docs: resolve release-plan formatting review notes

devlog/_plan/260906_release_244_followups/000_plan.md

@@ -1,6 +1,7 @@
 # Release 2.44 follow-up integration
 
 ## Loop contract
+
 - Archetype: spec-satisfaction repair; class C4 for governance, replay and release; C3 for bounded client changes.
 - Trigger: owner authorized the named backlog, bottom-up stacked PR integration, --no-verify pushes, admin merges, maintainer dev policy and release on 2026-09-06.
 - Goal: publish the verified next release after these narrowly scoped fixes.
@@ -12,6 +13,7 @@
 - Resources: existing GitHub account, repository and release OIDC only; no new credentials/purchases. Unlimited requested-model delegation within available concurrency; no owner token/cost cap. Each subprocess <=30 minutes, CI polls <=60 seconds, each phase investigation checkpoint at 60 minutes with evidence-based continuation. No implicit exhausted outcome.
 
 ## Snapshot and sequence
+
 Baseline dev: af344a28eabcee09a5e04c48ab897449792719c2, version 2.44.0. Latest published stable is 2.43.0. Refresh before every layer.
 
 | Work phase | Design | Dependency / independent proof |
@@ -32,4 +34,5 @@ Baseline dev: af344a28eabcee09a5e04c48ab897449792719c2, version 2.44.0. Latest p
 One work-phase is one PABCD cycle. Publish short dependency stacks; use merge commits for parents with live children, squash bounded terminal carries if safe, and recascade after any squash. Independent presentation/governance slices remain their own PRs even though execution is sequential. Every carried contributor receives account-linked Co-authored-by credit. Preserve snapshots of source heads.
 
 ## Evidence boundaries
+
 #3735/#3734 are public current-SHA reports; independently inspect code, author local-pass statements remain reports. Kiro proof is recorded-log shape plus synthetic CI tests, never a live quota-consuming request. #3644 has a network A/B report and landed diagnostic #3693; do not claim a Windows runtime reproduction from mocked tests. Detailed private logs are never committed.

devlog/_plan/260906_release_244_followups/010_policy.md

@@ -3,6 +3,7 @@
 Depends on roadmap. Class C4; spec-satisfaction. Owner authorizes maintain/admin integration through PRs without a second maintainer approval, including self-authored PRs. Actual inspected roles for both rostered maintainers are admin; current dev rules already permit role 5 PR bypass. The contradiction is primarily normative documentation, plus future Maintain role coverage.
 
 ## Exact change map
+
 - MODIFY MAINTAINERS.md review policy and dated change log: distinguish contributor approvals from explicitly opted-in maintainer integration to dev. Preserve actual independent technical/security review and CI duties; do not call self-integration a second-person approval. Main/preview promotions retain existing rules.
 - MODIFY AGENTS.md branch/review summary: align with maintainer dev exception; PRs still required, force pushes and deletions still blocked.
 - MODIFY scripts/ci/assert-mergeable-review.sh: parse explicit --maintainer-integration in any argv position, retaining optional repository positional argument. Default strict contributor-review path unchanged. Opt-in skips exactly the reviewDecision=APPROVED and qualified non-self approval checks, not review retrieval, objections or race checks. For override, require baseRefName=dev, current authenticated human actor from gh api user, membership in trusted base dev MAINTAINERS roster, and live maintain/admin role. Preserve complete review parsing, maintainer CHANGES_REQUESTED blocking and final head/base/actor authorization recheck. Print only a truthful validation snapshot with head/base/actor; do not emit a privileged merge recipe because head matching cannot atomically bind the PR base. Never accept a CLI-supplied actor, PR-authored roster, bot or unknown role.
@@ -11,14 +12,18 @@ Depends on roadmap. Class C4; spec-satisfaction. Owner authorizes maintain/admin
 - External UPDATE dev ruleset 20763889 only: add RepositoryRole actor_id=2 bypass_mode=pull_request, preserve actor_id=5 and all conditions/rules. Read snapshot immediately before update; compare after. Verify role names through GraphQL repositoryRoleName: maintain=2, admin=5; role4 is write and must never be added. Do not change main 20764415 or preview 20764486. Rollback is the saved before JSON projected to accepted API fields.
 
 ## Activation matrix and verifier
+
 CI test fixture: authorized admin and maintain actors with no second approval on dev pass ONLY opt-in; write/outsider/bot/missing actor/role API error fail; main/preview/stack base fail; pending maintainer objections, API pagination failures, head/base races fail. Default no flag retains all prior strict failures. shell syntax can be read/checked; Bun tests and typecheck run remotely. Live REST readback proves only dev actor list changed; compare main/preview snapshots unchanged.
 
 ## Trust / bypass record
+
 Assets repository integration history; entry script and authenticated GitHub rules API; boundary contributor metadata versus trusted dev roster/live permissions. E7 human policy plus E8 GitHub branch rules; admin can alter rules outside this helper, so helper is an early review check, not universal enforcement. PR bypass does not remove deletion/non-fast-forward rules outside PRs. Security review recorded independently in scratch; final disposition may be published after diff is public.
 
 
 ## Policy-cycle P refresh and delegation
+
 Current dev remains the roadmap baseline; source helper and ruleset snapshots were reread. The preceding D locked the roadmap and made policy the next cycle. Worker owns only scripts/ci/assert-mergeable-review.sh and tests/ci-workflows/assert-mergeable-review.test.ts; main owns MAINTAINERS.md, AGENTS.md, contributing.md, structure/06_docs-and-release.md and GitHub settings. No overlapping writes or local tests. An independent reviewer audits the final script/docs delta before remote CI and dev-only ruleset application.
 
 ## Dispatch repair / policy P amendment
+
 The first helper+tests worker repeatedly read unrelated plan pages and produced no source delta after a scope correction and bounded waits. It was retired without edits. Main now owns scripts/ci/assert-mergeable-review.sh in addition to documentation/settings; a fresh worker owns ONLY tests/ci-workflows/assert-mergeable-review.test.ts. The protocol remains --maintainer-integration in any argv slot, optional repo, actor from gh api user (login/type), baseRefName from PR metadata, maintain/admin role_name from collaborator permission. Final metadata rechecks head/base/author, then actor/role/roster authorization again. Default strict path does not require new fields. No expected evidence or scope was removed. This replan changes dispatch ownership, not the approved policy.

devlog/_plan/260906_release_244_followups/020_task_input.md

@@ -3,6 +3,7 @@
 Depends on policy; class C4 for protocol admission. Fix public issue #3735, observed on baseline dev. Preserve the existing unpaired-tool HTTP 400 guard from #3471.
 
 ## Diff-level change map
+
 - MODIFY src/responses/parser.ts at function_call_output classification before tool lookup: route only a complete external task-input envelope to an Ocx user message. Eligibility: type function_call_output, no call_id property (including inherited properties for direct helper calls), nonempty string id/name/namespace, nonempty fully representable text/image output. Do not require specific names, prefixes, namespaces or XML content. Existing standard tool results and custom_tool_call_output keep current path.
 - NEW src/responses/task-input.ts: pure recognition returning supported Ocx user content or undefined, no request mutation/network/storage. Reuse existing content converters only when they preserve every accepted output part and reject invalid mixed arrays rather than silently drop them.
 - MODIFY tests/responses/responses-parser.test.ts, tests/responses/responses-compaction-routing.test.ts and tests/responses/openai-responses-passthrough.test.ts with narrow positive/negative fixtures. No new test file or layout registry entry is needed.
@@ -11,20 +12,24 @@ Depends on policy; class C4 for protocol admission. Fix public issue #3735, obse
 Before: result-shaped external task input enters toolResult branch with undefined call id, then translated-adapter guard returns 400. After: the complete external shape enters user message with intact supported text/images; malformed/orphan tool results still fail. No secret or raw logged transcript is copied to tests.
 
 ## Activation / verifier
+
 Remote parser tests exercise arbitrary tool names/namespaces, blank/empty content remains ineligible, multiple ordered text parts and supported images; retain exact content without orphan marker. Explicit call_id empty/null/number/undefined-own-property remain invalid, as do custom outputs missing identity, partial provenance, unsupported/mixed malformed arrays. Existing genuine call ids remain tool results. Remote endpoint/compaction/passthrough fixtures prove unchanged raw body forwarding and guard failures. ci.yml runtime jobs + typecheck/privacy establish fresh proof. Local saved log provides provenance only; no live Kiro request.
 
 ## Boundary / alternatives
+
 No-op leaves current task creation unusable; configuration cannot distinguish this parser envelope; generic orphan-to-user repair would reverse #3471 and is rejected. Reuse current message types; no persisted schema fields. Classification is compatibility handling, not authentication: no privilege is granted by envelope metadata.
 
 
 ## Source follow-up folded at roadmap lock
+
 Author yrlan-montagnier (Yrlan), GitHub id 71253160: preserve Co-authored-by: Yrlan <71253160+yrlan-montagnier@users.noreply.github.com>. Posted helper may manufacture an encrypted-content-omitted marker that makes encrypted-only input look usable; reject encrypted-only and mixed opaque/unsupported input, never use placeholder text as eligibility. Keep every pre-existing #3471 regression, adding tests rather than replacing them. Add tests/responses/responses-compaction-routing.test.ts and tests/responses/openai-responses-passthrough.test.ts to explicit remote verification. Prefer a dedicated small predicate over relocating passthrough helpers unless byte-for-byte behavior is proved.
 
 ## Task-input cycle P refresh at 25c8d2b4e
+
 The preceding D landed policy #3739 and actual Maintain/Admin settings. Issue #3735 is still open and the author has no open PR; retain the account-linked Yrlan trailer. Source parser at lines 150-160 currently recognizes only message/agent_message as the continuation conversation boundary. Compute the optional external content once near effectiveType and include a recognized envelope in that existing boundary predicate. In the function_call_output branch, clear pendingReasoning, emit a user message and continue; leave the ordinary result branch and core guard unchanged.
 
 Concrete new leaf: src/responses/task-input.ts exports externalTaskInputContent(item: unknown): string | OcxContentPart[] | undefined. It imports only type OcxContentPart and existing isObj/inputContentParts. Require exact function_call_output, no call_id property, nonblank id/name/namespace, and a nonblank string or fully supported array. Array parts are input_text/text/output_text with string text or input_image with nonblank string image_url and optional auto/low/high/original detail. Normalize output_text to input_text before calling the existing input converter; original image detail maps to high by that converter. Require at least one nonblank text or usable image. Reject any unsupported/opaque/malformed member, invalid detail or file-id-only reference as a whole; placeholder text never establishes eligibility. Preserve accepted text bytes, order and image references; no raw-body mutation or helper relocation from passthrough.
 
-Field chain: external JSON shape -> pure leaf validation -> parser user message + existing _continuationConversationMessageIndex -> translated adapter's existing user-content serialization. No new persisted field/schema/config. Passthrough and compact use unchanged raw body. Tests include pending reasoning reset and previous_response_id boundary index=0 for a new envelope without a replay prefix, alongside all old #3471 controls.
+Field chain: external JSON shape -> pure leaf validation -> parser user message + existing `_continuationConversationMessageIndex` -> translated adapter's existing user-content serialization. No new persisted field/schema/config. Passthrough and compact use unchanged raw body. Tests include pending reasoning reset and previous_response_id boundary index=0 for a new envelope without a replay prefix, alongside all old #3471 controls.
 
 Dispatch: main owns new leaf, parser, endpoint/passthrough regressions and English/structure docs; a bounded worker owns only tests/responses/responses-parser.test.ts. Independent A/C reviewer reads named leaf/parser boundaries. No local tests/typecheck/build; remote ci.yml runtime/gates and existing parser/compaction/passthrough suites provide proof. Parser leaves add no core/Lab dependency. No-op/configuration cannot fix this shape; existing input converter is reused behind strict validation.

devlog/_plan/260906_release_244_followups/030_kiro_results.md

@@ -3,6 +3,7 @@
 Depends on task-input; class C4 for protocol identity. Fix #3734 from recorded Codex code-mode output shape, never by spending live Kiro quota.
 
 ## Diff-level change map
+
 - MODIFY src/adapters/kiro.ts pushUser/turn-construction helper: when adding results in immediately adjacent parsed messages, tracked separately from collapsed user turns, combine only adjacent results with identical normalized toolUseId. Append content in exact input order and propagate error if any constituent is error. Preserve images via the adapter's supported representation; ensure no image is dropped or reordered relative to supported content semantics.
 - Preserve the pendingToolUses.delete validation: call-a, call-b, call-a remains invalid. Do not globally deduplicate by id or merge across assistant/tool boundaries, intervening ordinary input, or unrelated result.
 - MODIFY tests/providers/kiro/kiro-adapter.test.ts and relevant kiro-images.test.ts fixtures for three adjacent results, error later in group, different ids and nonadjacent repeats, text+image preservation. No new fixture uses real call ids or messages.
@@ -11,16 +12,20 @@ Depends on task-input; class C4 for protocol identity. Fix #3734 from recorded C
 Before: pushUser appends each result, wire validation consumes the first matching toolUseId and rejects the next duplicate. After: consecutive same-call outputs become one ordered result before validation. Opaque encrypted output rejection remains unchanged.
 
 ## Activation / verifier
+
 CI tests feed one assistant exec call followed by notify/notify/final results; assert one toolResult and ordered content. Mixed error/success reduces to error; unrelated result boundaries cannot be crossed. Same-id nonadjacent repeat still throws matching error. Exercise retained images using existing adapter representation; enforce maximum/shape constraints already owned by Kiro wire. Run existing Kiro adapter/image suites through ci.yml, plus full typecheck/privacy. Saved local log shape is supporting evidence only; live Kiro correctness remains untested and explicitly reported.
 
 ## Non-goals
+
 No Kiro account/OAuth/quota changes, no aggressive malformed-history healing, no parser changes beyond prior layer, no global result deduplication.
 
 
 ## Source follow-up folded at roadmap lock
+
 Track adjacency in original message iteration; reset on every non-toolResult message including user/developer/assistant, even if pushUser collapses it into one user turn. Retain Kiro images on the current user image list as the existing wire format requires; do not promise unsupported text/image interleaving in the wire. Preserve Co-authored-by: Yrlan <71253160+yrlan-montagnier@users.noreply.github.com>. Local log metadata contains old Kiro activity and is not a current live reproduction.
 
 ## Kiro-cycle P refresh on parent b24ed35a
+
 Parent #3743 is verified and ready, still open as this branch base; fixture prerequisite #3745 is merged. Issue #3734 remains open without an author PR. kiroPayloadMessages currently returns parsed.context.messages unchanged, so tracking adjacency at the top of its loop observes original Ocx message barriers even when a reasoning-only assistant is later skipped or user/developer turns collapse.
 
 Concrete source edits in src/adapters/kiro.ts only: priorCalls values retain rawId alongside wireName; validate each result against that exact raw id after normalizing for wire lookup. This rejects different raw ids sharing a replacement/truncation result without banning legitimate paired non-wire ids. Track adjacentRawToolResultId, reset it for every non-toolResult before any early continue; for matching adjacent raw id and last user turn/last wire result, append text content and images, set status error if any constituent isError. Otherwise retain pushUser and final conversation validation. No global dedup, cross-turn merge or normalizer change.
@@ -34,6 +39,7 @@ Local evidence limit: saved Kiro conversation data and OCX diagnostic artifacts
 Dispatch: main owns adapter/docs; bounded worker owns only kiro-adapter.test.ts. Independent A/C reviewers inspect raw identity, original-message adjacency, error/image propagation and unchanged encrypted rejection. Full runtime CI is remote only, including existing Kiro image/adapter tests; live Kiro is forbidden.
 
 ## Resumed P after verified guidance parent b7e67d84d
+
 The separate task-guidance cycle is complete, parent3743 P1 is resolved and CI34014313740 is green. Its verified head was merged into this preserved Kiro branch before implementation. Prior Euler review is folded below and must be rechecked before B.
 
 A contiguous group is finalized before any non-toolResult (including skipped reasoning-only assistant), before a different raw id, and after the loop. Track only local bookkeeping: rawId, reference to the fresh KiroToolResult, count, raw text parts and whether this group carried images; never put these fields on wire objects. A single-result group keeps the exact existing normalized text/fallback. For 2+ results, preserve ordered raw text parts except successful empty-exec wrappers, append images and keep any isError sticky. If the whole group has meaningful text, use those parts and remove any first-chunk empty fallback. Preserve whitespace text parts when meaningful text exists. If all text is empty, retain one existing fallback; use the neutral KIRO_EMPTY_TOOL_RESULT_MESSAGE when images or an error flag make an empty-success exec hint inappropriate. Failed exec wrappers are meaningful failure information and remain raw text in multi-result groups even when the incoming isError flag is false; preserve existing FAILED_EXEC_OUTPUT_MESSAGE for a single result. No new normalizer or message template.
@@ -43,5 +49,6 @@ Read evidence: normalizeEmptyExecToolResultText distinguishes EMPTY_EXEC_OUTPUT_
 Additional regressions: later image-only/empty/success-empty wrapper does not inject placeholders into an already-populated result; initial empty then real text removes the empty hint; all-empty groups retain a valid nonblank result; multi-result failed wrapper retains its failure signal; later encrypted adjacent result still rejects; whitespace between meaningful chunks survives. Existing single empty/failed exec normalization tests must pass unchanged.
 
 ## Resumed A dispositions
+
 Accept whitespace concern: collect a nonzero-length raw text part when trim is empty OR the shared normalizer did not classify it as EMPTY_EXEC_OUTPUT_MESSAGE. This preserves whitespace between/before actual text while discarding only true empty-success wrapper text; failed wrappers are never in that drop category. Finalization decides whether the aggregate has meaningful text.
 Rebut the need for duplicated tool-name bookkeeping: create the first fresh wire result using the EXISTING normalizeEmptyExecToolResultText(text,{toolName,toolNamespace}) call before registering the group. A one-result group is never rewritten at finalization, so its exact precomputed fallback is retained; no normalization without identity occurs. Multi-result finalization replaces that initial content only with raw aggregate parts (or neutral empty text for image/error groups). Tests pin the existing single-result behavior and no bookkeeping keys on wire.