decisions in root-405 (Bash tool_call protection design) be put to maintainer vote before any implementation branch is created?
• Roots: root-405
• PRs: none
• Options: Put all five decisions to maintainer vote; Approve the design as-is and proceed to implementation; Defer to a separate design phase
• Recommendation: Put the five decisions to maintainer vote
• Rationale: The root cause explicitly states the protection cannot be achieved by extending existing matchers due to regex-only limitations, and the design requires explicit maintainer decisions on boundaries.
• Exact next action: Maintainer votes on the five Open Decisions documented in root-405 before any implementation branch exists

U-001

• Issues: fix(codegraph): add injectable shell-free runner #163
• Roots: root-160
• Description: PR 163's mergeable status is CONFLICTING due to main advancing after maintainer approval; it is unknown whether the conflict can be resolved without introducing regressions or whether the PR's foundational changes have been partially superseded.
• Evidence gap: No rebase attempt, conflict resolution, or post-approval CI verification exists for PR 163; the conflict hunk content between main's structural additions and PR 163's changes is unexamined.
• Exact next action: Obtain the actual diff between PR 163 head and current main to determine the conflict scope and assess whether resolution is straightforward

U-002

• Issues: fix(harness): refresh orchestrator prompt from disk on each access #51
• Roots: root-39
• Description: It is unknown whether PR 39's investigation has been superseded by post-v0.6.0 harness architecture changes, making its findings potentially stale or irrelevant.
• Evidence gap: No evidence of architectural changes between the PR investigation time and current main; no maintainer decision on the relevance of PR 39's findings exists.
• Exact next action: Compare PR 39's investigated mechanisms against current main's agent-loop or orchestrator implementation to confirm relevance or obsolescence

U-003

• Issues: fix(review): detect a linked worktree missing its project-local .pi config #371, fix(installer): give the Windows source build its own timeout and report timeouts precisely #420
• Roots: root-370, root-379
• Description: CI status is unavailable for PRs 371, 420, and 539, preventing verification of whether the changes pass existing tests and whether the PRs are viable for merge.
• Evidence gap: No GitHub Actions job states, logs, or check runs are included in the survey evidence for these PRs.
• Exact next action: Query GitHub API for the latest CI job states and check runs for PRs 371, 420, and 539 on their current heads

U-004

• Issues: fix(review): surface pre-pr binding failures #202, feat(skill-registry): index npm-package skills (node_modules/*/skills) #281
• Roots: none
• Description: Multiple related PRs (for fix(review): surface pre-pr binding failures #202: fix(review): unify worktree and native authority #132, fix(review): preserva diagnósticos ante módulos duplicados #151, feat(review): adopt review integration v1 end to end #164; for feat(skill-registry): index npm-package skills (node_modules/*/skills) #281: feat(skill-registry): index pi.skills project-declared skill dirs #285, fix(skill-registry): mirror Pi-resolved loaded skills #369) may have partially addressed the same diagnostic areas, making it unclear whether these PRs are still needed or are redundant.
• Evidence gap: No cross-PR analysis exists to determine which related PRs landed, what they changed, and whether the reported issues are now satisfied.
• Exact next action: Diff the changes in the related PRs against the issue requirements to determine redundancy or partial coverage

U-005

• Issues: test(agents): assert per-JD-agent model profiles can diverge #351, feat(runtime): add authoritative managed-task registry #441, feat(runtime): add isolated Pi AgentSession adapter #445
• Roots: root-440, root-444
• Description: It is unknown what check is producing UNSTABLE status on PR 351, whether branch protection requires human approval for PR 441, and whether the SUCCESS verify entries on PR 445 actually bound to the correct commit.
• Evidence gap: No check run details, branch-protection configuration, or commit-verification evidence exists for these PRs.
• Exact next action: Query GitHub for the specific check run statuses and branch-protection rules that apply to PRs 351, 441, and 445

U-006

• Issues: fix: clarify guarded command confirmations #504
• Roots: root-460
• Description: It is unclear whether the duplicate declaration reported at CodeRabbit line 747 is a real bug or a diff-context misplacement that will resolve during rebase.
• Evidence gap: No inspection of the rebased head or the actual declaration scope exists; the conflict surface area is unexamined.
• Exact next action: Rebase PR 504 onto current main and inspect line 747 (or equivalent) on the rebased head to confirm whether the duplicate persists

U-007

• Issues: feat(agents): execute managed tasks in linked worktrees #376
• Roots: root-376
• Description: The prerequisite state reported for background-subagent capability (task mode, worktree preflight, versioned adapter registry, selection mechanism) is asserted but not verified on current main.
• Evidence gap: No inventory of current main's background-subagent capability surface exists; the ready|absent claim is not confirmed against running code.
• Exact next action: Run the capability probe on current main and record whether task mode, worktree preflight, and adapter registry surfaces exist

U-008

• Issues: fix(sdd): remove provider-specific model pin from sdd-init #427, bug(release): latest v2.2.0 still strands corrected lineages before last-event migration #464
• Roots: root-216, root-464
• Description: It is unknown whether repository policy requires completed CodeRabbit review or human approval before merging PR 427 (a migration-related fix), and whether the Go binary review_facade.go correction preflight validation is present in the shipped v2.4.0 binary versus main.
• Evidence gap: No repository branch-protection configuration or maintainer policy documentation is included; no diff of review_facade.go between shipped v2.4.0 and main is available.
• Exact next action: Check repository branch-protection rules for PR 427 and diff review_facade.go between the shipped binary and main to confirm preflight validation presence

U-009

• Issues: review: a linked worktree loses the project-local .pi configuration, so review fails claiming pi is an unsupported runtime #370
• Roots: root-370
• Description: The exact scope and 400-line budget feasibility of the relay/handshake fix for root-370 is unknown; the required file inspection (review-host-relay.ts, review-integration-v2.ts, native-review-cli.ts, config loader) and inter-version diff may exceed budget.
• Evidence gap: No line-count estimate exists for the required changes across the multiple files and versions that must be diffed and reconciled.
• Exact next action: Perform the inter-version diff and line-count estimate for the relay/handshake fix before deciding on single-PR vs split-scope approach

U-010

• Issues: fix(startup-banner): eliminate flicker, freeze, and reload-damaged intro #355
• Roots: none
• Description: The Ghostty-reported symptoms on PR 355 have no tracked issue reference, and the reported 204/204 local suite result cannot be verified against any CI run.
• Evidence gap: No CI check runs, Ghostty issue references, or verifiable CI logs exist for PR 355.
• Exact next action: Search for a Ghostty-reported issue and verify the 204/204 local suite result against a reproducible CI run

[barbatdev]

Verified inventory delta — 2026-09-09

Baseline: strict-v1-88113a3da8ce1ac3212facb363d91144184de689f094ec1ea8c249dca3045472 (cutoff 2026-09-01T13:25:12.459361Z).
Observation window: 2026-09-09T17:27:32.833783Z–2026-09-09T17:27:51.571110Z; main cff894241f385ca1b8197a36872f4285548c02b4.
Two complete paginated open-inventory reads and main SHA agreed. This is an observed snapshot, not a transactional GitHub snapshot.

Kind	Baseline	Still open	No longer open	New to baseline, currently open	Current open
issue	85	68	17	92	160
pr	32	29	3	14	43

117 → 203 open records (+86; +73.5%). Includes tracker #545 as a newcomer. Excluding this tracker: 159 issues + 43 PRs = 202 records.

ISSUE sets

retained: #37, #39, #55, #57, #58, #59, #62, #63, #67, #83, #107, #116, #125, #213, #216, #234, #238, #277, #298, #299, #300, #304, #327, #347, #348, #354, #360, #369, #372, #373, #375, #376, #379, #381, #382, #383, #384, #391, #395, #405, #414, #419, #430, #432, #436, #438, #440, #442, #443, #444, #446, #447, #448, #449, #460, #463, #464, #469, #471, #476, #479, #497, #498, #500, #510, #511, #531, #544

new_open: #545, #546, #548, #551, #554, #555, #573, #577, #578, #579, #580, #581, #582, #586, #593, #596, #597, #600, #608, #614, #617, #619, #626, #636, #638, #644, #653, #654, #656, #670, #674, #675, #680, #681, #685, #686, #687, #688, #690, #693, #694, #695, #696, #700, #701, #706, #707, #713, #714, #715, #716, #718, #719, #721, #722, #723, #724, #727, #728, #730, #731, #732, #734, #736, #737, #740, #741, #742, #744, #745, #746, #748, #749, #750, #751, #752, #753, #754, #757, #758, #759, #760, #761, #763, #764, #765, #767, #769, #775, #784, #789, #791

no_longer_open: #160, #185, #203, #215, #230, #284, #314, #323, #343, #370, #455, #466, #503, #509, #525, #533, #543

PR sets

retained: #36, #51, #52, #53, #54, #68, #195, #202, #281, #351, #355, #371, #387, #420, #426, #427, #428, #429, #435, #439, #441, #445, #457, #501, #504, #512, #532, #539, #540

new_open: #564, #571, #572, #591, #592, #605, #632, #671, #711, #717, #743, #766, #768, #790

no_longer_open: #163, #481, #505

Limits and invariants

• Baseline = retained + no longer open; current open = retained + new open. Partitions are disjoint and matched by GitHub node ID, with issue/PR types kept separate.
• Newcomers means absent from the baseline and open at capture; it does not imply a creation date and may include reopened items.
• Items created and closed between snapshots are outside scope. This is not a complete intervening activity or merge history.
• Closure timestamps and GitHub state reasons are metadata, not proof of causal resolution. PR fix(codegraph): add injectable shell-free runner #163 and docs: proponer un modelo resiliente de supervisión watchdog y recuperación #481 are closed without merge; baseline PR fix(review): match the native eight-line abandon authorization binding #505 is merged.
• The original 84-root analysis and D-001–D-010 remain historical and are not re-adjudicated. No current root count, priority, conflict, supersession or resolved decision is inferred.
• The original issue body and five baseline comments are preserved. This comment is an append-only inventory delta, not a replacement machine state.

Machine-readable observation

Payload SHA-256: 51b740a22b2868c1946f75a66ed04861660f7b424fd372226af3eaa818f1c191. Canonical preimage: UTF-8 JSON with sorted object keys, compact separators (",", ":"), Unicode preserved, no trailing newline. Lists retain the order below.

{"base_snapshot_id":"strict-v1-88113a3da8ce1ac3212facb363d91144184de689f094ec1ea8c249dca3045472","baseline_closed_records":[{"closed_at":"2026-09-09T00:18:46Z","kind":"issue","merged_at":null,"node_id":"I_kwDOSZWLmc8AAAABI_I_UQ","number":160,"state":"closed","state_reason":"completed"},{"closed_at":"2026-09-06T09:39:37Z","kind":"issue","merged_at":null,"node_id":"I_kwDOSZWLmc8AAAABJWKeGA","number":185,"state":"closed","state_reason":"completed"},{"closed_at":"2026-09-07T05:48:54Z","kind":"issue","merged_at":null,"node_id":"I_kwDOSZWLmc8AAAABJfkX2Q","number":203,"state":"closed","state_reason":"completed"},{"closed_at":"2026-09-05T20:38:56Z","kind":"issue","merged_at":null,"node_id":"I_kwDOSZWLmc8AAAABJmUdJg","number":215,"state":"closed","state_reason":"completed"},{"closed_at":"2026-09-06T07:45:34Z","kind":"issue","merged_at":null,"node_id":"I_kwDOSZWLmc8AAAABKKIKYA","number":230,"state":"closed","state_reason":"completed"},{"closed_at":"2026-09-07T06:12:44Z","kind":"issue","merged_at":null,"node_id":"I_kwDOSZWLmc8AAAABLn6Kfg","number":284,"state":"closed","state_reason":"completed"},{"closed_at":"2026-09-05T21:00:32Z","kind":"issue","merged_at":null,"node_id":"I_kwDOSZWLmc8AAAABMzh3kg","number":314,"state":"closed","state_reason":"completed"},{"closed_at":"2026-09-05T21:38:50Z","kind":"issue","merged_at":null,"node_id":"I_kwDOSZWLmc8AAAABM2kMQg","number":323,"state":"closed","state_reason":"completed"},{"closed_at":"2026-09-07T16:13:49Z","kind":"issue","merged_at":null,"node_id":"I_kwDOSZWLmc8AAAABM-O0Ow","number":343,"state":"closed","state_reason":"completed"},{"closed_at":"2026-09-06T08:35:38Z","kind":"issue","merged_at":null,"node_id":"I_kwDOSZWLmc8AAAABNNwyQg","number":370,"state":"closed","state_reason":"completed"},{"closed_at":"2026-09-05T21:38:50Z","kind":"issue","merged_at":null,"node_id":"I_kwDOSZWLmc8AAAABOiaHmg","number":455,"state":"closed","state_reason":"completed"},{"closed_at":"2026-09-07T06:27:47Z","kind":"issue","merged_at":null,"node_id":"I_kwDOSZWLmc8AAAABOmo8oQ","number":466,"state":"closed","state_reason":"completed"},{"closed_at":"2026-09-05T19:12:45Z","kind":"issue","merged_at":null,"node_id":"I_kwDOSZWLmc8AAAABO2xXtQ","number":503,"state":"closed","state_reason":"completed"},{"closed_at":"2026-09-03T22:22:30Z","kind":"issue","merged_at":null,"node_id":"I_kwDOSZWLmc8AAAABO4j3Xw","number":509,"state":"closed","state_reason":"completed"},{"closed_at":"2026-09-02T01:48:59Z","kind":"issue","merged_at":null,"node_id":"I_kwDOSZWLmc8AAAABO9ePRQ","number":525,"state":"closed","state_reason":"completed"},{"closed_at":"2026-09-02T09:13:26Z","kind":"issue","merged_at":null,"node_id":"I_kwDOSZWLmc8AAAABPBprJg","number":533,"state":"closed","state_reason":"completed"},{"closed_at":"2026-09-05T22:10:27Z","kind":"issue","merged_at":null,"node_id":"I_kwDOSZWLmc8AAAABPJSpVg","number":543,"state":"closed","state_reason":"completed"},{"closed_at":"2026-09-09T00:18:14Z","kind":"pr","merged_at":null,"node_id":"PR_kwDOSZWLmc7yrcBX","number":163,"state":"closed","state_reason":null},{"closed_at":"2026-09-01T22:00:11Z","kind":"pr","merged_at":null,"node_id":"PR_kwDOSZWLmc8AAAABBdBCKA","number":481,"state":"closed","state_reason":null},{"closed_at":"2026-09-07T06:27:46Z","kind":"pr","merged_at":"2026-09-07T06:27:46Z","node_id":"PR_kwDOSZWLmc8AAAABBgtFcw","number":505,"state":"closed","state_reason":null}],"baseline_cutoff":"2026-09-01T13:25:12.459361Z","capture_end":"2026-09-09T17:27:51.571110Z","capture_start":"2026-09-09T17:27:32.833783Z","causal_reaudit":false,"main_sha":"cff894241f385ca1b8197a36872f4285548c02b4","open_records":[{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABEK6DAQ","number":37,"updated_at":"2026-06-02T20:45:12Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABEkiSxg","number":39,"updated_at":"2026-08-15T14:32:19Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABFsyBrA","number":55,"updated_at":"2026-06-16T20:02:43Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABFtm8uA","number":57,"updated_at":"2026-06-16T22:19:36Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABF0H-vA","number":58,"updated_at":"2026-09-07T14:19:41Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABHCKLUw","number":59,"updated_at":"2026-06-29T09:51:13Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABHcvK9A","number":62,"updated_at":"2026-08-25T22:04:10Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABH-oYQg","number":63,"updated_at":"2026-07-07T16:31:55Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABIHavNA","number":67,"updated_at":"2026-08-15T01:57:00Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABIUKebA","number":83,"updated_at":"2026-07-10T07:26:23Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABIfiE5A","number":107,"updated_at":"2026-09-03T19:59:40Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABIiVZTA","number":116,"updated_at":"2026-07-12T16:33:40Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABIi80Tg","number":125,"updated_at":"2026-07-13T19:05:10Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABJjgFwg","number":213,"updated_at":"2026-09-06T08:45:04Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABJofE-Q","number":216,"updated_at":"2026-08-22T04:40:44Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABKts13w","number":234,"updated_at":"2026-07-29T22:07:47Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABKvuh1w","number":238,"updated_at":"2026-09-08T18:09:12Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABLGJyXQ","number":277,"updated_at":"2026-08-01T23:53:07Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABMSDDsw","number":298,"updated_at":"2026-08-29T20:42:10Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABMSDc7g","number":299,"updated_at":"2026-09-02T12:46:37Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABMSDrsw","number":300,"updated_at":"2026-08-25T21:52:19Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABMfAhPQ","number":304,"updated_at":"2026-09-02T13:41:51Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABM5Gxsw","number":327,"updated_at":"2026-08-15T15:56:55Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABM_gHOA","number":347,"updated_at":"2026-08-24T20:45:40Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABNECviQ","number":348,"updated_at":"2026-09-04T14:02:47Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABNGNA3w","number":354,"updated_at":"2026-08-17T17:41:27Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABNHk1iA","number":360,"updated_at":"2026-08-22T04:32:03Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABNNZn8Q","number":369,"updated_at":"2026-08-20T22:15:33Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABNOF47A","number":372,"updated_at":"2026-09-05T02:32:46Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABNQA6SQ","number":373,"updated_at":"2026-08-18T16:59:25Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABNQtMtA","number":375,"updated_at":"2026-08-18T18:23:12Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABNR8nWg","number":376,"updated_at":"2026-08-26T20:48:26Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABNTO0SA","number":379,"updated_at":"2026-08-22T04:45:32Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABNnh2Pg","number":381,"updated_at":"2026-08-24T18:33:07Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABNntsQQ","number":382,"updated_at":"2026-08-24T18:33:15Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABNnttpA","number":383,"updated_at":"2026-08-24T18:33:43Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABNntvog","number":384,"updated_at":"2026-08-24T18:34:02Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABNxJIYg","number":391,"updated_at":"2026-08-22T11:57:10Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABNx_Pzg","number":395,"updated_at":"2026-08-22T11:57:08Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABN5ip-g","number":405,"updated_at":"2026-09-08T23:48:10Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABOCiKfg","number":414,"updated_at":"2026-08-24T16:35:50Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABOFczpQ","number":419,"updated_at":"2026-08-27T14:37:18Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABOQX9Bg","number":430,"updated_at":"2026-08-26T06:01:12Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABOSsFlA","number":432,"updated_at":"2026-08-26T14:18:12Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABOXx4TA","number":436,"updated_at":"2026-08-26T17:24:01Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABOYI6WQ","number":438,"updated_at":"2026-08-26T18:28:41Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABOY1zGQ","number":440,"updated_at":"2026-08-26T19:06:06Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABOZmN9g","number":442,"updated_at":"2026-08-26T20:48:46Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABOZm4Lw","number":443,"updated_at":"2026-08-31T13:17:16Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABOZ8yFQ","number":444,"updated_at":"2026-08-26T21:41:10Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABOaiDYw","number":446,"updated_at":"2026-08-27T03:54:16Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABOa83CA","number":447,"updated_at":"2026-08-27T00:50:14Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABObV7fw","number":448,"updated_at":"2026-08-27T03:22:28Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABOb2WMA","number":449,"updated_at":"2026-08-27T03:49:30Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABOjt1zQ","number":460,"updated_at":"2026-08-29T21:01:25Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABOlKXwQ","number":463,"updated_at":"2026-08-31T13:17:17Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABOmRlbQ","number":464,"updated_at":"2026-09-05T19:13:12Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABOrKjuA","number":469,"updated_at":"2026-08-28T15:41:00Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABOuLNKA","number":471,"updated_at":"2026-09-01T18:22:04Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABOxQeIQ","number":476,"updated_at":"2026-08-29T10:59:48Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABOx-Kjw","number":479,"updated_at":"2026-08-29T13:47:27Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABO0-0JQ","number":497,"updated_at":"2026-08-30T02:53:59Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABO1tA-Q","number":498,"updated_at":"2026-08-30T07:27:25Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABO17GxQ","number":500,"updated_at":"2026-09-06T07:56:17Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABO4zQ-A","number":510,"updated_at":"2026-08-30T19:52:04Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABO5qH8Q","number":511,"updated_at":"2026-08-30T23:29:33Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABPBDElw","number":531,"updated_at":"2026-09-03T17:25:09Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABPJSqEg","number":544,"updated_at":"2026-09-01T18:26:13Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABPLeArw","number":545,"updated_at":"2026-09-07T12:50:01Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABPMYcYg","number":546,"updated_at":"2026-09-09T15:54:02Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABPMiyRg","number":548,"updated_at":"2026-09-01T17:19:59Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABPNJwVw","number":551,"updated_at":"2026-09-01T18:38:05Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABPPgAXQ","number":554,"updated_at":"2026-09-01T23:22:00Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABPS1rog","number":555,"updated_at":"2026-09-02T07:59:22Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABPZrzGw","number":573,"updated_at":"2026-09-02T19:51:53Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABPbSoCw","number":577,"updated_at":"2026-09-07T16:54:13Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABPbqxxQ","number":578,"updated_at":"2026-09-03T00:11:17Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABPcONEg","number":579,"updated_at":"2026-09-03T01:46:53Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABPcONqg","number":580,"updated_at":"2026-09-03T01:46:55Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABPcON_w","number":581,"updated_at":"2026-09-03T01:46:56Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABPc2r1g","number":582,"updated_at":"2026-09-03T03:28:42Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABPf307g","number":586,"updated_at":"2026-09-03T09:49:30Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABPu3hJQ","number":593,"updated_at":"2026-09-04T16:46:52Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABPy6SBQ","number":596,"updated_at":"2026-09-05T02:25:21Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABPz1p6A","number":597,"updated_at":"2026-09-05T05:49:56Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABP1gXIQ","number":600,"updated_at":"2026-09-05T11:47:00Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABP2p2WQ","number":608,"updated_at":"2026-09-05T15:20:04Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABP4GXVA","number":614,"updated_at":"2026-09-05T19:30:31Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABP4P7Aw","number":617,"updated_at":"2026-09-05T20:01:14Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABP4Vzlg","number":619,"updated_at":"2026-09-05T20:19:23Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABP4453A","number":626,"updated_at":"2026-09-05T22:36:07Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABP6yjjw","number":636,"updated_at":"2026-09-06T06:11:10Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABP7fvTQ","number":638,"updated_at":"2026-09-08T22:49:12Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABP9t6UQ","number":644,"updated_at":"2026-09-06T16:46:10Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABP_1_yw","number":653,"updated_at":"2026-09-07T00:11:02Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQACq0g","number":654,"updated_at":"2026-09-08T11:30:40Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQA43dA","number":656,"updated_at":"2026-09-07T03:23:16Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQHi7zg","number":670,"updated_at":"2026-09-07T15:34:32Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQIK2eA","number":674,"updated_at":"2026-09-07T16:51:34Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQIOKYw","number":675,"updated_at":"2026-09-07T16:58:30Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQJWAtw","number":680,"updated_at":"2026-09-07T19:49:57Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQJcSIg","number":681,"updated_at":"2026-09-07T20:06:53Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQKAegg","number":685,"updated_at":"2026-09-07T21:33:32Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQKCe8w","number":686,"updated_at":"2026-09-08T01:26:32Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQKI_VQ","number":687,"updated_at":"2026-09-07T21:57:15Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQKfCJQ","number":688,"updated_at":"2026-09-07T22:52:04Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQLDG7w","number":690,"updated_at":"2026-09-08T14:05:08Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQLbZeg","number":693,"updated_at":"2026-09-08T01:18:55Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQLbaDQ","number":694,"updated_at":"2026-09-08T01:18:56Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQLbbYQ","number":695,"updated_at":"2026-09-08T01:18:58Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQLbcYQ","number":696,"updated_at":"2026-09-08T01:19:00Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQLxh8g","number":700,"updated_at":"2026-09-08T02:12:16Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQLxpLQ","number":701,"updated_at":"2026-09-09T05:22:01Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQPjJTQ","number":706,"updated_at":"2026-09-08T16:56:14Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQRMQjg","number":707,"updated_at":"2026-09-08T12:32:41Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQSup0w","number":713,"updated_at":"2026-09-08T14:59:20Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQTB_1w","number":714,"updated_at":"2026-09-08T15:27:20Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQTDtmw","number":715,"updated_at":"2026-09-09T13:35:31Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQTcDng","number":716,"updated_at":"2026-09-08T16:08:10Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQT_FUQ","number":718,"updated_at":"2026-09-08T17:05:39Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQT_GBA","number":719,"updated_at":"2026-09-08T17:05:40Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQUE4KQ","number":721,"updated_at":"2026-09-08T17:16:29Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQUJXpg","number":722,"updated_at":"2026-09-08T17:22:01Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQUKJ_A","number":723,"updated_at":"2026-09-08T17:23:25Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQUlKZQ","number":724,"updated_at":"2026-09-08T18:12:12Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQUzfCg","number":727,"updated_at":"2026-09-08T18:35:25Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQU42eg","number":728,"updated_at":"2026-09-08T18:44:20Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQVB7lw","number":730,"updated_at":"2026-09-09T17:12:57Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQVCOFw","number":731,"updated_at":"2026-09-08T19:00:57Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQVO64A","number":732,"updated_at":"2026-09-08T19:23:58Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQVSTvA","number":734,"updated_at":"2026-09-08T19:32:25Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQVX-fg","number":736,"updated_at":"2026-09-08T19:39:29Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQVZvdg","number":737,"updated_at":"2026-09-08T19:42:45Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQVmtLw","number":740,"updated_at":"2026-09-09T01:24:26Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQVpi3g","number":741,"updated_at":"2026-09-08T20:09:29Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQV2BrA","number":742,"updated_at":"2026-09-09T11:43:59Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQWFcTQ","number":744,"updated_at":"2026-09-08T23:56:04Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQWGGNg","number":745,"updated_at":"2026-09-08T22:42:52Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQWN1ag","number":746,"updated_at":"2026-09-08T21:18:29Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQWhu5g","number":748,"updated_at":"2026-09-08T21:58:27Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQWjg8w","number":749,"updated_at":"2026-09-08T22:02:16Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQWnOyA","number":750,"updated_at":"2026-09-08T22:10:14Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQW2EDQ","number":751,"updated_at":"2026-09-08T22:41:13Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQW5yYA","number":752,"updated_at":"2026-09-08T22:49:29Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQXgLIw","number":753,"updated_at":"2026-09-09T00:23:39Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQXmedQ","number":754,"updated_at":"2026-09-09T00:39:32Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQXtLPQ","number":757,"updated_at":"2026-09-09T00:57:30Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQXupFQ","number":758,"updated_at":"2026-09-09T01:01:29Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQXxBQA","number":759,"updated_at":"2026-09-09T06:39:05Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQYioLg","number":760,"updated_at":"2026-09-09T16:47:38Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQYmbGg","number":761,"updated_at":"2026-09-09T03:19:10Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQZNO0A","number":763,"updated_at":"2026-09-09T04:54:24Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQZNpiA","number":764,"updated_at":"2026-09-09T04:55:22Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQZZFIw","number":765,"updated_at":"2026-09-09T05:30:16Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQZ0q9A","number":767,"updated_at":"2026-09-09T06:38:17Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQa-gEQ","number":769,"updated_at":"2026-09-09T08:27:58Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQbyiYQ","number":775,"updated_at":"2026-09-09T15:15:08Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQc4Q3A","number":784,"updated_at":"2026-09-09T11:48:20Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQend4g","number":789,"updated_at":"2026-09-09T14:31:12Z"},{"kind":"issue","node_id":"I_kwDOSZWLmc8AAAABQe12Qw","number":791,"updated_at":"2026-09-09T14:51:33Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc7hpwDP","number":36,"updated_at":"2026-08-18T18:47:45Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc7nGh01","number":51,"updated_at":"2026-06-16T20:17:15Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc7nGh3_","number":52,"updated_at":"2026-06-16T20:16:07Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc7nGh7c","number":53,"updated_at":"2026-06-16T20:17:03Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc7nGiCL","number":54,"updated_at":"2026-06-16T20:25:24Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc7vY9F3","number":68,"updated_at":"2026-08-16T20:04:39Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc7zqMTk","number":195,"updated_at":"2026-08-18T17:07:07Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc7z-GKj","number":202,"updated_at":"2026-08-17T21:44:58Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc77H4sk","number":281,"updated_at":"2026-08-18T12:12:32Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABAB5dwg","number":351,"updated_at":"2026-09-06T23:06:18Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABADWzdw","number":355,"updated_at":"2026-09-09T12:19:05Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABAJWL1g","number":371,"updated_at":"2026-08-21T19:12:42Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABAmGGwQ","number":387,"updated_at":"2026-09-03T21:54:00Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABA7rzEA","number":420,"updated_at":"2026-08-25T20:43:52Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABA_ngvg","number":426,"updated_at":"2026-08-25T21:37:05Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABA_6gOw","number":427,"updated_at":"2026-08-25T21:37:32Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABBABcpg","number":428,"updated_at":"2026-09-08T15:13:49Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABBAHNfQ","number":429,"updated_at":"2026-09-06T22:44:15Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABBGIdOw","number":435,"updated_at":"2026-08-26T18:37:57Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABBHVO0A","number":439,"updated_at":"2026-08-26T18:35:45Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABBHwZwg","number":441,"updated_at":"2026-08-26T19:42:16Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABBIpDTg","number":445,"updated_at":"2026-08-26T22:36:20Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABBQCYLQ","number":457,"updated_at":"2026-08-27T19:19:21Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABBf0ulA","number":501,"updated_at":"2026-08-30T11:05:26Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABBgZHTQ","number":504,"updated_at":"2026-09-03T21:52:54Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABBiSMQw","number":512,"updated_at":"2026-09-05T13:24:24Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABBofAOQ","number":532,"updated_at":"2026-08-31T18:34:02Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABBq_wiA","number":539,"updated_at":"2026-08-31T23:44:48Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABBq_0SA","number":540,"updated_at":"2026-08-31T23:37:59Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABB5drng","number":564,"updated_at":"2026-09-02T12:50:39Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABB54KiQ","number":571,"updated_at":"2026-09-04T20:08:42Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABB8Xn1w","number":572,"updated_at":"2026-09-02T20:53:51Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABCMQEkA","number":591,"updated_at":"2026-09-04T11:38:45Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABCM06cQ","number":592,"updated_at":"2026-09-04T14:04:02Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABCVHJzg","number":605,"updated_at":"2026-09-07T13:44:10Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABCYG_dg","number":632,"updated_at":"2026-09-06T04:20:09Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABCi_Hvg","number":671,"updated_at":"2026-09-07T15:57:12Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABCsDSIw","number":711,"updated_at":"2026-09-08T14:44:55Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABCtPhHg","number":717,"updated_at":"2026-09-08T17:19:14Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABCu66ww","number":743,"updated_at":"2026-09-08T21:00:20Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABCyBodw","number":766,"updated_at":"2026-09-09T05:58:55Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABCyUS-w","number":768,"updated_at":"2026-09-09T06:46:35Z"},{"kind":"pr","node_id":"PR_kwDOSZWLmc8AAAABC2R0lw","number":790,"updated_at":"2026-09-09T17:22:45Z"}],"partitions":{"issue":{"new_open":[545,546,548,551,554,555,573,577,578,579,580,581,582,586,593,596,597,600,608,614,617,619,626,636,638,644,653,654,656,670,674,675,680,681,685,686,687,688,690,693,694,695,696,700,701,706,707,713,714,715,716,718,719,721,722,723,724,727,728,730,731,732,734,736,737,740,741,742,744,745,746,748,749,750,751,752,753,754,757,758,759,760,761,763,764,765,767,769,775,784,789,791],"no_longer_open":[160,185,203,215,230,284,314,323,343,370,455,466,503,509,525,533,543],"retained":[37,39,55,57,58,59,62,63,67,83,107,116,125,213,216,234,238,277,298,299,300,304,327,347,348,354,360,369,372,373,375,376,379,381,382,383,384,391,395,405,414,419,430,432,436,438,440,442,443,444,446,447,448,449,460,463,464,469,471,476,479,497,498,500,510,511,531,544]},"pr":{"new_open":[564,571,572,591,592,605,632,671,711,717,743,766,768,790],"no_longer_open":[163,481,505],"retained":[36,51,52,53,54,68,195,202,281,351,355,371,387,420,426,427,428,429,435,439,441,445,457,501,504,512,532,539,540]}},"schema":"gentle-pi-backlog-observed-inventory-delta/v1","scope":"open-inventory-only"}