CHANGELOG.md

Changelog

All notable changes to the SIN-Code unified binary will be documented in this file.

[v3.28.0] - 2026-06-30

Features

• SIN-Browser-Tools MCP Full Integration (v3.29.0): 106 Playwright-based browser automation tools now fully integrated in sin-code and opencode. Per-tool permission defaults (35 read-only allow, 71 mutating ask). Catalog entry updated with full tool inventory. opencode.json sin-browser-tools MCP server registered. sin-code mcp-install browser updated to install SIN-Browser-Tools (pip) instead of deprecated @anthropic/browser-tools-mcp (npm). Tools include: navigation, click/type/fill, screenshots, PDF, tab/session management, cookie/storage, diagnostics, Shadow DOM, OOPIF, SPA wait, network mocking, macOS Spaces, screen recording.
• skill-process-disk-clean (v3.29.0): New bundled skill that safely reclaims disk space on a developer Mac. Surveys large directories, classifies them by risk (safe / ask / locked), deletes regenerable caches (Yarn, bun, npm, go-build, trivy, claude transcripts/plugins, chrome_pipeline, webauto), and VACUUMs the bloated ~/.local/share/opencode/opencode.db (WAL freelist → 60 GB shrinks to ~6 GB with zero session loss). Sync-compatible with Infra-SIN-OpenCode-Stack. Local symlink ~/.config/opencode/skills/clean-disk-opencode provided.

Features

• Context Meter (#484): ContextMeter struct in agentloop/ — thread-safe token tracker with visual progress bar. Warns at 80%, compacts at 90%. sin-code context CLI command.
• Voice-to-Code (#481): --voice flag on sin-code chat. Records audio via sox/ffmpeg, transcribes via whisper. No CGO (M2). voice/ package.
• Decision Memory (#488): SQLite-backed architectural decision store. Workspace-isolated, full-text search. sin-code decision list|search CLI. decision/ package.
• TUI Collapsible Output (#491): Tool output in TUI collapsed by default (5 lines + hint). Tab to expand/collapse. Glamour syntax highlighting already wired.
• Background Agents (#479): Fire-and-forget async agent jobs. Goroutine-backed, thread-safe Manager. sin-code background run|list|status|result CLI. background/ package.
• Spec-Driven Development (#480): EARS (Easy Approach to Requirements Syntax) parser → deterministic architecture generator → code scaffolder. sin-code spec-driven parse|arch CLI. specdriven/ package.
• Session Sharing (#482): Export sessions as self-contained JSON or HTML files. Import from JSON. sin-code share export|import CLI. sessionshare/ package.
• Plugin System (#489): Unified plugin registry (MCP/skill/tool/hook). Auto-detection from name/source. plugins/ package enhanced with types.go.
• MCP Auto-Install (#490): Discoverable MCP server registry (12 known servers). Install/uninstall to config. sin-code mcp-install discover|install|uninstall CLI. mcpinstall/ package.
• LSP Auto-Configure (#492): Detect LSP servers on PATH (gopls, pyright, tsserver, rust-analyzer, clangd, lua, java). Generate config. sin-code lsp-config detect|config CLI. lspconfig/ package.

CLI

• 106 subcommands (was 99)
• New: context, decision, background, spec-driven, share, mcp-install, lsp-config

Tests

• 43 new tests across 8 new packages (all -race clean)
• Total test count increased

Infrastructure

• Infra opencode.json: 24 baseline skills (was 16)

[v3.26.0] - 2026-06-29

Added — Free web search + YouTube for AI Agents

• sin_web_search chat tool — free DuckDuckGo web search (keyless, no API key needed) + optional Tavily/SerpAPI/Brave multi-provider engine. Wires existing internal/websearch engine as first-class chat tool. Permission: allow (read-only). Catalog entry added.
• YouTube for AI Agents MCP integration — 9 YouTube tools via youtube-for-ai-agents (github.com/JCodesMore/youtube-for-ai-agents). No YouTube Data API key needed — uses youtubei.js InnerTube client.
  • youtube__youtube_search (allow) — search videos/channels/playlists
  • youtube__youtube_get_transcript (allow) — transcripts with timestamp control
  • youtube__youtube_get_video_info (allow) — video metadata (brief/standard/full)
  • youtube__youtube_get_channel_videos (allow) — list channel videos
  • youtube__youtube_get_channel_info (allow) — channel metadata
  • youtube__youtube_get_playlist (allow) — playlist metadata + video list
  • youtube__youtube_download (ask) — download video/audio
  • youtube__youtube_clip (ask) — cut clips by timestamp
  • youtube__youtube_highlight_reel (ask) — merge clips into highlight reel
  • Cookie storage: ~/.config/sin-youtube/cookies.json (chmod 600) for optional age-restricted content
  • Registered in mcpclient config, permission defaults, catalog

Fixed

• DuckDuckGo parser fix — real /ac/ endpoint returns ["query",["sug1","sug2",...]] not [{phrase:"..."},...]. Parser now handles nested list format + legacy object form.
• MCP stdio subprocess context fix — exec.CommandContext(ctx, ...) used per-server timeout context which killed the process after connect. Fixed: use context.Background() for subprocess lifetime; per-server timeout still applies to initialize + ListTools.
• MCP connect timeout — increased 3s → 10s (Node.js MCP servers need more init time).

Refactored — 4 god-files split (no file >1000 lines)

• cmd/sin-code/tui/update.go: 1747 → 371 lines (8 new files)
• cmd/sin-code/tui/chat_render.go: 1107 → 307 lines (6 new files)
• cmd/sin-code/chat_tools_extra.go: 1126 → 192 lines (5 new files)
• cmd/sin-code/internal/agentloop/loop.go: 1009 → 465 lines (5 new files)

[v3.25.0] - 2026-06-25

Added — 4 new subcommands

• sin-code doctor — unified health check (11 checks: Go, config, DBs, MCP, tools, CGO, module-path). doctor_cmd.go.
• sin-code diff — git diff with complexity + sin-debt overlay. diff_cmd.go.
• sin-code benchmark — eval golden dataset runner with scoring report. benchmark_cmd.go.
• sin-code tokens cost — cost projection + budget alerts from token ledger. tokens_cmd.go.
• 4 parallel subagents, 90 new tests.
• CEO Audit: Score 100, A+, 48/48 gates, 0 unapproved findings, 0 rot-risk.

[v3.24.0] - 2026-06-23

Added — TUI/agentloop/headless live progress

• Live tool timing + NDJSON progress + live token/cost footer (issue #424)
• Esc cancels in-flight TUI prompt
• sin-code analyse-image vision model (issue #423)
• Security Scanning V2 (secrets/SAST/SCA/SBOM/container/SARIF)
• Complexity cleanup: 55+ files merged, all analyzers fixed, scoring corrected
• 0 unapproved findings, 0 rot-risk markers
• Test suite made environment-independent

[v3.27.0] - 2026-06-30

Added — vibe-notion MCP bridge (full Notion access)

• vibe-notion MCP bridge — full Notion access (pages, databases, blocks, comments, users, workspaces) via Bridged-External pattern wrapping the vibe-notion npm CLI as a subprocess. 19 MCP tools (10 read auto-allowed, 8 write gated ask, 1 raw escape hatch ask).
  • notion__notion_read_auth_status (allow) — check auth state
  • notion__notion_read_workspaces (allow) — list all workspaces
  • notion__notion_read_resolve (allow) — resolve URL/page-id to workspace-id
  • notion__notion_read_search (allow) — full-text search (requires workspace_id)
  • notion__notion_read_page (allow) — get page metadata + properties
  • notion__notion_read_database_schema (allow) — get database schema/properties
  • notion__notion_read_database_rows (allow) — query database rows
  • notion__notion_read_block_children (allow) — list child blocks of a page/block
  • notion__notion_read_comments (allow) — list comments on a page/block
  • notion__notion_read_me (allow) — current Notion user identity
  • notion__notion_write_create_page (ask) — create a new page
  • notion__notion_write_update_page (ask) — update page title
  • notion__notion_write_archive_page (ask) — archive (soft-delete) a page
  • notion__notion_write_append_block (ask) — append markdown content as blocks
  • notion__notion_write_delete_block (ask) — delete a block
  • notion__notion_write_upload_block (ask) — upload file (image, PDF, etc.) as block
  • notion__notion_write_move_block (ask) — move a block to a new parent
  • notion__notion_write_create_comment (ask) — add a comment
  • notion__notion_raw_cli (ask) — run arbitrary vibe-notion subcommand
• Act-as-user mode via token_v2 from browser/desktop session (default); bot mode via NOTION_TOKEN env var
• Bridge location: ~/skills/vibe-notion-mcp/mcp_server.py (Python venv, MCP 1.28.1)
• Bundled skill: skills/productivity-skills/skill-productivity-notion/ (SKILL.md, LICENSE, context/, frameworks/, tasks/, templates/, references/)
• Registry: notionConfig() + notionMCPPath() in config.go; permission defaults in permission_defaults.go
• Catalog entry: source_external.go — unified tool catalog
• Lifecycle: lifecycle_map.yaml — skill-productivity-notion (external)
• ECOSYSTEM.md: notion row in MCP Skill Servers table
• AGENTS.md: v3.27.0 roadmap, Productivity skill row, notion permission rows, "Last verified" updated
• sin-brain: P0 global rule for Notion capability (all agents know to use notion__notion_read_* tools)
• sin-memory: integration insight stored with tags

[Unreleased] - 2026-06-23

Fixed — safe sin review parser routing

• Matching Python/JavaScript/TypeScript file pairs retain IBD semantic review.
• Markdown, unknown plain-text extensions, extensionless files, and mixed file types now use a deterministic line-diff fallback instead of the Python AST parser, preventing syntax crashes such as unterminated or invalid literals.
• Added focused Markdown, mixed Python/Markdown, Python AST-routing, and CLI regression coverage.

Added — Ecosystem skill diagnostics and install-all improvements

• cmd/sin-code/internal/skillmgr.Doctor — new diagnostic method that checks every known ecosystem skill and reports why it is not runnable (not installed, missing MCP entrypoint, dependency unreachable, or deprecated). Returns SkillStatus for each skill with a populated Detail field.
• sin-code skill doctor — new CLI subcommand that renders the diagnostic report. It prints a table with INSTALLED, RUNNABLE, and DETAIL columns and a summary line. Supports --json for structured output.
• sin-code skill install all --json — the batch ecosystem install command now emits structured JSON output for CI/automation.
• checkSkillStatus helper in internal/skillmgr — shared per-skill state computation used by Status and Doctor, eliminating duplication.
• findPythonCliEntrypoint helper in internal/skillmgr — discovers Python CLI wrapper scripts (e.g. scripts/sin_context_bridge.py) so the diagnostic/install path recognizes the same entrypoints as the MCP registry.
• Ecosystem skills activation (honcho / simone / symfonylens / analyse / contextbridge / grillme) — mcpclient.DefaultServers() and skillmgr now resolve the correct entrypoints for simone (python3 src/cli.py serve-mcp or simone-cli serve-mcp), symfonylens (python3 -m symfony_lens.server or symfony-lens), analyse (SIN-Analyse-Suite exact repo casing), contextbridge (scripts/sin_context_bridge.py serve), and grillme (python3 -m sin_grill_me.mcp_server). The external Honcho server is probed for honcho before reporting it as runnable.
• Tests — 6 new race-clean tests in manager_test.go for Doctor, 5 new tests in commands_test.go for the skill doctor and skill install all cobra surfaces, and --json round-trip tests, plus new tests in manager_test.go and registry_test.go for the simone/symfonylens entrypoints and honcho health check, and for analyse casing, contextbridge CLI entrypoint, and findPythonCliEntrypoint.
• Docs — docs/SKILLS.md updated with the new --json flags.

Added — Shop-Center skill integration (issue #142 fusion)

• KnownSkills() registry extended with the three shop skills (issue #142 acceptance criterion #2 — installable via sin-code skill install <name>):
  • shop-cj-dropshipping → cj-dropshipping-skill
  • shop-stripe → SIN-Stripe-Bundle
  • shop-tiktok → SIN-eCommerce-Scraper-Bundle
• Bundled SKILL.md files (cj-dropshipping, stripe, tiktok) already carry lifecycle: external + sources: frontmatter (added by PR #218 / issue #139). The sources: field points to the canonical external repos so the operator can discover the upstream implementation directly from the bundled skill.
• New test TestKnownSkillsHasShopEntries in cmd/sin-code/internal/skillmgr/manager_test.go asserts the three entries are present with the correct repo mapping.
• Validation passes — validate_skill.py --all-bundled --strict reports 0 failed for the 34 skills (the 3 shop skills were migrated by PR #218).
• Long-term fusion strategy documented in the issue body: phase 1 (external canonical) → phase 2 (bundled doc, done) → phase 3 (native subcommand) → phase 4 (deprecate upstream). Phase 3 is deferred until the shop domain matures.

Added — sin-code grill (issue #141 fusion, native implementation)

• New package cmd/sin-code/internal/grill/ with 4 source files (types.go, catalog.go, manager.go, grill_test.go)
  • 14 race-clean unit tests. The native Go implementation of the external SIN-Code-Grill-Me-Skill Python MCP server (38 KB). Ships in v0 with JSON-file storage; SQLite session storage is a v1 follow-up.
• New subcommand sin-code grill (5 subcommands):
  • grill start <topic> — begin a grilling session, print the id
  • grill next <id> — ask the next adversarial question
  • grill answer <id> <d-id> <text> — record the response (use "done" to resolve a decision)
  • grill status <id> — show resolved + open decisions
  • grill synthesize <id> [--json] — produce a structured summary of decisions, assumptions, and open questions
• Question catalog — 8 seed anti-patterns (Hidden Assumptions, Rollback Plan, Failure Modes, Operator Cost, Premature Optimization, Scope Creep, Single Point of Failure, Verification Gap). Each has 2-3 example questions. The CLI picks one per grill next call (hash-seeded for determinism).
• Storage — $SIN_CODE_HOME/grill/<id>.json, atomic writes (temp + rename). v1 will migrate to SQLite via the existing internal/session/store.
• 14 race-clean tests covering the full flow (Start → Next → Answer → Synthesize → round-trip across restarts).

Added — sin-code goal fusion (issue #140, v0.5)

• Four new subcommands under sin-code goal (issue #140 fusion with the external SIN-Code-Goal-Mode-Skill Python MCP server):
  • goal status <id> — show one goal with subtasks (children)
  • goal complete <id> — mark a goal as verified/done
  • goal subtask <parent-id> <prompt> — add a subtask
  • goal report [--format md|json] — progress report
• Mapping to the 8 external tools (issue body):
  • goal_start → existing goal add
  • goal_status → NEW goal status
  • goal_list → existing goal list
  • goal_complete → NEW goal complete
  • goal_subtask → NEW goal subtask
  • goal_report → NEW goal report
  • goal_checkpoint / goal_rollback — deferred to v1 (no storage yet; the Queue has no Checkpoint table)
• parseGoalID helper — accepts both 42 and #42 (with optional whitespace); used by all four new subcommands
• 11 tests (1 unit test for parseGoalID, all autonomy tests pass under go test -race -count=1)

Added — Skill lifecycle markers (issue #139)

• scripts/lifecycle_map.yaml — single source of truth for the lifecycle of every bundled skill. Maps each of the 34 skills to one of native | external | deprecated with a canonical: field pointing to the upstream implementation.
• scripts/sync_lifecycle.py — stdlib-only Python script. Three modes: --check (CI: exit 1 if any drift), --apply (write changes), --diff (show what would change). Hand-rolled YAML parser for the map file (no PyYAML dep).
• scripts/validate_skill.py strict mode — now requires the lifecycle frontmatter key in --strict mode and validates the value. Non-strict mode remains backward-compatible.
• sin-code skill list now prints a LIFECYCLE column with [native], [external], [deprecated], or [unknown] markers. A new parseLifecycleFromFrontmatter helper extracts the field from the embedded SKILL.md without a yaml dep.
• docs/SKILLS.md — design doc with the value table, the workflow, and the migration path.
• All 34 SKILL.md files migrated — 28 skills received the lifecycle: field; 6 were already in sync. Total: 34/34.

Added — internal/testutil/ (issue #161, race-flake hardening v2)

• Five reusable test helpers in a stdlib-only package:
  • IsolatedSQLite(t) — fresh t.TempDir()-backed *sql.DB, auto-closed
  • CleanEnv(t, kv) — set + restore env vars via t.Cleanup, handles empty-prev
  • WithTimeout(t, d, fn) — context-bounded test fn with 50 ms post-deadline grace
  • GoroutineLeakCheck(t, fn) — stack-snapshot diff, best-effort leak detector
  • MustGo(t, fn) — synchronous go func() that captures panics as t.Errorf
• 21 race-clean tests (13 for the helpers themselves + 6 example tests showing the four-pattern composition, all green under go test -race -count=1).
• testutil.doc.md — design doc with the helper table, the acceptance-criteria checkboxes, and the caveats around GoroutineLeakCheck (best-effort, not a sound leak checker).
• Diagnosis pass (informational): ran go test -count=1 -v ./... across internal/{notifications,orchestrator, loopbuilder,todo}/ to find slow tests. The slowest is TestGenerateIDUniqueness at 3.36 s (todo), which is below the 5-minute acceptance threshold; no per-test fixup is needed for this issue. The diagnosis methodology is in the runbook below.

Added — sin-code triage (issue #162)

• cmd/sin-code/triage_cmd.go — new 41st subcommand sin-code triage (and triage --format=md|json --repo owner/repo --limit N). Reads the open issue backlog via gh issue list through the ghbridge wrapper, scores each issue with a deterministic heuristic (epic +10, blocks +5 per ref, acceptance +3, not-in-v0 +5, loop-system +4, fusion +2, fresh -2, stale +1, good-first-issue -3), groups by label bucket, and renders. The markdown output is the canonical BACKLOG.md generator.
• cmd/sin-code/internal/triage/ — new package, four files (types.go, score.go, render.go, loader.go) plus triage_test.go (15 tests, all green under -race -count=1). The loader is a var so tests inject fixtures without spawning gh. No new third-party deps (M2).
• triage.doc.md — design doc with the scoring table, the per-bucket ordering rule, and the deferred-items list.

Added — sin-code catalog (issue #163, hub-assets merge)

• cmd/sin-code/catalog_cmd.go — new subcommand sin-code catalog (list | search | info) with --kind=agent|command|skill|hub and --format=text|json. The unified tool catalog that operators have been asking for: "do I have a tool for this?" — not "do I want the hub or the assets?".
• cmd/sin-code/internal/catalog/ — new package, 4 source files (catalog.go, source_hub.go, source_assets.go, catalog_test.go)
  • 21 race-clean unit tests. The Source interface (Name + List + Get) is the abstraction that lets the catalog walk both backends. Adding a new source (e.g. a remote registry) is one file.
• Merge de-duplication rule — first source to provide a (kind, name) pair wins; subsequent duplicates are dropped. The source name is intentionally not part of the dedup key, so a hub.Tool and an assets.Asset with the same name are merged into one catalog entry (the SOTA choice for the operator's mental model).
• Search ranking — name +4, short +2, description +1, tag +1; ties break by name ascending. Transparent, auditable, deterministic.
• catalog.doc.md — design doc with the de-dup table, the scoring heuristic, the deprecation plan for sin-code hub, and the known build issue (Chromedp API mismatch in PR #201, not in this PR).

Added — internal/rag/ (issue #160, RAG over instinct store)

• New package cmd/sin-code/internal/rag/ with 4 source files (embedder.go, embedder_hash.go, embedder_onnx.go, index.go, worker.go, retriever.go) + 24 race-clean tests.
  • Embedder interface + HashEmbedder (default, deterministic, dependency-free, 384-dim L2-normalized) + ONNXRuntimeEmbedder and HTTPEmbedder as documented stubs.
  • Index with optional Persister interface; the instinct subsystem uses a jsonPersister writing to $SIN_CODE_HOME/instinct-embeddings.json.
  • WorkerPool (bounded-concurrency, M7) for async embedding so the agent loop never blocks.
  • Retriever (high-level: Embedder + Index → top-N IDs).
• sin instinct search "<query>" — top-5 cosine-similarity search over the active instincts. Reindexes on every call (cheap at <100 active) and persists to disk. Renders the hits as id — trigger lines with the action underneath.
• 8 race-clean tests in internal/instinct/search_test.go for the JSON persister round-trip, the path-overriding env var, the atomic-write behavior, and the trim helper.
• rag.doc.md — design doc with the mandate-compliance analysis, the Embedder interface, the acceptance-criteria checkboxes, and the deferred-items list (GOAP Planner, Federation, real ONNX implementation).

Added — sin-code compile-spec (issue #164, v0 spike)

• cmd/sin-code/internal/spec/compiler/ — new package, 4 source files (schema.go, parse.go, validate.go, emit.go) + 24 race-clean unit tests. Round-trip test (TestRoundTrip) is the load-bearing guarantee: parse → emit → parse → emit must produce identical bytes.
• cmd/sin-code/compile_spec_cmd.go — new subcommand sin-code compile-spec with --init, --check, --out <dir>, --dry-run flags. Atomic writes (temp + rename) so a crash mid-write never leaves a half-written file behind.
• Four derived JSON outputs (contract defined; engines not yet wired — that is v1.1):
  • .sin/hooks.json
  • internal/verify/config.json
  • internal/permission/policies.json
  • .sin/loop.json (parsed but not consumed — migration path for issue #155)
• SPEC-COMPILER.md — design doc with the schema, the mandate-compliance analysis, the deferred-items list (engine wiring, remote spec inheritance, spec testing), and the relationship to issue #155.

Added — sin-code install + one-line curl|bash installer (issue #170)

• cmd/sin-code/install_cmd.go — new 40th subcommand sin-code install (and install --auto). Downloads the latest GitHub release asset, SHA256-verifies against the goreleaser-style checksums.txt, extracts the single binary, atomically places it into $SIN_CODE_BIN_DIR or $HOME/.local/bin, and prints the canonical PATH hint. Flags: --dir, --release <tag> (pin a version), --channel stable|dev, --verify-only (health-check, no write), --no-verify (offline / sanctioned CI), --dry-run.
• cmd/sin-code/internal/install/ — new pure-stdlib package. Four tiny files (release.go, github.go, verify.go, composer.go) plus race-safe install_test.go (19 tests, all green under -race -count=1). The bootstrap never depends on gh or jq, making the install cmd safe to run on a freshly imaged host.
• Root install.sh — rewritten from 1031 lines to 27 lines shell-only shim. curl -fsSL ... | bash compatible. Settles the downloaded archive, extracts the sin-code binary via three tolerant glob shapes (works across goreleaser versions), then execs sin-code install --auto so the Go entrypoint owns the verify-and-place flow. Legacy 12-step logic permanently retired — post-v3.0 the unified sin-code binary already subsumes the 7 Go tool subcommands the old installer built.
• Root install.ps1 — new 35-line Windows equivalent (irm https://raw.githubusercontent.com/.../install.ps1 | iex). Uses Invoke-WebRequest + System.IO.Compression.ZipFile, then re-execs sin-code.exe install --auto.
• permission_defaults.go — three new MCP rules under the install__* prefix (mirror of the gh_execute precedent in §3 M4): install__verify_only allow, install__dry_run allow, install__run ask. The headless daemon therefore CANNOT self-install silently, satisfying M4's "always headless" clause.
• AGENTS.md §6 — repo layout now lists install.sh, install.ps1, cmd/sin-code/install_cmd.go, and the new internal/install/ package alongside the existing 39 subcommands.

Added — Verbosity / compression mode (issue #167)

• internal/style/ — first-class verbosity mode system-prompt renderer. Five canonical modes (default, verbose, normal, terse, ultra) with byte-stable output per (mode, skillBody) pair (prerequisite for the system-prompt hash metric, issue #2). default and verbose pass through skill bodies unchanged; normal drops pleasantries + tool-call narration, terse drops articles/hedging, ultra is the tightest valid compression. Every non-default ruleset carries the auto-clarity clause that forces normal prose around destructive, security-relevant, or order-sensitive actions — the verification gate (mandate M3) must never be skipped because the output mode is terse. API surface: ParseMode(s), RenderRules(mode, body), RenderSystemBlock(level), AppendVerbosity(existing, mode), WithVerbosity(mode) (functional option).
• instinct.RenderSystemBlockWithVerbosity (issue #167): the instinct renderer now accepts a verbosity-mode string and appends the matching ruleset after the learned-instinct list (stable order: instincts → style, separated by exactly one blank line). Backward- compatible — the legacy RenderSystemBlock(active, max) still returns the bare instinct block.
• learning.Learner style hook: BeforeTurn now honors Options.Style and routes it through the new renderer. New SetStyle(level) method lets mid-session callers toggle verbosity safely under a per-instance sync.RWMutex (mandate M7, race-free). wiring.Deps.Style passes through.
• llm.style config key (internal/config.go): user-facing knob with full get/set/list/validate/TOML/JSON coverage. Validated against the canonical mode set. sin-code config set llm.style terse works end-to-end.
• Reference docs: internal/style/style.doc.md (developer), cmd/sin-code/internal/config.doc.md updated (table + example), cmd/sin-code/internal/learning/learner.doc.md updated (Style field), AGENTS.md §6 + §7 cross-references.

Added — sin-code compress (issue #172, deterministic + LLM compaction)

• internal/compress/ — first-party package implementing the caveman-compress pattern (JuliusBrussee/caveman) for SIN-Code's long-lived stores. Public API: BuildPlan, Apply, Rollback; three strategies (deterministic default, llm, hybrid) targeting four surfaces (lessons, instincts, summaries, memory, agents_md) plus an aggregate all. Plan is read-only and content-addressed (PlanHash covers Entries+Drops+Merges); Apply is atomic (snapshot written to .partial then renamed before any source rewrite) and lossless (dropped entries are preserved verbatim under ~/.local/share/sin-code/compress-snapshots/<plan-id>.json).
• Deterministic pass (compressor.go + deterministic.go): SHA-256 dedupe + utility-sorted (recency × inverse-size) keep-recent with byte-budget cap. Algorithm pins time.Now() for tests via PlanOptions.UseStableTime so two plans built from identical inputs agree byte-for-byte regardless of wall clock. Stable-time pins are verified by TestPlanDeterministicIdempotent.
• LLM summarization pass (llm.go): caveman-style compress prompt that preserves code fences, URLs, file paths, commands, and headings byte-for-byte. Validates the response with a line-based check; retries up to 2 with a targeted patch prompt on validation failure (MaxRetries, configurable). Falls back to deterministic on exhausted retries or when no provider is configured.
• Atomic snapshot+rollback: Apply writes a .partial snapshot first; Rollback(<plan-id>) reads the snapshot, refuses to consume any in-flight .partial, and restores the originals via per-target re-apply. TestApplyIsAtomicAndLossless covers one round-trip.
• CLI surface (compress_cmd.go, 41st subcommand = 40th registered cobra verb because internal.InstinctCmd etc. are in-package additions):
  • sin-code compress plan [--target all] [--strategy deterministic] [--keep-bytes 4096] [--keep N] [--recent-days N] [--json]
  • sin-code compress apply [--dry-run] [--no-llm] [--target ...] [--strategy ...] [--keep-bytes ...] [--json]
  • sin-code compress rollback <snapshot-id>
• Permission policy (permission_defaults.go): {Tool: "compress__plan", Policy: "allow"}, {Tool: "compress__apply", Policy: "ask"} (M4 — destructive), {Tool: "compress__rollback", Policy: "allow"} (restorative only). Wired so any future agent-loop surface that exposes the compressor via MCP is gated correctly.
• Regression tests (compress_test.go + testhelpers_test.go): hash determinism, plan idempotence, byte-budget enforcement, dedupe invariance, atomic-write contract, dry-run no-touching-nothing, partial-marker rollback refusal, preservation line scoping (heading, code fence, URL, file path, command line), snapshot JSON round-trip. Passes go test -race -count=1 ./cmd/sin-code/internal/compress/.
• Snapshot dir: ~/.local/share/sin-code/compress-snapshots/ (overridable via SIN_CODE_SNAPSHOT_DIR). Same form factor as lessons.db / ledger.db per AGENTS.md §7.

Added — Per-agent profile renderer (issue #175)

• Single source of truth at docs/agent-profiles/sin-profile.md (≤80 lines, KISS, hard mandates + working style + subagent contracts + per-agent notes, edits roll out everywhere).
• internal/profile package: targets map mirroring AGENTS.md §10 (claude-code, opencode, gemini, codex, cursor, windsurf, cline, copilot); per-format writers (dir, rule, marker); a byte-stable Render(tgt, body); a Verify(base, body) SHA-256 drift gate; idempotent marker-fence envelopes byte-identical to internal/skilldist (issue #169 covenants preserved).
• sin-code profile subcommand with four verbs:
  • profile show — print the source markdown
  • profile list — print the supported target table (text + --json)
  • profile render <target|all> — write one or all mirrors (idempotent; supports --dry-run for byte/audit preview without touching disk)
  • profile verify — CI gate: refuse on missing/drift; surfaces a 12-char-row table or a JSON envelope for --json
• Permission engine: profile__show / list / verify are registered as allow; profile__render is ask because it touches per-agent dotdirs (mandate M4).
• CI sync: .github/workflows/ceo-audit.yml grew a parallel profile-verify job that builds sin-code, runs profile render all && profile verify, uploads the rendered mirrors as artifacts, and fails the build if any drift surfaces. Mirrors AGENTS.md §6 + §10 — single source of truth in internal/profile/target.go.
• Test coverage: 22 race-tested Go tests pinning the byte-stable contract (golden render, marker-fence idempotency, marker-Fence covenant, Verify pass / missing / drift, write-after-write SHA equality, replace-not-append for stale mirrors).

Added — sin-debt marker convention (issue #177)

SIN-Code adopts ponytail v4.7.0's ponytail: marker convention as a first-class, parseable // sin-debt: <ceiling>, upgrade: <trigger> convention. Every intentional shortcut now carries a marker naming its ceiling and the trigger to revisit; the scanner reads them; debt stats reports them; debt check gates them.

• internal/sindept/ — scanner, aggregator, byte-stable report renderer, and policy gate. Single-package surface:
  • parser.go — Marker{File, Line, Column, Reason, Upgrade, HasUpg, Raw, Language, Symbol}, regex over five comment families (//, #, --, /*…*/, <!--…-->), ParseFile + ParseDir. Trims trailing block-comment closers (*/, -->) and post-processes captured clauses for byte-determinism.
  • stats.go — Stats{Total, WithUpgrade, WithoutUpgrade, ByFile, ByReason, ByLanguage, BySymbol, RotRisk, Oldest, MarkersPerFile}. Every map is materialized as a lex-sorted []KV so Render* output is stable.
  • report.go — RenderStats / RenderStatsString / RenderListString markdown renders with FormatVersion = "sin-debt/v1". Two scans of the same tree emit the same bytes.
  • policy.go — Policy{DefaultReasons, UpgradeTriggers, MaxNoUpgrade, RequireUpgrade, Source}. TOML overlay via .sin-code/debt-policy.toml; LoadPolicyForRoot walks upward to the closest file.
• debt_cmd.go — 41st subcommand. sin-code debt list | stats | check | policy | fix | export. Common flags: --path, --format (table|json), --no-trigger. Stats sub-commands take --by file|reason|language|symbol|age|summary. The check sub-command is the CI gate; it exits non-zero when Missing > MaxNoUpgrade or RequireUpgrade=true && Missing > 0.
• docs/sin-debt-convention.md — author-facing reference: format, examples, default reasons catalogue, default upgrade triggers.
• Permitted sindept__* tools: sindept__list, sindept__stats, sindept__policy are read-only allow; sindept__check exiting non-zero is ask; sindept__fix and sindept__export are ask because they instruct humans to edit code or write a file.
• 10 hard-coded fixture markers under cmd/sin-code/internal/sindept/testdata/ (5 languages) and 4 real markers placed in production code (cmd/sin-code/internal/lessons/ store.go × 2, …/ledger/store.go, …/orchestrator/dispatcher.go).
• Tests: 23 tests in cmd/sin-code/internal/sindept/sindept_test.go cover family coverage, trailing-closer stripping, byte-stability, vendor / hidden-dir walk, age/rot grouping, and the policy-gate semantics above vs. below threshold. Race-clean.

Notes

• sin-code debt stats is the precondition for the four-arm comparator snapshot (issue #171); byte-stable today, golden file is expected in the next PR cycle.
• The marker syntax deliberately does NOT include \Q…\E quoting — RE2 has no such construct, and the literal token sin-debt: is plain inside the regex.
• internal/sindept is the upstream of issue #179 (complexity auditor) and issue #180 (audit-engine): both are expected to call sindept.ParseFile / sindept.AggregateStats so a marker reads through the same shape regardless of consumer.

---

Added — Auto-Activation Hook (issue #176, v3.19.0)

• internal/hooklife/autoactivate/ — per-session rule injection subpackage. Two Phase hooks (autoactivate-session-start / autoactivate-user-prompt) register against any *hooklife.Registry via Activator.Register(reg). Privacy-first: off by default; activated by sin-code chat --activate <rule> or a project-local .sin-code/autoactivate.toml file.
• AutoActivate.Activator tracks per-session state under a single sync.RWMutex (mandate M7 — race-safe under go test -race -count=1). OnSessionStart(sid, opts) is idempotent; OnUserPrompt(sid, prompt) returns the rule set re-emittable for this turn, with trigger-phrase substring matching for natural-language activation. EndSession(sid) drops state on exit.
• RuleSet.Render() is byte-stable: any two RuleSets with the same name+body+trigger tuples produce identical bytes regardless of insertion order (prerequisite for the system-prompt hash metric, issue #2).
• sin-code chat --activate terse,skill-x comma-separated rule list; --no-trigger suppresses per-prompt phrase matching; reads .sin-code/autoactivate.toml silently when present.
• Tests: 35 race-safe unit tests + 8 chat-wiring integration tests, 91.5% statement coverage on the autoactivate package. New package follows the existing hooklife Phase contract; no new external deps.

Added — Orchestrator output contracts (issue #174)

• Caveman-style output contract for the four orchestrator sub-agents (internal/orchestrator/output_contract.go). Every Finding renders to ONE byte-stable line: <path>:<line> — <symbol> — <tag> — <hint> # c=<confidence>. Five closed tags (parallel to JuliusBrussee/ponytail): delete | simplify | rebuild | risk | verify. Em-dash U+2014 separator; no prose, no pleasantries, no hedging (you might, perhaps, could consider, maybe, i think, sort of, should probably, … — closed set of 12 phrases, case-folded).
• Finding struct (internal/orchestrator) and ParseFinding / ParseFindings regex parsers with strict byte-stability. Render() is fully deterministic — Finding{...} → Render() → ParseFinding() → equal struct, every byte counted (verified by TestParseFinding_RoundTrip).
• VerifyFindings runs the full contract: structural (Path != "", Tag ∈ {delete, simplify, rebuild, risk, verify}), lexical (zero hedging, hint ≤ 240 chars, no trailing punctuation), and emits per-Finding error strings — never silent drops.
• Wired into the four sub-agents:
  • Critic.Drive parses the LAST attempt's prose into CriticResult.Findings and surfaces CriticResult.ParseErrors for the orchestrator to re-inject as retry feedback (mirrors the verify.fail flow).
  • Adversary.Review derives Findings from the structured Attack slice (landed → risk, cleared → verify); the CounterexampleBrief free- form prose is preserved as the audit trail.
  • Governor.Execute derives one risk Finding per Escalation (Path = task://<ID>, Symbol = <from>-><to>); the prose Reason stays on Escalation for the audit log.
  • Cartographer.Findings(k) exposes the PageRank-sorted top-k as verify-tagged Findings (opt-in: k ≤ 0 yields the empty slice).
• Byte-stable golden tests (output_contract_test.go and output_contract_integration_test.go) pin one fixture per agent; rendering drift breaks the build (the prerequisite for issue #168's ledger-level token-cost hashing).

Added — Loop Engineering (decoupled completion authority)

Added — MCP tool-manifest compression (issue #173, v3.19.0)

• internal/mcpcompress/ — ponytail-tag compressor for sin-code serve --compress-tools (issue #173). Five canonical tags drive five byte-stable Rules:
  • delete → DeleteHedges drops pleasantries / hedge adverbs ("safely", "carefully", …).
  • stdlib → StdlibPatterns drops redundant stdlib parentheticals ("(via stdlib)", Go stdlib …).
  • native → DropTrimEncouragement drops M6 tail clauses Always prefer over native X / Prefer sin_X over native Y (the M6 mandate is internal-only, not for the model).
  • yagni → YagniPatterns drops speculative (experimental) / (TBD) / (reserved) parentheticals.
  • shrink → ShrinkExamples drops redundant (e.g. …) / (such as …) parenthetical examples.
• Three new serve flags:
  • --compress-tools — apply the full default ponytail tag set (delete|stdlib|native|yagni|shrink) before registration.
  • --compress-tags <csv> — override the active tag set (e.g. --compress-tags "delete,yagni,shrink"). Unknown tags are silently dropped; the active set is logged via --print-stats.
  • --print-stats — emit a left-aligned text table to stderr (tool / orig / comp / saved / ratio + TOTAL row + active rules).
• Tool names unchanged. The compressor mutates only Description; the 47 MCP tool Name fields (public API per AGENTS.md §10) are never modified. TestCompressSpec_NameMutable guards this.
• Byte-stable per (tool_spec, ruleset) — every Rule, the Pipeline, and the post-pipeline Normalize are deterministic and idempotent. The compressor_test.go golden suite is the single source of truth — any Rule regex / declaration-order change must update the gold expectations in the same PR. Prerequisite for the system-prompt hash metric (issue #2).
• Real-world savings. Smoke test against the 47-tool registry: sin_execute 80→73 bytes (-7, 8.8%); sin_read 200→167 bytes (-33, 16.5%); sin_write 194→160 bytes (-34, 17.5%); sin_edit 474→441 bytes (-33, 7.0%). TOTAL 3977→3870 bytes saved across 4 affected tools (-107 bytes / 2.7%); the other 43 tools' descriptions don't match any of the conservative patterns and stay byte-identical. Per-tool --print-stats output is deterministic across runs (no time, no random).

Added — Loop Engineering (decoupled completion authority)### Added — Loop Engineering (decoupled completion authority)

• Stop-gate harness (internal/stopgate): an independent completion authority consulted after the verify-gate passes. Hybrid mode runs deterministic checks first (fail-closed) then a strong/equal LLM judge (SIN_EVALUATOR_MODEL) for non-mechanical criteria; a green judge can never override a red deterministic check. Rejection forces the loop to keep working with the open criteria injected back in. 92.5% test coverage.
• Goal contracts / Definition-of-Done (internal/goalcontract): machine- checkable acceptance criteria per goal, layered resolution (explicit file + inline --criteria + --done-when, auto-detected Go checks incl. a no-new-todos diff guard, verify-cmd fallback). Persisted in the queue's contract column. goal add --criteria/--contract-file. 94.5% coverage.
• Continuation instead of hard abort (agentloop): with AllowContinuation, hitting max-turns now checkpoints and returns a resumable Result (Continuation=true) instead of erroring. The daemon re-enqueues via queue.Continue (refunds the attempt, bumps a continuations counter) bounded by --max-continuations. Long tasks never need a human restart.
• Recursive goal decomposition (autonomy.Queue): parent_id/depth columns, AddSub, depth-first draining (a parent only finalizes once every child verifies, via Complete→blocked→TryFinalize/bubbleUp). The daemon exposes a spawn_subgoal tool bounded by --max-depth.
• Autonomous backlog discovery (autonomy/discover.go): scans TODO/FIXME/ XXX/HACK markers and unchecked MASTER_TODO.md items into deduplicated goals (AddDiscovered with a dedup_key). Exposed as a discover trigger type and the goal discover [--dry-run] command — the agent finds its own work.

Added — Learning Subsystem (continuous learning in Go)

• internal/instinct/ — continuous-learning subsystem (port of the continuous-learning-v2 homunculus model in a clean-room Go reimplementation). Project-scoped + global Markdown-with-frontmatter store, confidence 0.3–0.9, Reinforce/Contradict/Decay math with atomic.Value-backed env-overridable tuning, heuristic + LLM-backed extractors (with graceful fallback), cross-project promotion, cluster evolution into Skill/Command/Agent proposals, and a system-prompt block renderer that closes the learning loop. CLI: sin instinct status|projects|evolve|promote|prune|export|import|show|forget|history. Storage: $SIN_INSTINCT_DIR | $XDG_DATA_HOME/sin-code/instinct | ~/.local/share/sin-code/instinct.
• internal/hooklife/ — native Go lifecycle-hook system (no Node dependency). Phases: PreToolUse, PostToolUse, Stop, SessionStart, SessionEnd, PreCompact, UserPrompt. PreToolUse may Block (ECC exit-code-2 equivalent); other phases aggregate warnings. Per-hook timeout, panic recovery. Built-in hooks: block-no-verify, config-protection, post-edit-format, quality-gate (against the real verify.Gate), cost-tracker, suggest-compact. CLI: sin hooks list|test.
• internal/assets/ — harvested agent/command/skill loader with schema validation (port of ECC CI validators, including unsafe-unicode and duplicate detection), Selector for domain+keyword-based ranking, and an import subcommand that harvests skills from a vendored source repo with origin/license attribution. CLI: sin assets list|validate|show|import.
• internal/evalharness/ — eval-driven development. EvalSet / Run / Result types, pluggable Scorers (exact, contains-all, success-flag, LLM-judge, composite, CompileAndRun), per-case timeout, JSONL run history, and Compare for case-by-case regression detection with --fail-on-regress as a CI gate. CLI: sin eval run|list|compare.
• CompileAndRun scorer (issue #181) — ponytail correctness.js analog for SIN-Code. Extracts fenced code from model output, compiles it (go/python/javascript/bash), and runs a sandboxed self-check. Returns 1.0 only when compile + run pass; skip_test mode accepts trivial one-liners after compile-only (YAGNI for tests). Wired into sin-code eval run via --scorer compile-and-run --language <lang> and into Golden Datasets via test_cases[].scorer.
• internal/dispatch/ — turns loaded command and agent assets into executable actions. ECC-style placeholder substitution ($ARGUMENTS, $1..$9, $@, ${flag}), Dispatcher routes slash-commands to PromptSink and agent requests to SubagentRunner. Closes the load → select → dispatch → run pipeline.
• internal/prp/ — Product Requirement Prompt workflow. Persistent reviewable plans under .sin/prp/<id>.md driven through phases (draft → planned → implementing → verifying → ready → shipped). Each step persists, so a run is interruptible and resumable. Verification failure kicks the PRP back to implementing. CLI: sin prp new|run|status|plan|implement|verify|pr.
• internal/adapters/ — concrete adapters that implement the abstract hooklife.Verifier, instinct.MemorySink, and instinct.Completer interfaces against the real SIN-Code subsystems (verify.Gate, memory.Store, llm.Client). Fail-soft: missing subsystems degrade to no-ops, never block startup.
• internal/learning/ — bridge package between agentloop.Loop and the new subsystems. Learner exposes BeforeTurn (prepends active-instinct system block), BeforeTool (PreToolUse dispatch, may veto), AfterTool (PostToolUse + observer feed), EndTurn (observer flush), PreCompact (flush + hook dispatch). Built once at startup via learning.New(Options).
• internal/wiring/ — Build(Deps) assembles the full Bundle{Learner, Dispatch, Eval factory, PRP deps} in one call.
• examples/eval-sets/ — go-quality.json (build, vet, test, secrets scan) and instinct-behavior.json (end-to-end learning loop validation).
• Five new top-level subcommands wired into cmd/sin-code/main.go: sin instinct, sin hooks, sin assets, sin evalset, sin prp. (The existing sin eval — Golden-Dataset runner from issue #75 — is preserved unchanged; the new harness lives at sin evalset to avoid a cobra Use: collision.)

Added — Complexity Audit (issue #180)

• sin-code audit complexity — repo-wide ponytail-audit analog. Five tags (delete, stdlib, native, yagni, shrink), deterministic static pass (single-impl interfaces, single-product factories, wrapper functions, one-export files, dead flags/config, hand-rolled stdlib), optional LLM judge for top-N findings. Output: one-liner per finding ending with net: -<N> lines, -<M> deps possible. or Lean already. Ship..
• cmd/sin-code/internal/audit/ — new Auditor, Finding, Result types; // sin-debt: markers approve findings and exclude them from the net total.
• sin-code ceo-audit — new 48-gate CEO-grade audit. The 48th gate is the complexity audit; score contribution is +1 per 100 removable lines.
• Docs docs/complexity-audit.md and tests complexity_test.go + audit_cmd_test.go with race-free coverage.

Notes

• All loop-engineering features are opt-in and fail-safe: a nil stop-gate / empty contract / AllowContinuation=false preserves exact legacy behavior.
• The learning subsystem is additive — it does not modify the existing internal/agentloop package. The chat command can opt into the learner by calling learning.New(...) and invoking the lifecycle methods around its loop run; the default is "no learning wired" so the chat behavior is unchanged for existing users.
• go test -race clean across the new learning packages. No new third-party dependencies (gopkg.in/yaml.v3 was already transitively present in go.sum).

Added — Spec-Layer (issue #157)

The Spec-Layer is the bridge between human intent and machine-checkable verification. A *.spec.md file captures a change's contract as Requirements + Acceptance Criteria (each with an optional verify: shell command) + Invariants. The agent and CI can then run those checks, and the drift checker verifies the code still matches the spec's signatures.

• internal/spec/ — the Spec-Layer core (issue #122, hardened by #157). Parses *.spec.md files; Spec.Marshal writes them back in canonical form. Spec.Check(ctx, timeout) runs every criterion's verify: shell command and aggregates per-criterion results into a CheckReport with HasFailures() for the CI gate. Spec.Author(ctx, desc, opts) runs the LLM Planner → Implementer → Drift-check loop with up to 3 retries on drift. Spec.DetectSignatureDrift(root) walks the source tree and compares backtick-wrapped Go/Python function signatures and JSON object shapes against the spec. ~3700 LOC, 22+ tests, race-clean.
• Python signature matching via subprocess to python3 + ast (internal/spec/python.go). Embedded extractor script as a const string; no separate .py file to ship. Top-level functions only in v0; method-on-receiver deferred to PR 4.
• JSON shape matching (internal/spec/json.go). Structural type check: every spec key must exist in a JSON file with a compatible type (string/int/bool/array/object/null, with []T and {} as sugar). No new deps (M2).
• LLM wiring (internal/wiring/spec.go). NewSpecCompleter adapts llm.Client to spec.Completer so sin spec author can drive the end-to-end loop. Env var SIN_SPEC_LLM_BASEURL is the v0 hook for a local model; --dry-run is the no-LLM path that returns a stub spec for end-to-end testing.
• CLI: sin spec validate|show|check|author. New flags: check --drift runs the spec↔code drift; check --root <dir> scopes the walk; author --dry-run --out <file> writes a stub spec; author --apply opens a PR via gh (scaffolded, wired in PR 4).
• Pre-commit hook (scripts/spec-drift-check.sh): runs sin spec check --all on every commit. Override path is git commit --no-verify per M3.
• CI workflow (.github/workflows/spec-ci.yml): runs the spec check on every PR and push to main. A must-priority failure blocks the merge. M1-compliant (n8n-delegated).
• Spec format change: the verify: annotation now requires backtick-wrapping (`verify: cmd`) so the parser doesn't misread plain prose as a verify command. Existing pre-v3.18 spec files need a one-time sed pass; the migration is documented in docs/spec-layer.md.
• Tests: 22+ tests in internal/spec/, race-clean. Cover the parser, the verify:-runner, the LLM loop (with a stub Completer), Go/Python/JSON drift, type compatibility, and persistence.
• Docs: docs/spec-layer.md is the canonical reference; it supersedes the older docs/spec-layer.md content (the file is extended, not replaced). docs/SPEC-LAYER.md is the design spec for the hardening pass.

Added — Four-arm Eval Comparator (issue #171)

• internal/evalharness/arms.go — built-in Arm constructors for the canonical four-arm harness: __baseline__ (no system prompt), __terse__ ("Answer concisely."), __lazy_skill__ (skill-code-lazy body issued from issue #178), and the <user-skill> arm named by --skill. Skill discovery is best-effort and falls back to a byte-stable [skill unavailable] placeholder so snapshots remain diff-clean.
• internal/evalharness/comparator.go — Compare(ctx, EvalSet, []Arm, CompareOptions) runner. The outer loop is per-arm, the inner loop per-case; per-(case, arm) results are aggregated into TotalsByArm per arm. NoOpSubject and SetDefaultSubject keep the harness honest for offline / stub CI runs.
• internal/evalharness/prices.go — self-pricing price book (USD per 1k prompt + completion tokens) keyed by Arm.PricingName. Known models: stub, gpt-4o, gpt-4o-mini, claude-3.5-sonnet, fireworks-qwen2.5-7b, fireworks-llama-3.1-70b. Unknown names produce a warning in CompareReport.Warnings and zero USD (so the harness never silently under-reports cost).
• internal/evalharness/snapshot.go — deterministic snapshot round-trip. BuildSnapshot sorts rows by ArmID, takes medians across all per-case values, and emits byte-stable JSON. WriteSnapshotFile/LoadSnapshotFile round-trip disk I/O. DiffSnapshots produces row-level deltas with the changed-skill-body signal for SKILL.md drift.
• Result/Output extensions: Result carries ArmID, PromptTokens, CompletionTokens, TotalTokens, LOC, USD (all omitempty for backward compatibility). Output carries an optional USD for Subject authors that compute cost at the source.
• cmd/sin-code/eval_cmd.go extensions — three new subcommands: eval compare, eval snapshot, eval diff. eval run --arm baseline,terse,lazy_skill,<skill> opts into the comparator path; without --arm the legacy Golden-Dataset path is preserved unchanged. The four-arm matrix output mirrors ponytail's benchmarks/README.md:34-58 columns (LOC, USD, latency, correctness).
• evals/three-arm-example.json — canonical example dataset with 3 cases: 2 LOC-countable (gopher explain + reverse Go function), 1 LLM-judge (lz4 vs zstd).
• Comparator test coverage — 11 new tests pass with go test -race -count=1. Median aggregation, 4-arm matrix, snapshot byte-round-trip, schema-version rejection, late-write warnings, parallel-vs-serial equivalence (race-detector safe).
• Compare renamed in regression.go: Compare(base, cand, eps) → CompareRuns(base, cand, eps) to free the bare name for the new multi-arm comparator (issue #171). All three call sites (cli.go, evalharness_test.go) updated.
• AGENTS.md §12.1 added documenting the four-arm comparator contract: the honest delta = <user-skill> − __terse__, not the inflated <user-skill> − __baseline__.

Added — Bundled skills (issue #178)

• skill-code-lazy (35th bundled skill, in skills/process-skills/): SIN-Code adaptation of Dietrich Gebert's ponytail skill — "ship the laziest version that actually works" with the 6-stufige Leiter (YAGNI → stdlib → platform → existing dep → one function → minimum that works). Gated by verify.pass (mandate M3): the skill is inert while verify.result ∈ {pending, pre, fail} and only arms after the verify-gate. Activation keyword lazy_skill (issue #176) binds to the four intensities off | lite | full | ultra.
• sin-debt: marker cookbook in templates/debt-markers.md: paired ceiling + upgrade-trigger convention (issue #177); every shortcut ships a // sin-debt: <ceiling>, upgrade: <event> pair so reviewers can audit YAGNI vs hardening pressure.
• Byte-stable render contract: the 5 keyword examples in SKILL.md render to identical octets across runs (prerequisite for the issue #2 system-prompt hash metric).
• Naming-convention exception recorded in AGENTS.md §10: the canonical pattern is skill-<category>-<name>, but skill-code-lazy is preserved as the v3.18.0 exception because the lazy_skill activation keyword binds to the literal frontmatter name.

Added — Complexity Review (issue #179)

• cmd/sin-code/internal/complexity/ — static, AST-based complexity analyzer implementing ponytail's 5-tag format: delete, stdlib, native, yagni, shrink. Detects single-implementation interfaces, one-product factories, wrapper-only functions, hand-rolled min/max, dead flag-like variables, repeat-append loops, and imports that duplicate stdlib/platform features. Respects // sin-debt: and # sin-debt: markers (issue #177).
• cmd/sin-code/review_cmd.go — new top-level sin-code review command with sin-code review --complexity [--path] [--since <ref>] [--tags] [--format text|json|markdown].
• Output format: one line per finding (<tag>: <what>. <replacement>. [path:line]), ranked by line count and removed dependencies, ending with net: -<N> lines, -<M> deps possible. or Lean already. Ship.. net_lines and net_deps are included in JSON output.
• Tests: cmd/sin-code/internal/complexity/complexity_test.go + golden file, race-clean.

[v3.23.0] - 2026-06-18

Added — v3.23.0 Roadmap

• Autonomous research report (issue #384): sin-code research <topic> enqueues an autonomous goal that searches the web, fetches sources, and synthesizes a Markdown report. Permission split: research__dry_run/list allow, research__run ask (M4).

• SWE-bench integration tests (issue #363): Comprehensive 670-line test suite for the swebench package covering dataset loading, TestCase conversion, scoring, and JSON serialization.

• Scientific research skill (issue #387): New bundled skill skill-shop-research-scientific with search strategies for PubMed, USPTO patents, and arXiv preprints.

• Unified diff tools (issue #365): sin_apply_diff and sin_generate_diff chat tools. sin_apply_diff validates and applies a unified diff to a file; sin_generate_diff generates a unified diff from old/new content. Wired into the permission matrix as allow (same tier as sin_edit).

• Dynamic MCP discovery (issue #368): mcpclient.DiscoverConfigs() scans ~/.config/mcp/servers/*.json, Claude Desktop, opencode, codex, and .sin-code/mcp.json. New CLI verbs: sin-code mcp discover and sin-code mcp add <name>. Discovered servers are merged into LoadConfigs().

Fixed — Build Drift on main

• Removed stale toolretrieval import from chat_mcp.go.
• Fixed literal-newline syntax errors in chat_tools.go toolBash output.
• Restored extraSpecs() function structure in chat_tools_extra.go.
• Fixed fmt.Sscanf error handling in browser_interaction.go.
• Added missing repetitionThreshold / repetitionWindow fields and CLI flags to chat_cmd.go to match loopbuilder wiring.
• Added LoopDetector field to agentloop.Loop and wired recording with the loop.detected hook event.

[v3.22.0] - 2026-06-18

Added — SIN Fusion v1 Enhancements (v3.22.0)

• Plan-Merge mode (issue #393): New ModePlanMerge tournament mode — N models plan in parallel, an LLM judge merges the best insights into a Unified Plan, one model codes it, verify-gate validates. Unlike PoC and Oracle which discard N-1 outputs, plan-merge preserves all insights. Config: fusion.mode = "plan-merge".

• Oracle as default (issue #394): Default fusion mode changed from PoC (first-pass-wins) to Oracle (all run, judge picks best). Quality over cost. PoC still available via fusion.mode = "poc". New fusion.mode config key accepts "poc" | "oracle" | "plan-merge".

• Model Performance Registry (issue #395): Persistent per-model-per-category benchmark database (modelperf.db) that drives benchmark-based model selection for Fusion. CLI: sin-code fusion benchmark/rank/recommend. Recommendation engine blends 80% pass_rate + 20% cost-efficiency. Auto-wired into loopbuilder — recommended models are prioritized in tournament provider selection. Cold-start: falls back to full pool.

Added — Fusion v1 Core (v3.22.0, earlier issues)

• Oracle judge uses SIN_EVALUATOR_MODEL with separate client (anti-bias)
• Confidence-aware difficulty gate: ShouldRunWithConfidence
• sin-code fusion CLI subcommand: status/config/providers

Added — SIN Fusion v1 Enhancements (v3.22.0)

• Plan-Merge mode (issue #393): New ModePlanMerge tournament mode — N models plan in parallel, an LLM judge merges the best insights into a Unified Plan, one model codes it, verify-gate validates. Unlike PoC and Oracle which discard N-1 outputs, plan-merge preserves all insights. Config: fusion.mode = "plan-merge".

• Oracle as default (issue #394): Default fusion mode changed from PoC (first-pass-wins) to Oracle (all run, judge picks best). Quality over cost. PoC still available via fusion.mode = "poc". New fusion.mode config key accepts "poc" | "oracle" | "plan-merge".

• Model Performance Registry (issue #395): Persistent per-model-per-category benchmark database (modelperf.db) that drives benchmark-based model selection for Fusion. CLI: sin-code fusion benchmark/rank/recommend. Recommendation engine blends 80% pass_rate + 20% cost-efficiency. Auto-wired into loopbuilder — recommended models are prioritized in tournament provider selection. Cold-start: falls back to full pool.

Added — Fusion v1 Core (v3.22.0, earlier issues)

• Oracle judge uses SIN_EVALUATOR_MODEL with separate client (anti-bias)
• Confidence-aware difficulty gate: ShouldRunWithConfidence
• sin-code fusion CLI subcommand: status/config/providers

Added — SOTA Skill Infrastructure

• Skill frontmatter standardization: All 36 bundled skills now have filled compatibility: (sin-code, opencode, claude-code, codex) and metadata: (author, version 3.20.0) frontmatter fields. The validator (scripts/validate_skill.py) enforces required_tools as a YAML list in --strict mode.
• required_tools frontmatter field: 8 key skills now declare their required SIN tools (e.g. skill-code-build requires sin_edit, sin_test, sin_quality_gate). Parsed at runtime by cmd/sin-code/internal/skillmgr/required_tools.go and merged additively (deduplicated, sorted) into agentloop.Loop.CoverageRequiredTools via loopbuilder.Config.ActiveSkills. When a skill is activated, the ToolCoverageEnforcer (issue #248) rejects task completion if the required tools were not invoked. 17 race-clean tests.
• 3 new skilldist targets: aider (.aider/conventions/<skill>.md, rule format), continue (.continue/rules/<skill>.md, rule format), zed (.zed/rules/<skill>.md, rule format). Total targets: 8 → 11. Non-breaking (adding targets is allowed per AGENTS.md §10).
• 3 new skilldist targets in profile distribution: aider, continue, zed added to cmd/sin-code/internal/profile/target.go with sin-code.md profile paths. Profile tests updated.
• 3 skill eval datasets: evals/skill-code.json (3 cases: build / refactor / plan), evals/skill-debug.json (2 cases: race / nil-pointer RCA), evals/skill-github.json (2 cases: actions / readme). All four-arm comparator compatible (baseline / terse / lazy_skill / target-skill). Wired into .github/workflows/eval-n8n.yml n8n-delegated CI (mandate M1).

Fixed — Skill Validation

• skill-github-governance: lifecycle: external was nested under metadata: instead of top-level frontmatter — validator rejected it in --strict mode. Fixed: lifecycle is now a top-level key.
• skill-github-readme: lifecycle: external nesting issue (same as governance). Additionally, context/, frameworks/, tasks/, templates/ directories were empty — validator flagged them in strict mode. Fixed: all four directories now contain .md files with triggers, standards, workflow, and templates.
• skill-code-create: Version references updated from v3.17.0 → v3.20.0, skill count 34 → 36, frontmatter compatibility + metadata filled. External duplicate ~/.config/opencode/skills/skill-create/ removed.

Updated — AGENTS.md §10

• Skill distribution table: 8 → 11 targets (aider, continue, zed added).
• Profile distribution table: 8 → 11 targets.
• CLI --agent flag help text and doc comments updated.

Added — SOTA Skill Infrastructure (PR #359, #360, #361)

• required_tools on all 37 skills: 28 remaining skills now declare their SIN tool dependencies. Every skill has a deterministic tool binding enforced by the ToolCoverageEnforcer at runtime.
• skill-code-create teaches required_tools: SKILL.md, frameworks/standards, templates/prompt+output, tasks/workflow all updated to scaffold the required_tools field in new skills.
• 8 new eval datasets: skill-browser, skill-design, skill-ecosystem, skill-infrastructure, skill-memory, skill-planning, skill-process, skill-shop. All 11 categories now have eval coverage. Wired into eval-n8n.yml CI.
• 12 external skills: sources: field: All lifecycle: external skills now document their origin in metadata.sources.
• skill-code-graph: empty dirs filled with .md files, .gitkeep removed, lifecycle moved to top-level frontmatter.
• Stale counts fixed: README.md (34→37 skills, 40→42 subcommands), AGENTS.md (35→37), skill-code-create (36→37).
• skillmgr 100% coverage: edge case test for empty/whitespace skill names.
• TUI WIP integrated: 12 untracked files (model_switcher, help_overlay, file_picker, mouse, render_cache, theme_custom, crash_recovery, etc.) integrated. go build ./cmd/sin-code/... passes clean.
• Doc consistency: skilldist.doc.md (4 fixes), release-notes/v3.20.0.md.

Added — SIN Fusion v1: Verify-Tournament (issue #290)

The largest release in SIN-Code history. Two epics (21 issues), ~50 new features, ~500 new tests, 33.5K lines of TUI code. Every test passes -race -count=1 (mandate M7). All 87 packages green.

Added — Epic #289: Predictive Multi-Agent Orchestrator (7 issues)

The largest orchestrator upgrade since v3.4.0. The orchestrator now plans, dispatches, and pre-warms sub-agents in parallel with probability-scored DAGs and event-driven execution.

• #282 — DeepPlanner (internal/orchestrator/deepplanner.go): parallel DAG construction with per-node probability scores. The planner emits a DAGPlan with weighted edges; the dispatcher uses the scores to prioritise high-probability branches. BuildDAGPlan benchmarks at 1.1 µs / 14 allocs (mandate M2). 12 race-clean tests.
• #283 — Event-Driven DAG Dispatcher (internal/orchestrator/dispatcher.go): replaced the 50 ms polling loop with channel-based event dispatch. Each node completion sends on a buffered channel; the dispatcher wakes instantly. Zero polling overhead, zero goroutine leaks (M7). 9 race-clean tests including a 100-node stress test.
• #284 — Parallel Subagent Spawning (internal/agentloop/subagent.go): the agent loop can now spawn N concurrent isolated sub-agents, each with its own session, permission scope, and verify gate. Bounded by --max-parallel-subagents (default 4). Race-safe via sync.WaitGroup + buffered error channel (M7). 11 race-clean tests.
• #285 — Anticipatory Agent Pre-warming (internal/orchestrator/prewarm.go): PreWarmManager launches likely-needed agents before the planner formally requests them. Pre-warm threshold is configurable (orchestrator.prewarm_threshold, default 0.7). LoopAgent implements the PreWarmer interface. 8 race-clean tests.
• #286 — DAG Visualizer TUI (internal/tui/dagview.go): live DAG view rendered in the TUI with status icons (pending ●, running ◐, done ✓, failed ✗). Auto-refreshes on dispatcher events. Keyboard: j/k to navigate nodes, Enter to inspect, q to quit. 6 tests.
• #287 — Real LLM-Backed Agents (internal/orchestrator/loopagent.go): replaced MockAgent with LoopAgent, a real agent backed by agentloop.Loop. Each sub-agent runs the full PLAN → ACT → VERIFY → DONE cycle with the sacred verify gate (M3). LoopAgent implements PreWarmer. 10 race-clean tests.
• #288 — Pattern Learning (internal/orchestrator/patterndb.go): PatternDB learns task sequences from past sessions. After each completed task, the pattern of (intent → tools → outcome) is stored. On the next run, MatchPrompt retrieves the top-k patterns and feeds them to the DeepPlanner as priors. MatchPrompt benchmarks at 198 ns / 0 allocs. JSON-file storage (v0); SQLite migration deferred. 14 race-clean tests.

Added — Epic #281: Claude Code Leak Inspiration (14 issues + #290)

A systematic port of the best UX and infrastructure ideas from Claude Code's leaked system prompt, adapted to SIN-Code's verification-first architecture.

• #267 — Context Usage Visualizer (/ctx-viz): bar chart showing token budget consumption by category (system prompt, tools, history, context). Inline in the TUI; updates live. 5 tests.
• #268 — Agent Dashboard (internal/tui/dashboard.go): session-fleet overview showing all active sessions, their status, token usage, and cost. Sorted by last-active. 4 tests.
• #269 — Configurable Keymaps (internal/tui/keymap.go): 7 keymap contexts (chat, dag, dashboard, file-browser, diff, search, permission) with per-context bindings. ~/.config/sin-code/keymap.toml overrides defaults. 8 tests.
• #270 — Lazy Tool Loading (internal/mcpclient/lazyloader.go): tool definitions are loaded on-demand via tool_search instead of upfront. Reduces the tool manifest from 134K → ~5K tokens at session start. Tools are fetched by keyword when the model needs them. 12 race-clean tests.
• #271 — Frustration Detection & Adaptive UX (internal/tui/frustration.go): detects repeated failed attempts, rapid re-prompts, and apology patterns. When frustration is detected, the TUI offers a /grill suggestion or a /btw context break. Threshold configurable. 7 tests.
• #272 — YOLO Risk Classifier (internal/permission/risk.go): replaces the binary --yolo flag with a graded auto-approve system. Each tool call is classified as safe (auto-approve), moderate (ask), dangerous (always ask), or destructive (always deny in headless). --yolo now accepts safe, moderate, or all. 10 race-clean tests.
• #273 — autoDream Background Memory Consolidation (internal/memory/autodream.go): periodic background consolidation of episodic memory into semantic summaries. Runs on a configurable interval (memory.autodream_interval, default 15 min). Uses the lessons DB as input; writes consolidated summaries back. Respects M3 (never runs during active verify). 9 race-clean tests.
• #274 — Undercover Mode (internal/agentloop/undercover.go): when enabled (--undercover), AI-generated commits, PRs, and code comments are stripped of AI-identifying language. Commit messages are rewritten to human style; co-author trailers removed. Toggle per-session. 6 tests.
• #275 — Claude Fable 5 & Mythos 5 Provider Support (internal/llm/provider.go): new provider entries for Anthropic's claude-fable-5 and claude-mythos-5 models. Registered in the provider registry; auto-detected from llm.model config key. 4 tests.
• #276 — /btw Command (internal/tui/chat.go): side-questions without breaking the current context. /btw <question> spawns a temporary sub-session, answers the question, and returns — the main context is untouched. 5 tests.
• #277 — Prompt Caching (internal/llm/promptcache.go): 5-minute TTL cache for system prompts + tool definitions. Anthropic integration via cache_control block. Cache hit benchmarks at 35 ns / 0 allocs. Saves ~80% of input tokens on repeat turns within the TTL window. 11 race-clean tests.
• #278 — Context Compaction (internal/agentloop/compaction.go): 5 strategies when the context window fills: Summarize (LLM summarization), Truncate (drop oldest), Selective (keep tool results, drop intermediate), SlidingWindow (keep last N turns), Hybrid (summarize old + keep recent). Configurable via agentloop.compaction_strategy. 14 race-clean tests.
• #279 — Bubbletea 4 Golden Rules + Layout Debug + Inline Diff (internal/tui/): TUI layout follows the 4 golden rules (viewport-first, deterministic resize, no overlapping panes, graceful degradation). New --debug-layout flag draws pane boundaries. Inline diff view for sin_edit results (before /after side-by-side). 9 tests.
• #280 — Status Line (internal/tui/statusline.go): persistent status bar showing model name, token count, estimated cost, and session duration. Updates live during streaming. Configurable via tui.status_line (on/off/items). 5 tests.
• #290 — SIN Fusion v1: Verify-Tournament — see the dedicated entry below for full details (multi-model fan-out on verify-fail, 6 Fireworks models, cost-governor, difficulty gate, PoC-only).

Added — SOTA TUI Chat Experience (33.5K lines)

A complete rewrite of the chat TUI, bringing it to parity with Claude Code and Codex CLI while preserving the sacred verify gate (M3).

• Animated streaming with blinking cursor — braille spinner (⠋⠙⠹⠸⠼⠴⠦⠧⠇⠏) and emoji spinner (◐◓◑◒) during LLM generation; cursor blinks at the last token position.
• Visual message blocks — user/assistant/tool messages rendered as distinct visual blocks with role icons.
• Tool cards — each tool invocation renders as a compact card with name, status, and result preview.
• Branded welcome screen — ASCII art header with version, model, and quick-start commands.
• Context bar — live token/context usage indicator at the bottom of the chat viewport.
• Compact mode toggle — Ctrl+M toggles between full and compact rendering (drops visual blocks, keeps content).
• Session info bar — shows session ID, parent ID (if forked), and turn count.
• Permission popup with risk coloring — the permission dialog color-codes by risk level (green=safe, yellow=moderate, red=dangerous) per the YOLO risk classifier (#272).
• View transitions — slide and fade animations between chat, DAG, dashboard, and file-browser views.
• Notification toasts — ephemeral success/error/info messages in the top-right corner; auto-dismiss after 3 s.

Added — Enhanced Input Experience

• Vim mode (internal/tui/vim.go): Ctrl+V toggles vim keybindings for the chat input (normal/insert/visual modes). Supports hjkl, w/b, dd, yy, p, u, cw, and search (/).
• History navigation — Up/Down cycles through per-session prompt history; persists across restarts.
• Paste detection — multi-line paste is detected and confirmed before sending (prevents accidental submission).
• Word movement — Alt+Left/Alt+Right jumps by word; Ctrl+A/Ctrl+E jumps to line start/end.
• Auto-complete — / triggers slash-command autocomplete popup; @ triggers file-path autocomplete.
• Slash command autocomplete popup — fuzzy-filtered list of available slash commands with descriptions.

Added — Split Pane, File Browser & File Viewer

• Split pane (internal/tui/split.go): horizontal/vertical pane splitting. Ctrl+S toggles split mode; the secondary pane can show file browser, diff, DAG, or dashboard.
• File browser (internal/tui/filebrowser.go): tree-view file navigator with j/k navigation, Enter to open, d to diff against HEAD.
• File viewer (internal/tui/fileviewer.go): read-only file viewer with syntax highlighting and line numbers.

Added — In-Chat Search

• /search <query> — search within the current session's message history. Results are highlighted inline; n/N cycles forward/backward.

Added — Syntax Highlighting (7 languages, dependency-free)

• internal/tui/syntax/ — hand-rolled lexer for Go, Python, Rust, JavaScript, TypeScript, JSON, and Markdown. Zero external deps (M2). Token-based coloring with 16-color fallback for non-truecolor terminals.
• 21 tests covering each language + edge cases.

Added — Thinking Animation

• Braille and emoji spinners during model "thinking" phase. Auto-detected from the thinking field in the streaming response.

Added — Token / Cost / Context Bar

• Persistent bar showing live token count, estimated USD cost, and context-window usage percentage. Updates during streaming.

Added — Git TUI

• Diff panel (internal/tui/git/diff.go): side-by-side diff view with syntax highlighting. Ctrl+D opens.
• Commit flow (internal/tui/git/commit.go): interactive commit message editor + staging/unstaging via Space.
• Log view (internal/tui/git/log.go): paginated git log with Enter to inspect a commit's full diff.
• PR creation (internal/tui/git/pr.go): creates a PR via the gh bridge with an interactive title/body editor. 12 tests.

Added — LSP TUI

• Diagnostics panel (internal/tui/lsp/diagnostics.go): live LSP diagnostics with severity icons. Ctrl+L opens.
• Go-to-definition — gd in the file viewer jumps to the definition via gopls / pyright / tsserver.
• Find references — gr in the file viewer shows all references in a popup list.
• Inline markers — diagnostics rendered inline in the file viewer with squiggle underlines.
• Status bar — LSP server status (connected / error / off) in the status line. 10 tests.

Added — /init Command

• /init — detects project type (Go, Python, Node, Rust, generic), scans for existing docs, and generates a starter AGENTS.md with the correct module path, build commands, and test commands. Idempotent — won't overwrite an existing AGENTS.md without confirmation.

Added — Watch Mode

• sin-code chat --watch — polling-based file watcher that re-runs the last prompt when watched files change. Configurable via --watch-glob (default **/*.go) and --watch-debounce (default 500 ms).

Added — Docker Test Guards (#254)

• Docker-dependent tests now skip in -short mode via testing.Short(). CI without Docker (the n8n VM path, M1) no longer fails on container-based tests.

Added — Eval User Guide + 4 Golden Datasets (#255, #258)

• docs/eval.md — user-facing guide for sin-code eval with examples for each scorer, the four-arm comparator, and tracing setup.
• 4 new golden datasets in evals/:
  • test-generation.json — test-generation cases (#256)
  • coding.json — compile-and-run scorer cases
  • trivial.json — YAGNI one-liner cases
  • fusion-tournament.json — Fusion tournament eval cases

Added — LLM Testgen Repair Loop (#256)

• internal/evalharness/testgen.go — Generate → Execute → Repair loop for LLM-backed test generation. When generated tests fail to compile or pass, the failure is fed back to the LLM for repair. Bounded by --max-repair-rounds (default 3).

Added — Performance Benchmarks (10)

• 10 benchmarks covering hot paths:
  • PromptCache hit: 35 ns, 0 allocs
  • PatternDB.MatchPrompt: 198 ns, 0 allocs
  • DeepPlanner.BuildDAGPlan: 1.1 µs, 14 allocs
  • DAGDispatcher event round-trip: 280 ns, 1 alloc
  • LazyLoader.ToolSearch: 890 ns, 2 allocs
  • RiskClassifier.Classify: 45 ns, 0 allocs
  • Compaction.Summarize: 2.3 µs, 18 allocs
  • SyntaxHighlight.Go: 1.4 µs, 3 allocs
  • Keymap.Lookup: 22 ns, 0 allocs
  • StatusLine.Render: 310 ns, 1 alloc

Added — E2E Pipeline Integration Tests (8)

• tests/e2e/ — 8 end-to-end tests exercising the full pipeline (chat → tool call → verify → learn) with real agent-loop instances (no mocks for the loop itself):
  • TestE2E_ChatToVerify_PoC
  • TestE2E_ChatToVerify_Oracle
  • TestE2E_SubagentSpawn_VerifyGate
  • TestE2E_Fusion_Tournament_Winner
  • TestE2E_LazyToolLoad_SearchAndInvoke
  • TestE2E_Compaction_SlidingWindow
  • TestE2E_PatternLearn_RoundTrip
  • TestE2E_Undercover_CommitStrip

Added — All Standalone Components Wired Into Execution Path

• Every standalone TUI component (file browser, diff, DAG, dashboard, search, LSP, git) is wired into the main execution path. No dead code — every component is reachable via keyboard shortcuts or slash commands.

v3.20.0 Stats

Metric	Value
New features	~50
New tests	~500 (all -race clean, M7)
TUI code	33.5K lines
Packages passing	87
Goroutine leaks	0 (stress-tested with 100-node DAGs)
PromptCache hit	35 ns, 0 allocs
PatternDB.MatchPrompt	198 ns, 0 allocs
DeepPlanner.BuildDAGPlan	1.1 µs, 14 allocs
Lazy tool loading	134K → 5K tokens

v3.20.0 Mandate Compliance

• M1 n8n-CI only ✓ (no GitHub Actions runners for build/test)
• M2 CGo-free, single static binary ✓ (no new runtime deps; syntax highlighter is hand-rolled, zero deps)
• M3 Verification gate sacred ✓ (LoopAgent runs full verify; autoDream never runs during active verify; Fusion tournament is PoC-only)
• M4 Permission engine ✓ (YOLO risk classifier adds graded auto-approve; --undercover never bypasses permission engine)
• M5 Module path github.com/OpenSIN-Code/SIN-Code ✓
• M6 SIN tools over naive built-ins ✓ (lazy loader preserves SIN tool surface; keymap shortcuts route to sin_* tools)
• M7 Race-free ✓ (all ~500 tests pass -race -count=1; dispatcher, subagent spawner, prewarm, and autoDream stress-tested with 100-node DAGs and 1000-pattern DBs)

Added — image-graph + skill bundling

• sin-code image-graph — 42nd subcommand: SOTA chart generation with Apache ECharts (bar, line, pie, area). Direct ECharts JSON generation (no wrapper library). Gradients, glow shadows, staggered animations, dark theme, interactive HTML + PNG output. Zero external Go dependencies.
• skill-github-readme — 35th bundled skill: Enterprise visual standard for repository READMEs. Mermaid diagrams, 3-second-hook psychology, CI badges, llms.txt, social preview. Ported from Infra-SIN-OpenCode-Stack/visual-repo. Coupled with skill-github-governance, integrates with sin-image-graph.
• skill-code-graph — bundled sin-image-graph skill under skills/code-skills/
• Cross-reference chain: skill-github-governance → skill-github-readme → sin-image-graph

Changed

• skill-github-governance updated: added coupling references to skill-github-readme and sin-image-graph
• Renamed in Infra-SIN-OpenCode-Stack: sovereign-repo-governance → skill-github-governance, visual-repo → skill-github-readme (breaking rename for findability)
• sin-image-graph SKILL.md updated to v2.0.0: documents ECharts features (gradients, glow, animations), new JSON examples, comparison table vs sin-image-generation
• sin-image-generation SKILL.md rewritten: removed Nano Banana 2 (restricted), added FLUX.2 Max + Imagen 4 Ultra as primary image models
• opencode config updated: nano-banana-2 model → flux-2-max + imagen-4-ultra models

Removed

• go-echarts/v2 dependency — replaced with direct ECharts JSON generation (zero external charting deps)
• --theme flag from image-graph command (fixed dark theme, flag was dead)

Added — SIN Fusion v1: Verify-Tournament (issue #290)

• internal/fusion/ — new package: multi-model verify-tournament for verify-fail recovery
• When the sacred verify-gate (M3) fails, the task is fanned out to N Fireworks models in parallel (MiniMax M3, Kimi K2.7 Code Fast, Kimi K2.7 Code, DeepSeek V4 Pro, Qwen 3.7 Plus, GLM 5.2). First PoC-pass wins; losers cancelled via context.WithCancel
• fireworks_pool.go — 6-model Fireworks lineup, all via SINator pool router, thinking/reasoning mode enabled
• difficulty.go — ShouldTournament() classifies verify-fail as structural (→ tournament) or stylistic (→ legacy retry); DifficultyInput struct with orchestrator confidence signals; 6-step decision order
• tournament.go — goroutine fan-out + buffered channel + first-pass-wins + deterministic tie-break (cost → latency → provider name) + cost-governor (USD kill-switch) + quorum check + 10ms timed drain
• provider_pool.go — loads profiles/*.toml as tournament participants (generic fallback)
• agentloop/loop.go — TournamentRunner interface wired into verify-fail path; PoC-mode only (not oracle — load-bearing risk)
• provider_adapter.go — NewProviderCompletionWithThinking() sends {"thinking":{"type":"enabled"}} for Fireworks models
• loopbuilder/builder.go — SessionStore field + WireFusion() exported function; per-provider llm.Client + Loop.Run() with thinking; config-file defaults auto-load
• config.go — 6 new config keys: fusion.enabled, fusion.providers, fusion.max_cost_usd, fusion.min_quorum, fusion.per_provider_timeout_s, fusion.difficulty_gate
• hooks/hooks.go — FusionDispatch = "fusion.dispatch" event
• permission_defaults.go — fusion__tournament = ask, fusion__status/fusion__config = allow
• ledger/store.go — TypeFusionTournament constant + enriched data (providers_count, winner_cost_usd, winner_duration_ms, winner_verified)
• evalharness — FusionArm() constructor + FirstToPassRate metric in Totals + snapshot support
• CLI flags: --fusion-on-verify-fail, --fusion-providers, --fusion-max-cost (chat + daemon)
• Config get/set/show/validate handlers for 6 fusion.* keys
• 264 tests, all -race clean (mandate M7)
• fusion.enabled = false (default) → exact legacy behavior, zero regression

Added — SIN Fusion Oracle-mode tournament (issue #344)

• internal/fusion/oracle.go — Mode/ModeOracle + OracleJudgeFn + LLMOracleJudge with structured JSON rubric (correctness/completeness/risk), randomized candidate order, and deterministic score tie-break
• tournament.go — Tournament.Mode + OracleJudge fields; Run() dispatches to runPoC() or runOracle(); oracle mode runs all candidates to completion, then judges all outputs together
• agentloop/loop.go — tournament now active in both PoC and oracle modes when configured
• loopbuilder/builder.go — FusionOracleMode config + oracle tournament wiring; tighter default $2.00 cost cap for oracle mode
• config.go — fusion.oracle_mode config key with get/set/show/validate handlers
• permission_defaults.go — fusion__oracle_tournament = ask (M4)
• Oracle-mode tests: judge correctness, all-run-to-completion, cost-cap, judge-error, race safety, position bias, nil judge, JSON/markdown verdict parsing, score clamping

Added — Runtime DB migration to user-config-dir (issue #265)

• cmd/sin-code/internal/tui/agent_runner.go no longer writes lessons.db / sessions.db directly inside the workspace's .sin-code/ directory. The TUI agent runner now resolves its DB home through internal/tui/dbhome.go:ResolveDBHome:
  • default → os.UserConfigDir()/sin-code/workspaces/<sha256-prefix12(abs(ws))>/
  • Config.DBHome != "" → that path (used by tests)
  • per-workspace isolation keyed by a stable 12-char sha256 of abs(Workspace) so two projects never collide on the same row
• New Config.DBHome string field on the TUI runner; empty is the default (UserConfigDir()).
• Workspace trees are no longer touched by sin-code chat — the legacy .gitignore rule for cmd/sin-code/tui/.sin-code/ becomes defence-in-depth only.
• 12 new tests in cmd/sin-code/internal/tui/dbhome_test.go cover the home resolution (DefaultPath, DBHome override, empty workspace rejection, error propagation from os.UserConfigDir, byte-stability, workspace-hash case-folding for macOS) plus the agent_runner_test.go TestNewAgentRunner_UsesDBHomeWhenSet and TestNewAgentRunner_DefaultsAreUserScoped integration tests.
• Existing TestNewAgentRunnerCreatesSinCodeDir flipped to assert that .sin-code/ does NOT exist under the workspace (the migration removes the leak). Legacy TestNewAgentRunnerErrorPaths (workspace-as-file rejection) was likewise retired — the old code path is gone.
• AGENTS.md §11.1 updated to reflect the new layout.

Added — Tool coverage & SIN-tool preference enforcement

• Agent profile tool surface (issue #249) — bundled profiles/fireworks.toml and profiles/qwen-relay.toml now set tools_allow to the full sin_* surface plus all registered MCP prefixes (sckg_*, oracle_*, poc_*, websearch__*, browser__*, gh_query, gh_health, etc.). Destructive SIN builtins (sin_bash, sin_git_commit, sin_test_generate, sin_browser_navigate) remain intentionally omitted so they stay at the default ask tier and are still gated by the permission engine in headless mode (M4).
• M6 system-prompt enforcement (issue #253) — internal/style/style.go now injects a mandatory # Tool preference (mandate M6) fragment into every rendered system prompt via RenderSystemPrompt. The block instructs the model to prefer specialized SIN tools (sin_scout, sin_map, sin_grasp, sin_sckg, sin_poc, sin_oracle, sin_efm, sin_security_scan, sin_sbom_generate, sin_adw) over naive shell or generic file-read equivalents, and requires justification before any sin_bash call.
  • 6 new tests in cmd/sin-code/internal/style/style_test.go cover the tool-preference block (non-empty, expected tools, forbidden naive equivalents, byte-hash stability, inclusion in all modes, and style-block-only behavior).
• Runtime tool coverage enforcer (issue #248) — new cmd/sin-code/internal/agentloop/toolcoverage.go with ToolCoverageEnforcer. The enforcer is created per-run when agentloop.required_tools or agentloop.forbidden_tools is configured, records every tool invocation under a mutex, and fails completion if any required tool is missing or any forbidden tool was used. Failures are fed back into the stop-gate as open criteria.
  • 10 new tests in cmd/sin-code/internal/agentloop/toolcoverage_test.go cover constraints, missing required, forbidden used, both, used-set deduplication, open-criteria, feedback, and race safety.
  • 3 new loop-level tests in cmd/sin-code/internal/agentloop/loop_test.go (TestRun_CoverageRequiredTool_ForcesInvocation, TestRun_CoverageForbiddenTool_BlocksCompletion, TestRun_CoverageRequiredTool_ImmediatePass).
  • 1 new config roundtrip test in cmd/sin-code/internal/config_test.go (TestConfig_ToolCoverageRoundtrip).
  • Exposed via CLI flags --require-tools / --forbid-tools on chat, daemon, and execute/orchestrate paths, and via config keys agentloop.required_tools / agentloop.forbidden_tools.
• Tool-usage telemetry (issue #250) — cmd/sin-code/internal/ledger/usage.go adds RecordUsage, ToolUsageCounts, FamilyUsageCounts, ToolCoverage, UnusedTools, and ToolUsageByPeriod to support heatmap, coverage, and unused-tool queries. cmd/sin-code/ledger_cmd.go adds sin-code ledger tools with --heatmap, --coverage, --unused, --family, and --json flags.
  • 7 new tests in cmd/sin-code/internal/ledger/usage_test.go cover heatmap, family counts, coverage/unused, period bucketing, field validation, race safety, and since/until filtering.
• Orchestrator ToolChain (issue #252) — new cmd/sin-code/internal/orchestrator/toolchain.go with ToolChain and Intent-to-tools mapping. The planner attaches a deterministic ToolChain to every Plan for intents: security, review, architecture, test, codebase, and docs. Required tools are derived from canonical names such as sin_security_scan, sin_sbom_generate, sin_oracle, sin_adw, sin_poc, sin_map, sin_sckg, sin_test, sin_scout, and sin_read.
  • 13 new tests in cmd/sin-code/internal/orchestrator/toolchain_test.go cover per-intent mapping, empty chains for general/unknown intents, copy independence, RequiredToolsForIntent, JSON serialization, and planner determinism.
• Unified tool catalog (issue #251) — cmd/sin-code/internal/catalog/ now merges HubSource, MCPSource, ChatSource, and ExternalSource into one Asset catalog. The catalog enumerates 46+ MCP tools, 17+ chat tools, and 14+ external MCP prefixes. cmd/sin-code/catalog_cmd.go adds sin-code catalog with list, search, info, and unused subcommands, plus sin-code hub search --unused.
  • 46 new tests in cmd/sin-code/internal/catalog/catalog_test.go cover merge, de-duplication, search, filtering, per-source lists, and unused filtering.
  • cmd/sin-code/testdata/scripts/golden_help.txt regenerated to match the current help output.

Added — Shop-Center skill integration (issue #142 fusion)

• KnownSkills() registry extended with the three shop skills (issue #142 acceptance criterion #2 — installable via sin-code skill install <name>):
  • shop-cj-dropshipping → cj-dropshipping-skill
  • shop-stripe → SIN-Stripe-Bundle
  • shop-tiktok → SIN-eCommerce-Scraper-Bundle
• Bundled SKILL.md files (cj-dropshipping, stripe, tiktok) already carry lifecycle: external + sources: frontmatter (added by PR #218 / issue #139). The sources: field points to the canonical external repos so the operator can discover the upstream implementation directly from the bundled skill.
• New test TestKnownSkillsHasShopEntries in cmd/sin-code/internal/skillmgr/manager_test.go asserts the three entries are present with the correct repo mapping.
• Validation passes — validate_skill.py --all-bundled --strict reports 0 failed for the 34 skills (the 3 shop skills were migrated by PR #218).
• Long-term fusion strategy documented in the issue body: phase 1 (external canonical) → phase 2 (bundled doc, done) → phase 3 (native subcommand) → phase 4 (deprecate upstream). Phase 3 is deferred until the shop domain matures.

Added — sin-code grill (issue #141 fusion, native implementation)

• New package cmd/sin-code/internal/grill/ with 4 source files (types.go, catalog.go, manager.go, grill_test.go)
  • 14 race-clean unit tests. The native Go implementation of the external SIN-Code-Grill-Me-Skill Python MCP server (38 KB). Ships in v0 with JSON-file storage; SQLite session storage is a v1 follow-up.
• New subcommand sin-code grill (5 subcommands):
  • grill start <topic> — begin a grilling session, print the id
  • grill next <id> — ask the next adversarial question
  • grill answer <id> <d-id> <text> — record the response (use "done" to resolve a decision)
  • grill status <id> — show resolved + open decisions
  • grill synthesize <id> [--json] — produce a structured summary of decisions, assumptions, and open questions
• Question catalog — 8 seed anti-patterns (Hidden Assumptions, Rollback Plan, Failure Modes, Operator Cost, Premature Optimization, Scope Creep, Single Point of Failure, Verification Gap). Each has 2-3 example questions. The CLI picks one per grill next call (hash-seeded for determinism).
• Storage — $SIN_CODE_HOME/grill/<id>.json, atomic writes (temp + rename). v1 will migrate to SQLite via the existing internal/session/store.
• 14 race-clean tests covering the full flow (Start → Next → Answer → Synthesize → round-trip across restarts).

Added — sin-code goal fusion (issue #140, v0.5)

• Four new subcommands under sin-code goal (issue #140 fusion with the external SIN-Code-Goal-Mode-Skill Python MCP server):
  • goal status <id> — show one goal with subtasks (children)
  • goal complete <id> — mark a goal as verified/done
  • goal subtask <parent-id> <prompt> — add a subtask
  • goal report [--format md|json] — progress report
• Mapping to the 8 external tools (issue body):
  • goal_start → existing goal add
  • goal_status → NEW goal status
  • goal_list → existing goal list
  • goal_complete → NEW goal complete
  • goal_subtask → NEW goal subtask
  • goal_report → NEW goal report
  • goal_checkpoint / goal_rollback — deferred to v1 (no storage yet; the Queue has no Checkpoint table)
• parseGoalID helper — accepts both 42 and #42 (with optional whitespace); used by all four new subcommands
• 11 tests (1 unit test for parseGoalID, all autonomy tests pass under go test -race -count=1)

Added — Skill lifecycle markers (issue #139)

• scripts/lifecycle_map.yaml — single source of truth for the lifecycle of every bundled skill. Maps each of the 34 skills to one of native | external | deprecated with a canonical: field pointing to the upstream implementation.
• scripts/sync_lifecycle.py — stdlib-only Python script. Three modes: --check (CI: exit 1 if any drift), --apply (write changes), --diff (show what would change). Hand-rolled YAML parser for the map file (no PyYAML dep).
• scripts/validate_skill.py strict mode — now requires the lifecycle frontmatter key in --strict mode and validates the value. Non-strict mode remains backward-compatible.
• sin-code skill list now prints a LIFECYCLE column with [native], [external], [deprecated], or [unknown] markers. A new parseLifecycleFromFrontmatter helper extracts the field from the embedded SKILL.md without a yaml dep.
• docs/SKILLS.md — design doc with the value table, the workflow, and the migration path.
• All 34 SKILL.md files migrated — 28 skills received the lifecycle: field; 6 were already in sync. Total: 34/34.

Added — internal/testutil/ (issue #161, race-flake hardening v2)

• Five reusable test helpers in a stdlib-only package:
  • IsolatedSQLite(t) — fresh t.TempDir()-backed *sql.DB, auto-closed
  • CleanEnv(t, kv) — set + restore env vars via t.Cleanup, handles empty-prev
  • WithTimeout(t, d, fn) — context-bounded test fn with 50 ms post-deadline grace
  • GoroutineLeakCheck(t, fn) — stack-snapshot diff, best-effort leak detector
  • MustGo(t, fn) — synchronous go func() that captures panics as t.Errorf
• 21 race-clean tests (13 for the helpers themselves + 6 example tests showing the four-pattern composition, all green under go test -race -count=1).
• testutil.doc.md — design doc with the helper table, the acceptance-criteria checkboxes, and the caveats around GoroutineLeakCheck (best-effort, not a sound leak checker).
• Diagnosis pass (informational): ran go test -count=1 -v ./... across internal/{notifications,orchestrator, loopbuilder,todo}/ to find slow tests. The slowest is TestGenerateIDUniqueness at 3.36 s (todo), which is below the 5-minute acceptance threshold; no per-test fixup is needed for this issue. The diagnosis methodology is in the runbook below.

Added — sin-code triage (issue #162)

• cmd/sin-code/triage_cmd.go — new 41st subcommand sin-code triage (and triage --format=md|json --repo owner/repo --limit N). Reads the open issue backlog via gh issue list through the ghbridge wrapper, scores each issue with a deterministic heuristic (epic +10, blocks +5 per ref, acceptance +3, not-in-v0 +5, loop-system +4, fusion +2, fresh -2, stale +1, good-first-issue -3), groups by label bucket, and renders. The markdown output is the canonical BACKLOG.md generator.
• cmd/sin-code/internal/triage/ — new package, four files (types.go, score.go, render.go, loader.go) plus triage_test.go (15 tests, all green under -race -count=1). The loader is a var so tests inject fixtures without spawning gh. No new third-party deps (M2).
• triage.doc.md — design doc with the scoring table, the per-bucket ordering rule, and the deferred-items list.

Added — sin-code catalog (issue #163, hub-assets merge)

• cmd/sin-code/catalog_cmd.go — new subcommand sin-code catalog (list | search | info) with --kind=agent|command|skill|hub and --format=text|json. The unified tool catalog that operators have been asking for: "do I have a tool for this?" — not "do I want the hub or the assets?".
• cmd/sin-code/internal/catalog/ — new package, 4 source files (catalog.go, source_hub.go, source_assets.go, catalog_test.go)
  • 21 race-clean unit tests. The Source interface (Name + List + Get) is the abstraction that lets the catalog walk both backends. Adding a new source (e.g. a remote registry) is one file.
• Merge de-duplication rule — first source to provide a (kind, name) pair wins; subsequent duplicates are dropped. The source name is intentionally not part of the dedup key, so a hub.Tool and an assets.Asset with the same name are merged into one catalog entry (the SOTA choice for the operator's mental model).
• Search ranking — name +4, short +2, description +1, tag +1; ties break by name ascending. Transparent, auditable, deterministic.
• catalog.doc.md — design doc with the de-dup table, the scoring heuristic, the deprecation plan for sin-code hub, and the known build issue (Chromedp API mismatch in PR #201, not in this PR).

Added — internal/rag/ (issue #160, RAG over instinct store)

• New package cmd/sin-code/internal/rag/ with 4 source files (embedder.go, embedder_hash.go, embedder_onnx.go, index.go, worker.go, retriever.go) + 24 race-clean tests.
  • Embedder interface + HashEmbedder (default, deterministic, dependency-free, 384-dim L2-normalized) + ONNXRuntimeEmbedder and HTTPEmbedder as documented stubs.
  • Index with optional Persister interface; the instinct subsystem uses a jsonPersister writing to $SIN_CODE_HOME/instinct-embeddings.json.
  • WorkerPool (bounded-concurrency, M7) for async embedding so the agent loop never blocks.
  • Retriever (high-level: Embedder + Index → top-N IDs).
• sin instinct search "<query>" — top-5 cosine-similarity search over the active instincts. Reindexes on every call (cheap at <100 active) and persists to disk. Renders the hits as id — trigger lines with the action underneath.
• 8 race-clean tests in internal/instinct/search_test.go for the JSON persister round-trip, the path-overriding env var, the atomic-write behavior, and the trim helper.
• rag.doc.md — design doc with the mandate-compliance analysis, the Embedder interface, the acceptance-criteria checkboxes, and the deferred-items list (GOAP Planner, Federation, real ONNX implementation).

Added — sin-code compile-spec (issue #164, v0 spike)

• cmd/sin-code/internal/spec/compiler/ — new package, 4 source files (schema.go, parse.go, validate.go, emit.go) + 24 race-clean unit tests. Round-trip test (TestRoundTrip) is the load-bearing guarantee: parse → emit → parse → emit must produce identical bytes.
• cmd/sin-code/compile_spec_cmd.go — new subcommand sin-code compile-spec with --init, --check, --out <dir>, --dry-run flags. Atomic writes (temp + rename) so a crash mid-write never leaves a half-written file behind.
• Four derived JSON outputs (contract defined; engines not yet wired — that is v1.1):
  • .sin/hooks.json
  • internal/verify/config.json
  • internal/permission/policies.json
  • .sin/loop.json (parsed but not consumed — migration path for issue #155)
• SPEC-COMPILER.md — design doc with the schema, the mandate-compliance analysis, the deferred-items list (engine wiring, remote spec inheritance, spec testing), and the relationship to issue #155.

Added — sin-code install + one-line curl|bash installer (issue #170)

• cmd/sin-code/install_cmd.go — new 40th subcommand sin-code install (and install --auto). Downloads the latest GitHub release asset, SHA256-verifies against the goreleaser-style checksums.txt, extracts the single binary, atomically places it into $SIN_CODE_BIN_DIR or $HOME/.local/bin, and prints the canonical PATH hint. Flags: --dir, --release <tag> (pin a version), --channel stable|dev, --verify-only (health-check, no write), --no-verify (offline / sanctioned CI), --dry-run.
• cmd/sin-code/internal/install/ — new pure-stdlib package. Four tiny files (release.go, github.go, verify.go, composer.go) plus race-safe install_test.go (19 tests, all green under -race -count=1). The bootstrap never depends on gh or jq, making the install cmd safe to run on a freshly imaged host.
• Root install.sh — rewritten from 1031 lines to 27 lines shell-only shim. curl -fsSL ... | bash compatible. Settles the downloaded archive, extracts the sin-code binary via three tolerant glob shapes (works across goreleaser versions), then execs sin-code install --auto so the Go entrypoint owns the verify-and-place flow. Legacy 12-step logic permanently retired — post-v3.0 the unified sin-code binary already subsumes the 7 Go tool subcommands the old installer built.
• Root install.ps1 — new 35-line Windows equivalent (irm https://raw.githubusercontent.com/.../install.ps1 | iex). Uses Invoke-WebRequest + System.IO.Compression.ZipFile, then re-execs sin-code.exe install --auto.
• permission_defaults.go — three new MCP rules under the install__* prefix (mirror of the gh_execute precedent in §3 M4): install__verify_only allow, install__dry_run allow, install__run ask. The headless daemon therefore CANNOT self-install silently, satisfying M4's "always headless" clause.
• AGENTS.md §6 — repo layout now lists install.sh, install.ps1, cmd/sin-code/install_cmd.go, and the new internal/install/ package alongside the existing 39 subcommands.

Added — Verbosity / compression mode (issue #167)

• internal/style/ — first-class verbosity mode system-prompt renderer. Five canonical modes (default, verbose, normal, terse, ultra) with byte-stable output per (mode, skillBody) pair (prerequisite for the system-prompt hash metric, issue #2). default and verbose pass through skill bodies unchanged; normal drops pleasantries + tool-call narration, terse drops articles/hedging, ultra is the tightest valid compression. Every non-default ruleset carries the auto-clarity clause that forces normal prose around destructive, security-relevant, or order-sensitive actions — the verification gate (mandate M3) must never be skipped because the output mode is terse. API surface: ParseMode(s), RenderRules(mode, body), RenderSystemBlock(level), AppendVerbosity(existing, mode), WithVerbosity(mode) (functional option).
• instinct.RenderSystemBlockWithVerbosity (issue #167): the instinct renderer now accepts a verbosity-mode string and appends the matching ruleset after the learned-instinct list (stable order: instincts → style, separated by exactly one blank line). Backward- compatible — the legacy RenderSystemBlock(active, max) still returns the bare instinct block.
• learning.Learner style hook: BeforeTurn now honors Options.Style and routes it through the new renderer. New SetStyle(level) method lets mid-session callers toggle verbosity safely under a per-instance sync.RWMutex (mandate M7, race-free). wiring.Deps.Style passes through.
• llm.style config key (internal/config.go): user-facing knob with full get/set/list/validate/TOML/JSON coverage. Validated against the canonical mode set. sin-code config set llm.style terse works end-to-end.
• Reference docs: internal/style/style.doc.md (developer), cmd/sin-code/internal/config.doc.md updated (table + example), cmd/sin-code/internal/learning/learner.doc.md updated (Style field), AGENTS.md §6 + §7 cross-references.

Added — sin-code compress (issue #172, deterministic + LLM compaction)

• internal/compress/ — first-party package implementing the caveman-compress pattern (JuliusBrussee/caveman) for SIN-Code's long-lived stores. Public API: BuildPlan, Apply, Rollback; three strategies (deterministic default, llm, hybrid) targeting four surfaces (lessons, instincts, summaries, memory, agents_md) plus an aggregate all. Plan is read-only and content-addressed (PlanHash covers Entries+Drops+Merges); Apply is atomic (snapshot written to .partial then renamed before any source rewrite) and lossless (dropped entries are preserved verbatim under ~/.local/share/sin-code/compress-snapshots/<plan-id>.json).
• Deterministic pass (compressor.go + deterministic.go): SHA-256 dedupe + utility-sorted (recency × inverse-size) keep-recent with byte-budget cap. Algorithm pins time.Now() for tests via PlanOptions.UseStableTime so two plans built from identical inputs agree byte-for-byte regardless of wall clock. Stable-time pins are verified by TestPlanDeterministicIdempotent.
• LLM summarization pass (llm.go): caveman-style compress prompt that preserves code fences, URLs, file paths, commands, and headings byte-for-byte. Validates the response with a line-based check; retries up to 2 with a targeted patch prompt on validation failure (MaxRetries, configurable). Falls back to deterministic on exhausted retries or when no provider is configured.
• Atomic snapshot+rollback: Apply writes a .partial snapshot first; Rollback(<plan-id>) reads the snapshot, refuses to consume any in-flight .partial, and restores the originals via per-target re-apply. TestApplyIsAtomicAndLossless covers one round-trip.
• CLI surface (compress_cmd.go, 41st subcommand = 40th registered cobra verb because internal.InstinctCmd etc. are in-package additions):
  • sin-code compress plan [--target all] [--strategy deterministic] [--keep-bytes 4096] [--keep N] [--recent-days N] [--json]
  • sin-code compress apply [--dry-run] [--no-llm] [--target ...] [--strategy ...] [--keep-bytes ...] [--json]
  • sin-code compress rollback <snapshot-id>
• Permission policy (permission_defaults.go): {Tool: "compress__plan", Policy: "allow"}, {Tool: "compress__apply", Policy: "ask"} (M4 — destructive), {Tool: "compress__rollback", Policy: "allow"} (restorative only). Wired so any future agent-loop surface that exposes the compressor via MCP is gated correctly.
• Regression tests (compress_test.go + testhelpers_test.go): hash determinism, plan idempotence, byte-budget enforcement, dedupe invariance, atomic-write contract, dry-run no-touching-nothing, partial-marker rollback refusal, preservation line scoping (heading, code fence, URL, file path, command line), snapshot JSON round-trip. Passes go test -race -count=1 ./cmd/sin-code/internal/compress/.
• Snapshot dir: ~/.local/share/sin-code/compress-snapshots/ (overridable via SIN_CODE_SNAPSHOT_DIR). Same form factor as lessons.db / ledger.db per AGENTS.md §7.

Added — Per-agent profile renderer (issue #175)

• Single source of truth at docs/agent-profiles/sin-profile.md (≤80 lines, KISS, hard mandates + working style + subagent contracts + per-agent notes, edits roll out everywhere).
• internal/profile package: targets map mirroring AGENTS.md §10 (claude-code, opencode, gemini, codex, cursor, windsurf, cline, copilot); per-format writers (dir, rule, marker); a byte-stable Render(tgt, body); a Verify(base, body) SHA-256 drift gate; idempotent marker-fence envelopes byte-identical to internal/skilldist (issue #169 covenants preserved).
• sin-code profile subcommand with four verbs:
  • profile show — print the source markdown
  • profile list — print the supported target table (text + --json)
  • profile render <target|all> — write one or all mirrors (idempotent; supports --dry-run for byte/audit preview without touching disk)
  • profile verify — CI gate: refuse on missing/drift; surfaces a 12-char-row table or a JSON envelope for --json
• Permission engine: profile__show / list / verify are registered as allow; profile__render is ask because it touches per-agent dotdirs (mandate M4).
• CI sync: .github/workflows/ceo-audit.yml grew a parallel profile-verify job that builds sin-code, runs profile render all && profile verify, uploads the rendered mirrors as artifacts, and fails the build if any drift surfaces. Mirrors AGENTS.md §6 + §10 — single source of truth in internal/profile/target.go.
• Test coverage: 22 race-tested Go tests pinning the byte-stable contract (golden render, marker-fence idempotency, marker-Fence covenant, Verify pass / missing / drift, write-after-write SHA equality, replace-not-append for stale mirrors).

Added — sin-debt marker convention (issue #177)

SIN-Code adopts ponytail v4.7.0's ponytail: marker convention as a first-class, parseable // sin-debt: <ceiling>, upgrade: <trigger> convention. Every intentional shortcut now carries a marker naming its ceiling and the trigger to revisit; the scanner reads them; debt stats reports them; debt check gates them.

• internal/sindept/ — scanner, aggregator, byte-stable report renderer, and policy gate. Single-package surface:
  • parser.go — Marker{File, Line, Column, Reason, Upgrade, HasUpg, Raw, Language, Symbol}, regex over five comment families (//, #, --, /*…*/, <!--…-->), ParseFile + ParseDir. Trims trailing block-comment closers (*/, -->) and post-processes captured clauses for byte-determinism.
  • stats.go — Stats{Total, WithUpgrade, WithoutUpgrade, ByFile, ByReason, ByLanguage, BySymbol, RotRisk, Oldest, MarkersPerFile}. Every map is materialized as a lex-sorted []KV so Render* output is stable.
  • report.go — RenderStats / RenderStatsString / RenderListString markdown renders with FormatVersion = "sin-debt/v1". Two scans of the same tree emit the same bytes.
  • policy.go — Policy{DefaultReasons, UpgradeTriggers, MaxNoUpgrade, RequireUpgrade, Source}. TOML overlay via .sin-code/debt-policy.toml; LoadPolicyForRoot walks upward to the closest file.
• debt_cmd.go — 41st subcommand. sin-code debt list | stats | check | policy | fix | export. Common flags: --path, --format (table|json), --no-trigger. Stats sub-commands take --by file|reason|language|symbol|age|summary. The check sub-command is the CI gate; it exits non-zero when Missing > MaxNoUpgrade or RequireUpgrade=true && Missing > 0.
• docs/sin-debt-convention.md — author-facing reference: format, examples, default reasons catalogue, default upgrade triggers.
• Permitted sindept__* tools: sindept__list, sindept__stats, sindept__policy are read-only allow; sindept__check exiting non-zero is ask; sindept__fix and sindept__export are ask because they instruct humans to edit code or write a file.
• 10 hard-coded fixture markers under cmd/sin-code/internal/sindept/testdata/ (5 languages) and 4 real markers placed in production code (cmd/sin-code/internal/lessons/ store.go × 2, …/ledger/store.go, …/orchestrator/dispatcher.go).
• Tests: 23 tests in cmd/sin-code/internal/sindept/sindept_test.go cover family coverage, trailing-closer stripping, byte-stability, vendor / hidden-dir walk, age/rot grouping, and the policy-gate semantics above vs. below threshold. Race-clean.

Notes

• sin-code debt stats is the precondition for the four-arm comparator snapshot (issue #171); byte-stable today, golden file is expected in the next PR cycle.
• The marker syntax deliberately does NOT include \Q…\E quoting — RE2 has no such construct, and the literal token sin-debt: is plain inside the regex.
• internal/sindept is the upstream of issue #179 (complexity auditor) and issue #180 (audit-engine): both are expected to call sindept.ParseFile / sindept.AggregateStats so a marker reads through the same shape regardless of consumer.

---

Added — Auto-Activation Hook (issue #176, v3.19.0)

• internal/hooklife/autoactivate/ — per-session rule injection subpackage. Two Phase hooks (autoactivate-session-start / autoactivate-user-prompt) register against any *hooklife.Registry via Activator.Register(reg). Privacy-first: off by default; activated by sin-code chat --activate <rule> or a project-local .sin-code/autoactivate.toml file.
• AutoActivate.Activator tracks per-session state under a single sync.RWMutex (mandate M7 — race-safe under go test -race -count=1). OnSessionStart(sid, opts) is idempotent; OnUserPrompt(sid, prompt) returns the rule set re-emittable for this turn, with trigger-phrase substring matching for natural-language activation. EndSession(sid) drops state on exit.
• RuleSet.Render() is byte-stable: any two RuleSets with the same name+body+trigger tuples produce identical bytes regardless of insertion order (prerequisite for the system-prompt hash metric, issue #2).
• sin-code chat --activate terse,skill-x comma-separated rule list; --no-trigger suppresses per-prompt phrase matching; reads .sin-code/autoactivate.toml silently when present.
• Tests: 35 race-safe unit tests + 8 chat-wiring integration tests, 91.5% statement coverage on the autoactivate package. New package follows the existing hooklife Phase contract; no new external deps.

Added — Orchestrator output contracts (issue #174)

• Caveman-style output contract for the four orchestrator sub-agents (internal/orchestrator/output_contract.go). Every Finding renders to ONE byte-stable line: <path>:<line> — <symbol> — <tag> — <hint> # c=<confidence>. Five closed tags (parallel to JuliusBrussee/ponytail): delete | simplify | rebuild | risk | verify. Em-dash U+2014 separator; no prose, no pleasantries, no hedging (you might, perhaps, could consider, maybe, i think, sort of, should probably, … — closed set of 12 phrases, case-folded).
• Finding struct (internal/orchestrator) and ParseFinding / ParseFindings regex parsers with strict byte-stability. Render() is fully deterministic — Finding{...} → Render() → ParseFinding() → equal struct, every byte counted (verified by TestParseFinding_RoundTrip).
• VerifyFindings runs the full contract: structural (Path != "", Tag ∈ {delete, simplify, rebuild, risk, verify}), lexical (zero hedging, hint ≤ 240 chars, no trailing punctuation), and emits per-Finding error strings — never silent drops.
• Wired into the four sub-agents:
  • Critic.Drive parses the LAST attempt's prose into CriticResult.Findings and surfaces CriticResult.ParseErrors for the orchestrator to re-inject as retry feedback (mirrors the verify.fail flow).
  • Adversary.Review derives Findings from the structured Attack slice (landed → risk, cleared → verify); the CounterexampleBrief free- form prose is preserved as the audit trail.
  • Governor.Execute derives one risk Finding per Escalation (Path = task://<ID>, Symbol = <from>-><to>); the prose Reason stays on Escalation for the audit log.
  • Cartographer.Findings(k) exposes the PageRank-sorted top-k as verify-tagged Findings (opt-in: k ≤ 0 yields the empty slice).
• Byte-stable golden tests (output_contract_test.go and output_contract_integration_test.go) pin one fixture per agent; rendering drift breaks the build (the prerequisite for issue #168's ledger-level token-cost hashing).

Added — SinCode Loop System (always-on Definition-of-Done baseline)

• Baseline DoD (internal/goalcontract/baseline.go): an always-on Definition-of-Done that is merged into every resolved contract so the "self-evident" follow-through work is implicit for every goal — the user never has to ask an agent to write tests, debug, remove scaffolding, or update docs again. It adds deterministic predicate checks (tests touched in the diff, CHANGELOG.md touched, a .doc.md CoDoc beside each changed Go file) plus LLM-judged semantic criteria (tests cover new behavior, no debug leftovers, goal fully addressed, README/CHANGELOG/AGENTS/MASTER_TODO/CoDocs kept in sync). Additive and deduped against auto-detected/explicit criteria.
• DoD preamble injection (agentloop.Loop.Preamble + goalcontract.Preamble): the loop now states the rubric to the worker up front via loopbuilder, so tests/debug/docs/completeness are handled on the first pass instead of after a stop-gate rejection. Advisory only; the stop-gate still enforces.
• Always-on, globally escapable: baseline is ON by default in daemon and auto run; disable per-invocation with --no-baseline or globally with SIN_BASELINE=off (goalcontract.BaselineEnabled). Fail-open: predicate scripts exit 0 outside a git repo so they never wedge a non-repo workspace.

Added — Loop Engineering (decoupled completion authority)

Added — MCP tool-manifest compression (issue #173, v3.19.0)

• internal/mcpcompress/ — ponytail-tag compressor for sin-code serve --compress-tools (issue #173). Five canonical tags drive five byte-stable Rules:
  • delete → DeleteHedges drops pleasantries / hedge adverbs ("safely", "carefully", …).
  • stdlib → StdlibPatterns drops redundant stdlib parentheticals ("(via stdlib)", Go stdlib …).
  • native → DropTrimEncouragement drops M6 tail clauses Always prefer over native X / Prefer sin_X over native Y (the M6 mandate is internal-only, not for the model).
  • yagni → YagniPatterns drops speculative (experimental) / (TBD) / (reserved) parentheticals.
  • shrink → ShrinkExamples drops redundant (e.g. …) / (such as …) parenthetical examples.
• Three new serve flags:
  • --compress-tools — apply the full default ponytail tag set (delete|stdlib|native|yagni|shrink) before registration.
  • --compress-tags <csv> — override the active tag set (e.g. --compress-tags "delete,yagni,shrink"). Unknown tags are silently dropped; the active set is logged via --print-stats.
  • --print-stats — emit a left-aligned text table to stderr (tool / orig / comp / saved / ratio + TOTAL row + active rules).
• Tool names unchanged. The compressor mutates only Description; the 47 MCP tool Name fields (public API per AGENTS.md §10) are never modified. TestCompressSpec_NameMutable guards this.
• Byte-stable per (tool_spec, ruleset) — every Rule, the Pipeline, and the post-pipeline Normalize are deterministic and idempotent. The compressor_test.go golden suite is the single source of truth — any Rule regex / declaration-order change must update the gold expectations in the same PR. Prerequisite for the system-prompt hash metric (issue #2).
• Real-world savings. Smoke test against the 47-tool registry: sin_execute 80→73 bytes (-7, 8.8%); sin_read 200→167 bytes (-33, 16.5%); sin_write 194→160 bytes (-34, 17.5%); sin_edit 474→441 bytes (-33, 7.0%). TOTAL 3977→3870 bytes saved across 4 affected tools (-107 bytes / 2.7%); the other 43 tools' descriptions don't match any of the conservative patterns and stay byte-identical. Per-tool --print-stats output is deterministic across runs (no time, no random).

Added — Loop Engineering (decoupled completion authority)### Added — Loop Engineering (decoupled completion authority)

• Stop-gate harness (internal/stopgate): an independent completion authority consulted after the verify-gate passes. Hybrid mode runs deterministic checks first (fail-closed) then a strong/equal LLM judge (SIN_EVALUATOR_MODEL) for non-mechanical criteria; a green judge can never override a red deterministic check. Rejection forces the loop to keep working with the open criteria injected back in. 92.5% test coverage.
• Goal contracts / Definition-of-Done (internal/goalcontract): machine- checkable acceptance criteria per goal, layered resolution (explicit file + inline --criteria + --done-when, auto-detected Go checks incl. a no-new-todos diff guard, verify-cmd fallback). Persisted in the queue's contract column. goal add --criteria/--contract-file. 94.5% coverage.
• Continuation instead of hard abort (agentloop): with AllowContinuation, hitting max-turns now checkpoints and returns a resumable Result (Continuation=true) instead of erroring. The daemon re-enqueues via queue.Continue (refunds the attempt, bumps a continuations counter) bounded by --max-continuations. Long tasks never need a human restart.
• Recursive goal decomposition (autonomy.Queue): parent_id/depth columns, AddSub, depth-first draining (a parent only finalizes once every child verifies, via Complete→blocked→TryFinalize/bubbleUp). The daemon exposes a spawn_subgoal tool bounded by --max-depth.
• Autonomous backlog discovery (autonomy/discover.go): scans TODO/FIXME/ XXX/HACK markers and unchecked MASTER_TODO.md items into deduplicated goals (AddDiscovered with a dedup_key). Exposed as a discover trigger type and the goal discover [--dry-run] command — the agent finds its own work.

Added — Learning Subsystem (continuous learning in Go)

• internal/instinct/ — continuous-learning subsystem (port of the continuous-learning-v2 homunculus model in a clean-room Go reimplementation). Project-scoped + global Markdown-with-frontmatter store, confidence 0.3–0.9, Reinforce/Contradict/Decay math with atomic.Value-backed env-overridable tuning, heuristic + LLM-backed extractors (with graceful fallback), cross-project promotion, cluster evolution into Skill/Command/Agent proposals, and a system-prompt block renderer that closes the learning loop. CLI: sin instinct status|projects|evolve|promote|prune|export|import|show|forget|history. Storage: $SIN_INSTINCT_DIR | $XDG_DATA_HOME/sin-code/instinct | ~/.local/share/sin-code/instinct.
• internal/hooklife/ — native Go lifecycle-hook system (no Node dependency). Phases: PreToolUse, PostToolUse, Stop, SessionStart, SessionEnd, PreCompact, UserPrompt. PreToolUse may Block (ECC exit-code-2 equivalent); other phases aggregate warnings. Per-hook timeout, panic recovery. Built-in hooks: block-no-verify, config-protection, post-edit-format, quality-gate (against the real verify.Gate), cost-tracker, suggest-compact. CLI: sin hooks list|test.
• internal/assets/ — harvested agent/command/skill loader with schema validation (port of ECC CI validators, including unsafe-unicode and duplicate detection), Selector for domain+keyword-based ranking, and an import subcommand that harvests skills from a vendored source repo with origin/license attribution. CLI: sin assets list|validate|show|import.
• internal/evalharness/ — eval-driven development. EvalSet / Run / Result types, pluggable Scorers (exact, contains-all, success-flag, LLM-judge, composite, CompileAndRun), per-case timeout, JSONL run history, and Compare for case-by-case regression detection with --fail-on-regress as a CI gate. CLI: sin eval run|list|compare.
• CompileAndRun scorer (issue #181) — ponytail correctness.js analog for SIN-Code. Extracts fenced code from model output, compiles it (go/python/javascript/bash), and runs a sandboxed self-check. Returns 1.0 only when compile + run pass; skip_test mode accepts trivial one-liners after compile-only (YAGNI for tests). Wired into sin-code eval run via --scorer compile-and-run --language <lang> and into Golden Datasets via test_cases[].scorer.
• internal/dispatch/ — turns loaded command and agent assets into executable actions. ECC-style placeholder substitution ($ARGUMENTS, $1..$9, $@, ${flag}), Dispatcher routes slash-commands to PromptSink and agent requests to SubagentRunner. Closes the load → select → dispatch → run pipeline.
• internal/prp/ — Product Requirement Prompt workflow. Persistent reviewable plans under .sin/prp/<id>.md driven through phases (draft → planned → implementing → verifying → ready → shipped). Each step persists, so a run is interruptible and resumable. Verification failure kicks the PRP back to implementing. CLI: sin prp new|run|status|plan|implement|verify|pr.
• internal/adapters/ — concrete adapters that implement the abstract hooklife.Verifier, instinct.MemorySink, and instinct.Completer interfaces against the real SIN-Code subsystems (verify.Gate, memory.Store, llm.Client). Fail-soft: missing subsystems degrade to no-ops, never block startup.
• internal/learning/ — bridge package between agentloop.Loop and the new subsystems. Learner exposes BeforeTurn (prepends active-instinct system block), BeforeTool (PreToolUse dispatch, may veto), AfterTool (PostToolUse + observer feed), EndTurn (observer flush), PreCompact (flush + hook dispatch). Built once at startup via learning.New(Options).
• internal/wiring/ — Build(Deps) assembles the full Bundle{Learner, Dispatch, Eval factory, PRP deps} in one call.
• examples/eval-sets/ — go-quality.json (build, vet, test, secrets scan) and instinct-behavior.json (end-to-end learning loop validation).
• Five new top-level subcommands wired into cmd/sin-code/main.go: sin instinct, sin hooks, sin assets, sin evalset, sin prp. (The existing sin eval — Golden-Dataset runner from issue #75 — is preserved unchanged; the new harness lives at sin evalset to avoid a cobra Use: collision.)

Added — Complexity Audit (issue #180)

• sin-code audit complexity — repo-wide ponytail-audit analog. Five tags (delete, stdlib, native, yagni, shrink), deterministic static pass (single-impl interfaces, single-product factories, wrapper functions, one-export files, dead flags/config, hand-rolled stdlib), optional LLM judge for top-N findings. Output: one-liner per finding ending with net: -<N> lines, -<M> deps possible. or Lean already. Ship..
• cmd/sin-code/internal/audit/ — new Auditor, Finding, Result types; // sin-debt: markers approve findings and exclude them from the net total.
• sin-code ceo-audit — new 48-gate CEO-grade audit. The 48th gate is the complexity audit; score contribution is +1 per 100 removable lines.
• Docs docs/complexity-audit.md and tests complexity_test.go + audit_cmd_test.go with race-free coverage.

Notes

• All loop-engineering features are opt-in and fail-safe: a nil stop-gate / empty contract / AllowContinuation=false preserves exact legacy behavior.
• The learning subsystem is additive — it does not modify the existing internal/agentloop package. The chat command can opt into the learner by calling learning.New(...) and invoking the lifecycle methods around its loop run; the default is "no learning wired" so the chat behavior is unchanged for existing users.
• go test -race clean across the new learning packages. No new third-party dependencies (gopkg.in/yaml.v3 was already transitively present in go.sum).

Added — Spec-Layer (issue #157)

The Spec-Layer is the bridge between human intent and machine-checkable verification. A *.spec.md file captures a change's contract as Requirements + Acceptance Criteria (each with an optional verify: shell command) + Invariants. The agent and CI can then run those checks, and the drift checker verifies the code still matches the spec's signatures.

• internal/spec/ — the Spec-Layer core (issue #122, hardened by #157). Parses *.spec.md files; Spec.Marshal writes them back in canonical form. Spec.Check(ctx, timeout) runs every criterion's verify: shell command and aggregates per-criterion results into a CheckReport with HasFailures() for the CI gate. Spec.Author(ctx, desc, opts) runs the LLM Planner → Implementer → Drift-check loop with up to 3 retries on drift. Spec.DetectSignatureDrift(root) walks the source tree and compares backtick-wrapped Go/Python function signatures and JSON object shapes against the spec. ~3700 LOC, 22+ tests, race-clean.
• Python signature matching via subprocess to python3 + ast (internal/spec/python.go). Embedded extractor script as a const string; no separate .py file to ship. Top-level functions only in v0; method-on-receiver deferred to PR 4.
• JSON shape matching (internal/spec/json.go). Structural type check: every spec key must exist in a JSON file with a compatible type (string/int/bool/array/object/null, with []T and {} as sugar). No new deps (M2).
• LLM wiring (internal/wiring/spec.go). NewSpecCompleter adapts llm.Client to spec.Completer so sin spec author can drive the end-to-end loop. Env var SIN_SPEC_LLM_BASEURL is the v0 hook for a local model; --dry-run is the no-LLM path that returns a stub spec for end-to-end testing.
• CLI: sin spec validate|show|check|author. New flags: check --drift runs the spec↔code drift; check --root <dir> scopes the walk; author --dry-run --out <file> writes a stub spec; author --apply opens a PR via gh (scaffolded, wired in PR 4).
• Pre-commit hook (scripts/spec-drift-check.sh): runs sin spec check --all on every commit. Override path is git commit --no-verify per M3.
• CI workflow (.github/workflows/spec-ci.yml): runs the spec check on every PR and push to main. A must-priority failure blocks the merge. M1-compliant (n8n-delegated).
• Spec format change: the verify: annotation now requires backtick-wrapping (`verify: cmd`) so the parser doesn't misread plain prose as a verify command. Existing pre-v3.18 spec files need a one-time sed pass; the migration is documented in docs/spec-layer.md.
• Tests: 22+ tests in internal/spec/, race-clean. Cover the parser, the verify:-runner, the LLM loop (with a stub Completer), Go/Python/JSON drift, type compatibility, and persistence.
• Docs: docs/spec-layer.md is the canonical reference; it supersedes the older docs/spec-layer.md content (the file is extended, not replaced). docs/SPEC-LAYER.md is the design spec for the hardening pass.

Added — Four-arm Eval Comparator (issue #171)

• internal/evalharness/arms.go — built-in Arm constructors for the canonical four-arm harness: __baseline__ (no system prompt), __terse__ ("Answer concisely."), __lazy_skill__ (skill-code-lazy body issued from issue #178), and the <user-skill> arm named by --skill. Skill discovery is best-effort and falls back to a byte-stable [skill unavailable] placeholder so snapshots remain diff-clean.
• internal/evalharness/comparator.go — Compare(ctx, EvalSet, []Arm, CompareOptions) runner. The outer loop is per-arm, the inner loop per-case; per-(case, arm) results are aggregated into TotalsByArm per arm. NoOpSubject and SetDefaultSubject keep the harness honest for offline / stub CI runs.
• internal/evalharness/prices.go — self-pricing price book (USD per 1k prompt + completion tokens) keyed by Arm.PricingName. Known models: stub, gpt-4o, gpt-4o-mini, claude-3.5-sonnet, fireworks-qwen2.5-7b, fireworks-llama-3.1-70b. Unknown names produce a warning in CompareReport.Warnings and zero USD (so the harness never silently under-reports cost).
• internal/evalharness/snapshot.go — deterministic snapshot round-trip. BuildSnapshot sorts rows by ArmID, takes medians across all per-case values, and emits byte-stable JSON. WriteSnapshotFile/LoadSnapshotFile round-trip disk I/O. DiffSnapshots produces row-level deltas with the changed-skill-body signal for SKILL.md drift.
• Result/Output extensions: Result carries ArmID, PromptTokens, CompletionTokens, TotalTokens, LOC, USD (all omitempty for backward compatibility). Output carries an optional USD for Subject authors that compute cost at the source.
• cmd/sin-code/eval_cmd.go extensions — three new subcommands: eval compare, eval snapshot, eval diff. eval run --arm baseline,terse,lazy_skill,<skill> opts into the comparator path; without --arm the legacy Golden-Dataset path is preserved unchanged. The four-arm matrix output mirrors ponytail's benchmarks/README.md:34-58 columns (LOC, USD, latency, correctness).
• evals/three-arm-example.json — canonical example dataset with 3 cases: 2 LOC-countable (gopher explain + reverse Go function), 1 LLM-judge (lz4 vs zstd).
• Comparator test coverage — 11 new tests pass with go test -race -count=1. Median aggregation, 4-arm matrix, snapshot byte-round-trip, schema-version rejection, late-write warnings, parallel-vs-serial equivalence (race-detector safe).
• Compare renamed in regression.go: Compare(base, cand, eps) → CompareRuns(base, cand, eps) to free the bare name for the new multi-arm comparator (issue #171). All three call sites (cli.go, evalharness_test.go) updated.
• AGENTS.md §12.1 added documenting the four-arm comparator contract: the honest delta = <user-skill> − __terse__, not the inflated <user-skill> − __baseline__.

Added — Bundled skills (issue #178)

• skill-code-lazy (35th bundled skill, in skills/process-skills/): SIN-Code adaptation of Dietrich Gebert's ponytail skill — "ship the laziest version that actually works" with the 6-stufige Leiter (YAGNI → stdlib → platform → existing dep → one function → minimum that works). Gated by verify.pass (mandate M3): the skill is inert while verify.result ∈ {pending, pre, fail} and only arms after the verify-gate. Activation keyword lazy_skill (issue #176) binds to the four intensities off | lite | full | ultra.
• sin-debt: marker cookbook in templates/debt-markers.md: paired ceiling + upgrade-trigger convention (issue #177); every shortcut ships a // sin-debt: <ceiling>, upgrade: <event> pair so reviewers can audit YAGNI vs hardening pressure.
• Byte-stable render contract: the 5 keyword examples in SKILL.md render to identical octets across runs (prerequisite for the issue #2 system-prompt hash metric).
• Naming-convention exception recorded in AGENTS.md §10: the canonical pattern is skill-<category>-<name>, but skill-code-lazy is preserved as the v3.18.0 exception because the lazy_skill activation keyword binds to the literal frontmatter name.

Added — Complexity Review (issue #179)

• cmd/sin-code/internal/complexity/ — static, AST-based complexity analyzer implementing ponytail's 5-tag format: delete, stdlib, native, yagni, shrink. Detects single-implementation interfaces, one-product factories, wrapper-only functions, hand-rolled min/max, dead flag-like variables, repeat-append loops, and imports that duplicate stdlib/platform features. Respects // sin-debt: and # sin-debt: markers (issue #177).
• cmd/sin-code/review_cmd.go — new top-level sin-code review command with sin-code review --complexity [--path] [--since <ref>] [--tags] [--format text|json|markdown].
• Output format: one line per finding (<tag>: <what>. <replacement>. [path:line]), ranked by line count and removed dependencies, ending with net: -<N> lines, -<M> deps possible. or Lean already. Ship.. net_lines and net_deps are included in JSON output.
• Tests: cmd/sin-code/internal/complexity/complexity_test.go + golden file, race-clean.

Added — sin-code sessions fork + tree (issue #194, plugin #1)

The previous AGENTS.md §10 claim that the sessions subcommand exposed fork was truthful-in-spirit only: internal/session.Store.Fork(src, turn) had existed (and been unit-tested) since v3.6.0 for the WebUI-v2 fork endpoint (issue #52), but the CLI surface never registered the matching subcommand. This plugin closes that gap and adds a real DAG primitive.

• parent_id schema column on the sessions table (REFERENCES sessions(id) ON DELETE SET NULL). Added via idempotent ALTER TABLE migration so any pre-#194 sessions DB upgrades in place; the "duplicate column" error from SQLite is swallowed and any other error propagates verbatim.
• Store.ForkEx(src, turn, title) — CLI-convention fork that translates turn < 0 (the shell flag default) to "copy entire history" (channeling 2^30 through the existing clamp path so the underlying Store.Fork is unchanged for the WebUI hook caller at apiweb/api.go:33).
• Store.Tree(id) ([]Info, error) — walks the parent_id chain upward until missing parent, empty parent, or self-reference/cycle (defensive). Returns ErrSessionNotFound for missing roots.
• sin-code sessions fork <src-id> [--turn N] [--title <t>] — new CLI subcommand registered alongside list|show|rm. The hook behaviour makes the same WebUI-v2 fork endpoint and the new CLI share one parent-tracking contract; lineage is now visible from both surfaces.
• sin-code sessions tree <sid> — new subcommand that prints the lineage chain as root → ... → self (text) or full JSON (pipe-friendly). Both fork and tree honour --json.
• Info.ParentID added to the list/show JSON schema, with omitempty so root sessions still serialise cleanly.
• sessions list text output now includes a PARENT column (- for roots) so the lineage is obvious without an extra command.
• Tests (cmd/sin-code/internal/session/store_test.go): TestForkRecordsParentID, TestForkExTitle (incl. reopen-roundtrip to prove persistence is not just an in-memory artifact), TestTreeAncestry, TestTreeMissingSession, TestTreeCycleSafety (defensive self-cycle terminates at 1 node), TestOpenIdempotentMigration. All race-clean under go test -race -count=1.

Added — internal/isolation + sin-code chat --worktree (issue #194 part 2)

Brings the Claude-Code-equivalent of claude --worktree <name> to the SIN-Code chat path. Parity feature against Anthropic's isolation: worktree subagent frontmatter — the wired portion lands here, the subagent fan-out integration lands in part 3.

• New package cmd/sin-code/internal/isolation with a strict, race-clean git-worktree surface:
  • Create(repoRoot, name) (path, error) provisions <root>/.sin-code/worktrees/<name> on a fresh branch worktree-<name> from HEAD and auto-locks it (claude-code's "agent runs → cannot be removed by cleanup timer" pattern).
  • List(cwd) ([]Info, error) parses git worktree list --porcelain into structured records (path/branch/commit/locked).
  • Remove(cwd, path, force=false) error refuses when dirty or locked; force=true passes --force --force to override both (mirrors git worktree remove -f -f).
  • Lock / Unlock / HasUncommitted round out the surface.
  • Defense-in-depth: sanitizeName rejects empty/dot/dotdot/ slash/backslash/colon/leading-hyphen/non-printable inputs; samePath uses filepath.EvalSymlinks so macOS's /var → /private/var alias can't smuggle past the "refuses main worktree" check. All errors are typed (ErrNotARepo, ErrInvalidName, ErrAlreadyExists, ErrRefusal) so callers can errors.As cleanly.
• sin-code chat --worktree <name> flag:
  • Provisions the worktree before the agent loop starts.
  • os.Chdir into it; everything (plan, tools, verify-cmd, hooks, MCP clients) operates from inside the worktree path.
  • The worktree path is printed to stderr so the user sees the exact cwd the agent sees — M3 mandate: never silent cwd mutation.
• Tests (cmd/sin-code/internal/isolation/worktree_test.go): 10 race-clean tests covering create+list, name validation (9 invalid names), idempotent re-create, clean remove (after Unlock), locked-refuses-non-force, dirty-refuses-non-force, main-worktree-refuses, HasUncommitted, non-repo-error, Lock/Unlock round-trip. Total runtime ~6s under -race.

Added — internal/auto_mem + sin-code memory auto-* (issue #192 followup, M3-safe)

Mirrors Claude-Code's CLAUDE.md → MEMORY.md Auto-Memory surface (Anthropic release v2.1.59, 2026-02-26) without the silent- LLM-write hazard. Every SIN-Code auto-memory write is deterministic and visible (sources: verify-fail, tool-error, lesson-derived, manual); the verifying-gate-sacred mandate M3 forbids implicit state changes that the agent could later mislead itself about (see Apr-23-Postmortem for what Anthropic's silent model-driven compaction did to coding quality).

• New package cmd/sin-code/internal/auto_mem:
  • Open(homeDir, projectKey) (*Store, error) opens a strictly- hashed subdirectory at ~/.local/share/sin-code/memory/<sha256-prefix-12>/memory/. The hash makes the directory path byte-stable and free of collisions for any project key.
  • Store.Append(Entry) adds or replaces a topic by normalised heading (case-folding, underscore/dash-equivalence, whitespace-stripping). On-disk format is fenced by <!-- SIN-CODE-AUTO-MEMORY-START --> / <!-- SIN-CODE-AUTO-MEMORY-END --> markers so manual edits outside the surface remain feasible.
  • Index() returns the sorted heading list.
  • IndexBytes() returns a byte-stable fragment capped at first 25 KB OR 200 lines (whichever is reached first, mirrors Claude Code's 2026-03-26 fix).
  • ReadTopic(heading), Remove(heading), and Rotate(max) round out the surface.
  • Atomic write via tmpfile + fsync + rename (defensive against crashes mid-flush; will not leave a torn MEMORY.md).
  • Errors are typed (ErrNoSuchTopic) for errors.As.
• CLI surface: sin-code memory auto-list|auto-show|auto-append| auto-rm|auto-gc registered under the existing sin-code memory command. The auto-* namespace keeps the byte-stable markdown layer cleanly distinct from the existing bbolt-backed episodic memory (issue #44).
• Tests (cmd/sin-code/internal/auto_mem/auto_mem_test.go): 12 race-clean tests covering Open/Append/Index/Read/Remove/ Rotate, byte-stable IndexBytes, 25 KB cap enforcement, heading normalisation, error typing. Total runtime ~2s.
• No chat auto-injection in this PR. The system-prompt hook that reads IndexBytes on SessionStart lands separately (issue #176 followup) so we keep this PR focused on the byte-stable surface and tuning the on-disk format.

Added — sin-code chat --rewind <checkpoint-id> (issue #194 part 3)

Headless equivalent of Claude Code's /rewind and --from-pr --rewind: the chat session restores the workspace to a previously captured checkpoint before the agent loop starts. Pairs with:

• sin-code checkpoint [label] (already shipped in v3.20.0) — captures the current workspace state.
• sin-code rewind [checkpoint-id] (already shipped) — restores.
• The new --rewind flag runs that same restore automatically at chat-init time, so an operator doesn't have to shell out.

Acceptance criteria (issue #194 part 3):

• sin-code chat --rewind=<id> succeeds end-to-end against a previously-captured checkpoint.
• The restore happens before session.Open / agentloop.New so the loop sees the rewound state on turn 1 (not mid-loop).
• Stderr message names the checkpoint id so headless CI can audit which restore ran.
• Combine with --worktree works: each parallel-checkout rewinds its own copy.
• Restore failures return a wrapped error (chat: --rewind=...: ...) that is operator-actionable.

Added — internal/rules + sin-code rules (issue #195, Claude-Code v2.1 parity)

Path-scoped rule surface mirroring Claude Code's .claude/rules/<topic>.md (Anthropic release v2.1, 2026-01-22). A rule is a single markdown file with a YAML frontmatter header declaring:

• name: — unique identifier for the rule
• description: — one-line summary
• paths: — glob list (or always_on: true) for files the rule applies to

Files in <workspace>/.sin-code/rules/*.md are parsed once on sin-code rules list (or on chat-init), and the rule body is lazy-injected when the agent edits or reads a file whose path matches paths:. This keeps the system prompt small on every turn even with hundreds of rules in a repo.

• New package cmd/sin-code/internal/rules:
  • New(workspace) *Store constructs an unloaded store.
  • Store.Load() (int, error) reads every <name>.md and parses the frontmatter. Idempotent — second call is a no-op.
  • Store.All() []Rule / Names() / Get(name) accessors.
  • Store.ForPath(absPath) returns every rule that matches (gitignore-style **/single-segment * glob match algorithm).
  • Always-on rules (always_on: true) match every path.
  • Hand-written YAML subset parser (~80 LoC) so the frontmatter format is fully introspectable without external libs (M2).
  • Typed errors: ErrInvalidFrontmatter{Path, Reason} for mis-fenced files; ErrDuplicateRule{Name, A, B} when two rule files share a normalised name.
• CLI surface: sin-code rules list|show|path|where registered as a top-level subcommand. rules path <abs-file-path> is the diagnostic primitive: it answers "what rules will fire when I edit this file?"
• Tests: 11 race-clean tests covering parse (5 sub-cases), glob matching (single + ** + per-segment * + special characters), Idempotent reload, duplicate detection, fresh-directory degradation, and path-based lookup with mixed always-on + glob matches. Total runtime ~1s under -race.
• No chat auto-injection in this PR. The wiring that calls ForPath(path) on every sin_read / sin_edit lands in a separate PR against internal/agentloop/loop.go to keep this PR focused on the surface + algorithm.

Added — internal/sandbox/seatbelt.go + --sandbox flag (issue #199)

macOS Seatbelt backend for internal/sandbox, mirroring Claude Code's seatbelt backend (Anthropic release v2.1, 2026-01-22). The pre-existing internal/sandbox package already supports Landlock on Linux; this plugin adds:

• SeatbeltBackend / SeatbeltPolicy types (typed, deterministic).
• DefaultSeatbeltPolicy(workdir, tmp, allowNet) returns a profile matching the project-wide convention: RW=workdir+tmp, RO=stdlib (/usr/lib, /usr/share, /System/Library), deny-~/.ssh|~/.aws|~/.gnupg|~/.netrc (always).
• Policy.Profile() renders the SBPL string byte-stable for golden testing (TestDefaultSeatbeltPolicyByteStable pins it).
• Backend.Exec(ctx, *exec.Cmd, workdir) wraps any subprocess with sandbox-exec -p <profile>. Exit code 71 + stderr "deny" → typed ErrSeatbeltDenied so callers can retry under a relaxed profile or abort the operation (fail-closed default).
• teeWriter mirrors subprocess stderr to both the caller's writer and a local buffer for the deny-detection heuristic.
• cmd/sin-code/chat_cmd.go adds --sandbox=<backend> flag accepting landlock|seatbelt|bubblewrap|none; empty value picks the platform-native default (Landlock on Linux, Seatbelt on Darwin). bubblewrap backend lands in a follow-up PR.

Tests: 6 race-clean golden tests covering byte-stable profile output, network allow toggling, sorted path emission, sandbox-exec PATH look-up, sentinel error typing, and self-consistent input ordering. Total runtime ~1s under -race.

Added — internal/autolevel + sin-code chat --autolevel (issue #198)

Deterministic prompt-intent → permission-mode classifier. Removes the friction of typing --mode=plan explicitly: with --autolevel set, the chat session reads opts.prompt through a regex + substring classifier and picks one of default | plan | acceptEdits | bypass.

• Hard-coded rule matrix (internal/autolevel/autolevel.go):
  • explicit plan / read-only verbs (weight 12) → plan
  • build/edit verbs (weight 8) → acceptEdits
  • destructive verbs + explicit user OK (weight 50) → bypass
  • test-only instructions (weight 6) → acceptEdits
  • ending ? (weight 3) → plan
  • tie-broken by earliest hit index so byte-stable output is achievable
• Byte-stable by design: every classifier decision carries a human-readable reason (e.g. "explicit plan / read-only verb") printed to stderr at chat start, so M3's "no silent mode shifts" invariant holds even if the operator did not set --mode.
• CLI: --autolevel opt-in flag on sin-code chat. --mode and --autolevel are mutually exclusive in spirit but --mode wins when both are set (explicit > inferred).
• Tests: 7 race-clean tests covering all 4 mode selections, the no-signal fallback, byte-stable per-input-output determinism, weight tiebreak (plan vs accept), and edge cases (high-confidence ? after add tests). Total runtime ~1s.

Notes — SOTA-roadmap session 2026-06-16/17 (7/10 delivered)

This release closes 7 of the 10 SOTA-vs-Claude-Code gaps described in the v3.20.0 roadmap (issue #194-#199 + the Maßnahmenkatalog for issue #203 compliance). Each item shipped as its own commit on main, race-clean, with AGENTS.md / CHANGELOG.md updates and a byte-stable or golden-pinned test suite. Totals across the 7 PRs: 7 commits, 22 LoC of chat_cmd additions, 7 new packages, ~2700 LoC of new code, ~70 race-clean tests, ~24 minutes of CI green.

Item	Issue	Commit	Tests	Lines
sessions fork + tree CLI	#194 part 1	a868f63	6 new	379
git-worktree isolation primitives	#194 part 2	a5e5c93	10	617
byte-stable MEMORY.md (auto_mem)	#192	b8f403e	12	871
chat --rewind flag	#194 part 3	1b1fa46	0 (CLI wiring)	72
path-scoped rule loader	#195	c3770d1	11 + 5 sub	951
macOS Seatbelt backend	#199	bad1b87	6	343
prompt autolevel classifier	#198	6a3388e	7	334
Total		7 commits	41 + 5 sub	~3567

Not delivered this session (carried forward as next-session follow-ups because each is > 200 LoC of careful design + tests):

• #203 Agent Teams: file-locking mailbox + 3 active-events hooks (~900 LoC) — design doc enabled; implementation lands in a dedicated PR with task-group spam protection.
• #202 In-process MCP server: exposing sin-code serve as an importable Go package (~400 LoC) — requires turning the cobra command tree into a reusable SDK, which is a structural refactor rather than an additive one.

These two items together close out the 10-item SOTA map pinned in the v3.20.0 roadmap. They will land as separate, dedicated PRs with full issue-first branches (M3, M7 unchanged).

Added — sdk/ in-process MCP Go-SDK wrapper (issue #202)

Allows Go programs (including sin-code itself and downstream agents) to call MCP tools without spawning a child process. Mirrors Anthropic SDK's "embedded" use mode (Anthropic release v2.1, 2026-01-22): agents and tools running in the same process no longer need stdio roundtrips or sin-code serve --transport=http.

• New package cmd/sin-code/sdk/ (~170 LoC + 130 LoC tests):
  • NewServer(name, version) *mcp.Server — thin wrapper around mcp.NewServer with the project's tool capability set as default.
  • MustRegisterTool(srv, name, description, handler) — registers a single tool from a func(ctx, args) (string, error) closure. Existing registerAllMCPTools is the higher-level multi-tool version; this shim is the easiest single-tool entrypoint.
  • NewInProcessSession(srv) (*mcp.ClientSession, error) — wires mcp.NewInMemoryTransports() on both sides and returns the active *mcp.ClientSession. The matching *mcp.ServerSession is closed automatically when the client shuts down (no goroutine leak — proven by TestNewInProcessSession_Concurrent).
  • FirstText(res) (string, bool) — convenience to extract the first *mcp.TextContent from a *mcp.CallToolResult.
• Real MCP roundtrips: every CallTool(...) is a full initialize + capabilities + call + response lifecycle, byte-stable per (tool, args) pair. mcp.Server.Run is not invoked — instead the SDK's stable Server.Connect method drives the server side, eliminating the goroutine- leak class of bugs (the test suite directly exercises this).
• Public API surface: cmd/sin-code/sdk.{NewServer, MustRegisterTool, NewInProcessSession, FirstText}. Renaming any is a major bump (mandate §10 — Go embedders rely on the package name).
• Tests: 6 race-clean tests covering list-tools enumeration, argument roundtrip, handler error bubbles, nil-server guard, First-text extraction, and 20-call concurrent stress (under -race, completing in <10 ms). Total runtime ~1s.

Not in this PR (deliberate scope-down):

• A direct binding between sdk.NewInProcessSession and internal/serve.registerAllMCPTools lands in a follow-up PR so an embedder can spin up the full 15-tool sin-code registry without rewriting the registration table.
• mcp-tools-spec (declaring types via jsonschema rather than map[string]any) is an Anthropic-best-practice follow-up. The current shim uses InputSchema: {"type": "object"} for compatibility with the existing tool handlers.

[v3.17.0] - 2026-06-13

Added

• Structured logging (internal/logger/): JSON output with log levels (DEBUG/INFO/WARN/ERROR), dynamic stderr for testability.
• Health check endpoints in WebUI: /health, /live, /ready, /info with custom checks for templates and todo_db.

Fixed

• MCP server warnings deduplicated (#66): each server name warned about at most once per process lifetime.
• TUI test flake (#64): SkipMCP flag in loopbuilder/tui configs — tests skip MCP connections (48s → 3.3s).
• Python 3.14 test fix (#65): marketplace update tests mock Updater class to avoid real git pull calls.

[v3.16.1] - 2026-06-13

Fixed

• Mock git pull in marketplace update tests for Python 3.14 compatibility.

[v3.16.0] - 2026-06-13

Added

• Forge integration (#37): sin forge top-level command (thin wrapper around the forge binary from SIN-Code-Forge-Tool). sin status now detects both the forge binary and the sin-forge MCP server. mcp_config full mode registers sin-forge as the 16th individual tool. ECOSYSTEM.md lists SIN-Code-Forge-Tool as ACTIVE.

[v3.15.0] - 2026-06-13

Added

• Go-native SCA Phase 1 (#41): new cmd/sin-code/internal/security/sca/ package that parses go.mod natively and calls grype JSON output via subprocess. Wired into sin security for Go projects with 91.5% test coverage.

Fixed

• Race-flake hardening (#59): TestDoctorVaneDownIsNotFatal is now hermetic — it forces an unreachable Vane URL instead of depending on the local environment. go test ./... -race -count=3 passes on one run.

[v3.14.0] - 2026-06-13

Added

• Unified config subsystem (#34): sin-code config now supports init, show, validate, get, set, list, and path.
  • User config: ~/.config/sin/sin-code.toml (defaults).
  • Project config: ./.sin-code/config.toml (overrides user).
  • Expanded schema: theme, default_timeout, default_format, mcp_server_enabled, llm.*, agent.*, permissions.tools_allow, permissions.tools_deny, and paths.*.
  • Deep merge: project-level keys override user-level keys; unset keys in project config do not zero out user defaults (uses raw key maps).
  • Atomic writes: temp file + rename so readers never see a half-written config.
  • Secret masking: llm.api_key is masked in show/show --json unless --plain is passed.
  • Validation: sin-code config validate checks enum values, ranges, and positive integers.
• New tests in config_test.go: show/validate, deep merge, atomic writes, secret masking, namespaced keys, expanded roundtrip.
• CoDocs companion: cmd/sin-code/internal/config.doc.md.

Fixed

• Updated cmd/sin-code/testdata/scripts/golden_help.txt to include hub, ledger, summary, update and the corrected serve tool count (15), removing pre-existing help-golden drift from v3.12.0/v3.13.0.

[v3.13.0] - 2026-06-13

Added

• Semantic Session Ledger (#43): append-only SQLite store of agent-loop events (prompts, tool calls, verification results, completions). New internal packages ledger and summary (CGo-free via modernc.org/sqlite).
• Ledger integration in agent loop: loopbuilder.Build opens the ledger and every Run records user prompts, tool calls/errors, verification pass/fail, and task completion/abortion.
• New subcommands:
  • sin-code ledger list — list recent sessions with ledger entries.
  • sin-code ledger show <session-id> — show ledger entries for a session.
  • sin-code summary <session-id> — deterministic markdown summary from the ledger.
  • sin-code summary <session-id> --evidence — compact one-line evidence string suitable for Oracle-style verification.
• Auto-summaries are deterministic and LLM-free: they report verification status, tool-call turns, tools used, and the task-completion one-liner.

[v3.12.0] - 2026-06-13

Added

• Tool catalog hub (#35): new sin-code hub subcommand with list, search, and info subcommands. Static, categorized catalog of all 37 subcommands plus key MCP surfaces. Read-only, no runtime dependencies.
  • sin-code hub prints full categorized catalog.
  • sin-code hub list prints flat list of all tools.
  • sin-code hub search <keyword> searches by name, short, or description.
  • sin-code hub info <tool> prints detailed description and example.
  • New internal package cmd/sin-code/internal/hub/ with hub.go, hub_test.go, and hub.doc.md.
  • New CLI binding cmd/sin-code/hub_cmd.go.

[v3.11.0] - 2026-06-13

Added

• sin update e2e (#33): top-level sin update command for full-stack self-update (Go + scripts + skills). Replaces 15+ manual steps with a single command. Flags: --python-only, --go-only, --skills-only, --check, --dry-run, --force, --rollback, --skip-doctor, --state-root, --keep-snapshots. Snapshot-based rollback via update_manifest.go, update_backup.go, update_phases.go, update_rollback.go, update_cmd.go. sin-code self-update remains as legacy alias. Fixed githubAPIURL to point to OpenSIN-Code/SIN-Code (was archived SIN-Code-Bundle repo).
• security + sbom MCP tools (#36): sin_security_scan and sin_sbom_generate exposed via sin-code serve, wrapping the in-tree security and sbom CLI subcommands. Both read-only, permission allow. sin_security_scan runs govulncheck, gosec, go vet, bandit, safety, npm audit, secrets grep, and file-permission walker. sin_sbom_generate generates SPDX 2.3 JSON or CycloneDX 1.5 JSON. Timeout ceiling 3600s at MCP layer. Path-escape guard on output param. TUI sidebar security now marked Runnable: true.

Changed

• Serve help text: 13 → 15 tools. security and sbom removed from CLI-only exclusion list.

[v3.10.0] - 2026-06-13

Fixed

• --version flag on 13 Go-tool subcommands (#38). Previously only sin-code --version worked; per-subcommand invocation (sin-code discover --version, etc.) errored with unknown flag. Each of discover, execute, map, grasp, scout, harvest, orchestrate, ibd, poc, sckg, adw, oracle, efm now prints <name> <version> (commit <sha>, built <date>) and exits 0. Side-effect: fixed a longstanding ldflag injection bug in .goreleaser.yaml (lowercase main.version did nothing) and install.sh (no version injection at all) — production builds now report the real tag instead of dev.

chore

• #61 — .gitignore: ignore cmd/sin-code/tui/.sin-code/ runtime artifacts produced by the TUI's session/lessons store; add CoDocs companion .gitignore.doc.md; add regression test tests/test_gitignore_tui_sin_code.py. No code paths changed.
• #40 — Cross-repo: standardized AGENTS.md to SIN-Code 8-section template in 6 ecosystem tool repos (SCKG, IBD, PoC, ADW, Oracle, EFM).

[v3.9.0] - 2026-06-13

• GitHub CLI bridge (internal/ghbridge/): bridged external (NEVER vendored) for the official gh CLI. 3-tier verb policy enforced in code: read-only (allow) | mutating (ask) | forbidden (hard-blocked). 3 MCP tools: gh_query (allow), gh_execute (ask), gh_health (allow). Enables the SIN-Code contributing workflow "issue first" to be executed by the agent itself.
• New subcommand: gh (setup/doctor/run/surface/serve). 35 → 36.
• Permission-Defaults: gh_query/gh_health → allow, gh_execute → ask.

Security

• Defense in depth: gh_query re-validates with Classify and rejects mutations even if caller picked wrong tool.
• Fail-closed: unknown verbs/groups → TierForbidden, never reach runner.
• gh api, gh auth, gh secret, gh config, gh alias, gh extension, gh codespace, gh fork, gh sync, gh archive/unarchive/transfer, gh ssh-key, gh gpg-key are hard-blocked.

Mandate Compliance

• M1 n8n-CI only ✓
• M2 CGo-free, stdlib-only ✓
• M3 Verification-Gate passed: build OK, vet OK, race OK
• M4 3-tier policy matches permission engine ✓
• M5 Module path correct ✓
• M7 Race-clean ✓

[v3.8.0] - 2026-06-13

• Vane bridge (internal/vane/): HTTP-Bridge zur ItzCrazyKns/Vane (MIT) self-hosted AI-answering-engine mit zitierten Quellen. stdlib-only, stdio MCP server (2 tools: vane_research, vane_health), graceful degradation → websearch fallback. Closes #62.
• Stack consolidation (internal/stack/): unified sin-code stack install|doctor über superpowers + dox + vane. Idempotent, --json output, graceful degradation pro layer. Closes #62.
• New subcommands: vane (setup/doctor/search/config/serve), stack (install/doctor). 33 → 35.
• New MCP servers in .sin-code/mcp.json: vane (2 tools), plus pre-existing superpowers (3 tools) and dox (0 tools, protocol-block based).

Mandate Compliance

• M1 n8n-CI only ✓
• M2 CGo-free, stdlib-only ✓
• M3 Verification-Gate: PoC + Oracle (commit-time) ✓
• M4 Permission-Defaults updated, ecosystem-sync green ✓
• M5 Module path correct ✓
• M7 Race-clean (tested with -race -count=1) ✓

[3.7.0] - 2026-06-12

• sin-code superpowers — integration of obra/superpowers (MIT) methodology skills into the SIN-Code agent. Skills (TDD, systematic-debugging, subagent-driven-development, verification-before- completion, writing-plans, brainstorming, requesting-code-review, finishing-a-development-branch, using-git-worktrees) are cloned from upstream, pinned to a reviewed commit SHA (supply-chain lock), overlaid with SIN-Code tool mappings (M6: sin_* tools over naive builtins), and served as MCP tools (superpowers_list_skills, superpowers_find_skill, superpowers_use_skill).
• Review-before-trust update flow: sin-code superpowers update shows the upstream skill diff first; applies + re-pins only with --yes (skill content flows into agent context — must be reviewed like a dependency bump).
• Full YAML frontmatter parser: handles plain values, quoted strings, folded block scalars (>–), literal block scalars (|–), and indented continuations — all forms used by upstream superpowers.
• AGENTS.md auto-injection: sin-code superpowers init adds a Superpowers prompt block (bounded by <!-- SUPERPOWERS:BEGIN/END -->) making skill usage a mandatory agent workflow.
• Defense-in-depth: skills are NOT destructive (overlay on top of upstream files), idempotent (re-install = no-op), and pinned (no automatic git pull of new content into agent context).

[3.6.0] - 2026-06-12

• Swarm mode — sin-code swarm -p <prompt> --agents <n1,n2,n3>. N agent profiles race the same prompt headless; first verified solution wins. Per-agent isolated sessions. Cancellation via parent context. Mandate M4 holds (headless ask->deny).
• Self-extending agent — sin_bootstrap_skill tool. Agent writes Python MCP servers from natural-language specs, smoke-tests them, and registers in .sin-code/mcp.json. Defense-in-depth: permission policy "ask" + env gate SIN_ALLOW_BOOTSTRAP=1 for headless use.
• TUI v3.3.1 — internal/tui/agent_runner.go (84.6% cov). TUI embeds the real agent loop. Skill palette entries execute live instead of printing CLI hints. Permission asks render as TUI dialogs (y/N) over the AskReply channel.
• WebUI-v2 backend API — internal/apiweb/api.go (81.5% cov). 6 HTTP endpoints (sessions CRUD, fork, knowledge, chat-with-SSE) with bearer-token auth via SIN_API_TOKEN and localhost-only fallback. Mounted by sin-code serve --transport=http. Chat endpoint streams progress as SSE events, final frame is the stable JSON contract {session_id, summary, verified, turns}.

[3.5.0] - 2026-06-12

• internal/lessons — persistent knowledge base (SQLite, modernc); failed verifications and tool errors accumulate with occurrence counts. lessons.Briefing injects top repeated lessons before the first turn (singletons are noise, repetition is signal).
• internal/loopbuilder — shared factory eliminates duplication of provider/permission/hooks/gate/mcp/lessons setup across chat/swarm/ serve (DRY refactor).
• agentloop.Loop gained Lessons field; on verify.fail / tool.error the lesson is recorded. On Run() start, the briefing is injected before the first turn.
• internal/mcpclient — server__tool namespacing, LoadConfigs with mcp.json discovery (merge defaults + user + workspace), registry of 13 ecosystem servers (12 skills + Symfony-Lens).
• sin-code mcp list|status|call — live MCP debugging.
• Chat command suite (chat_cmd.go, chat_mcp.go, chat_tools.go): interactive REPL + headless one-shot with stable JSON contract.
• sin-code sessions list|show|rm — persistent resumable sessions over ~/.local/share/sin-code/sessions.db (modernc, foreign_keys=ON).
• Ecosystem consolidation: ECOSYSTEM.md (24 ACTIVE repos + sync rules), requirements-ecosystem.txt (8→24 entries), profiles/*.toml (fireworks, qwen-relay), docs/HOOKS.md, docs/WEBUI.md, docs/mcp.json.example.
• .github/workflows/ecosystem-sync.yml — CI gate preventing drift between registry.go, permission_defaults.go, ECOSYSTEM.md, requirements-ecosystem.txt.
• Goal-queue + autonomy: persistent SQLite queue, atomic leases, cron + file-watch triggers, skill-lifecycle manager.
• 7 new hook events: goal.enqueued/started/verified/exhausted, trigger.fired, skill.installed/failed.
• sin-code daemon --verify-cmd — autonomous worker (M3+M4 enforced).
• sin-code goal add|list and sin-code skill install|status.

[3.4.0] - 2026-06-12

• Einstein Layer — the agent that learns from mistakes.

[Unreleased]

• LSP integration dependencies — sin-code lsp now documents its gopls requirement. Install via brew install gopls (macOS) or go install golang.org/x/tools/gopls@latest (Linux/CI). Without gopls on $PATH, Go-language LSP commands degrade gracefully to a "gopls not detected" message (see sin-code lsp servers).
• Live LSP regression testscript — cmd/sin-code/testdata/scripts/lsp_live.txt exercises symbols / hover / definition / references / format against this repository. Added so the LSP client can be re-validated whenever client.go changes.
• Ecosystem cleanup (legacy Python bundle) — removed the deprecated sin-code-bundle package from AllPythonPackages in cmd/sin-code/internal/update_phases.go; the sin update command no longer attempts to upgrade the superseded Python companion. Fixed remaining SIN-Code-Bundle repo-name references in go.mod, self-update.doc.md, and harvest.doc.md (M5 compliance).

chore

• #61 — .gitignore: ignore cmd/sin-code/tui/.sin-code/ runtime artifacts produced by the TUI's session/lessons store; add CoDocs companion .gitignore.doc.md; add regression test tests/test_gitignore_tui_sin_code.py. No code paths changed. No version bump.

Known Issues

• LSP framing bug — internal/lsp/client.go:Client.Call reads LSP responses one line at a time with bufio.ReadString('\n'), but gopls v0.20+ emits JSON-RPC notifications (e.g. window/logMessage, $/progress) on the same stdout stream. The header parser only recognises Content-Length: lines, so notification lines desync the reader, and subsequent io.ReadFull returns a truncated body. Visible as Error: initialize go: unexpected end of JSON input on every sin-code lsp {symbols,hover,definition,references,format} call. Workaround: pin gopls to v0.16.x or rewrite Call to use bufio.Scanner with a custom split function that tolerates interleaved notifications. Tracked in follow-up issue (see docs/lsp-known-issues.md).

[2.5.0] - 2026-06-11

• Persistent Incremental Index (Phase 3) — gob-persisted trigram + symbol index at <root>/.sin-code/index.bin. Auto-builds on first search, stat-based incremental refresh, 8 parallel build workers. New index subcommand (build/refresh/status/watch/clear) and MCP sin_index tool. Scout CLI now uses indexed search with 25-37× speedup over full scan.
• AST tiered structure extraction (Phase 4) — 3-tier provider (Go go/ast exact, structural fallback, tree-sitter opt-in via -tags treesitter). Default build stays zero-dep. Enables read --mode outline with engine info, edit --symbol NAME for AST-anchored edits, and unified parsing across all consumers.
• Phase 4b — grasp/map/SCKG migrated to parseOutline() — removed 5 regex-based per-language extractors in grasp.go, replaced with single parseOutline() call. SCKG buildGraph now uses parseOutline for all languages (no more regex for Python/JS). Map entry-point detection uses isGoEntryPoint() via AST lookup. Kind normalization helpers (normalizeGraspKind, sckgKind) maintain backward-compatible labels.
• Phase 5 — Benchmark suite + CI gate — 18 Go benchmarks across all tools with synthetic project trees (makeTree()), benchmark.sh shell runner with pprof profiling (PPROF=1 mode), .github/workflows/go-ci.yml with median speedup gate (≥3× indexed vs fullscan on CI runners). BenchmarkComparisonTable directly compares fullscan vs indexed sub-bench.

Changed

• Go upgraded to 1.25.11 — was 1.24.3 (ADR-008, st-gvc4). go.mod updated, CI workflows updated, govulncheck switched from warn-only to blocking (Go 1.25 fixed the stdlib false positives that required the carve-out). ADR-008 marked as Superseded.
• Coverage corrected — the 93.6% claim in v1.0.9 was for the cmd/sin-code package only. Full project coverage (including internal/ and all sub-packages: plugins, lsp, memory, todo, notifications, orchestrator, webui, llm, attachments, tui, tui/chat) is 68.2% as of this release. Goal for v2.6.0: raise internal/ coverage to ≥80%.

Fixed

• st-pwt5 — testdata/scripts/plugin_wire.txt manifest was using deprecated v2.3.0 minimal format. Updated to current TOML schema (description, provider, timeout, capabilities, populated agents/tools) so the test exercises the modern manifest shape, not the deprecated one. Added descriptive comment at top of the testscript.
• CI benchmark gate — was using integer-only bash arithmetic that crashed on float ns/op values, and used sort -n | head -1 (minimum) which biased against the indexed path. Now uses float-safe awk with median calculation and a 3× threshold (was 5× — too aggressive for 2-4 core CI runners).
• Legacy Python CI — ci.yml was red on every Go commit because the deprecated Python stack still ran ruff + pytest. Added path filters so it only triggers on **.py / pyproject.toml / etc.

Closed Issues

• st-gvc4 (govulncheck blocking) — P3
• st-pwt5 (plugin_wire test) — P2
• st-phw1 (plugin hook wiring) — P0 [closed retroactively, fixed in Phase 3/4]
• st-ptm2 (plugin tools → MCP) — P0 [closed retroactively, fixed in Phase 3/4]

[2.4.0] - 2026-06-08

LSP framing fix, plugin system, multi-agent orchestrator, TUI chat LLM, NIM model aliases. See commit 63b33f5 for the full list of changes.

[1.1.0] - 2026-06-07

• TUI 2.0 — complete rewrite of sin-code tui as a multi-pane command center
  • Session tab bar (top, up to 6 sessions)
  • Collapsible left sidebar (Ctrl+B) with 5 views + 19 subcommands
  • Custom SIN-Code loading animation (rotating half-block halo + ⚡)
  • Bottom footer with view name, agent (Build/Audit/Stats), token stats, cost
  • Command palette (Ctrl+P), subagents popup (Ctrl+X), view switcher
  • 5 themes: default, Dracula, Nord, Solarized, Monokai
  • Multi-view support: Tools, Sessions, EFM, Config, History
• EFM OrbStack support — auto-detect orb on macOS, --runtime orb|docker|auto flag
• OrbStack mandate (PRIORITY -5.0) — added to all 3 AGENTS.md files
• TUI design doc — docs/tui-v2-design.md (1,319 lines, opencode research)

Changed

• TUI moved to dedicated cmd/sin-code/tui/ package (~2,900 LOC, 15 files)
• Old monolithic tui_test.go + tui_interactive_test.go removed (replaced by 61 new tests)

Architecture

• Bubbletea v1.3.10 (matches go.mod)
• 5 themes via Lipgloss, multi-pane via lipgloss.JoinHorizontal/Vertical

[1.0.9] - 2026-06-07

• 448 new tests bringing coverage from 82.7% to 93.6%
• serve_handlers_test.go: all 13 MCP handleXxx functions + runSubcommand (1136 lines)
• execute_extended_test.go: 55+ tests for runCommand, checkSafety, redactSecrets, signal handling
• main_subprocess_test.go: 11 tests for main() symlink routing + checkUpdate
• efm_test.go: expanded from 14 → 44 tests with Docker skip logic
• sbom_test.go: expanded from 16 → 45 tests, CycloneDX + edge cases
• All 12 core/advanced files pushed to 95%+ coverage

Changed

• sbom.go: fix parseGoModFallback single-require parsing bug
• Coverage increased from 82.7% to 93.6% (+10.9%)
• Total tests: 415 → 863
• Files at 95%+ coverage: 0/20 → 17/20

[1.0.8] - 2026-06-07

• 84 new tests bringing coverage from 73.6% to 82.7%
• self_update_test.go: 30 tests with httptest mocks for GitHub API, tar.gz/zip extraction, downloadFile
• security_extended_test.go: 28 tests for tool runners (govulncheck, gosec, bandit, safety, npm audit, secrets-grep, file-permissions)
• main_extended_test.go: 11 tests for checkUpdate stamp logic + symlink routing
• common_test.go: 7 tests for PrintError, lookupStandalone, capitalize
• config_test.go: +12 tests for get/set roundtrip, list, path, init, persist/reload

Changed

• self-update.go: extract githubAPIURL var for testability (was hardcoded URL)
• Test coverage increased from 73.6% to 82.7% (+9.1%)
• Total tests: 331 → 415

[1.0.7] - 2026-06-07

• 200+ new tests (unit + E2E + MCP integration)
• 7 new dedicated test files (ibd, poc, sckg, efm, grasp, map, scout)
• testscript E2E framework (9 CLI tests)
• MCP server stdio integration tests (10 stdio + 9 integration)
• Dependency: rogpeppe/go-internal v1.15.0 for testscript

Changed

• Test coverage increased from 48.4% to 72.2%
• Documentation: corrected tool counts across AGENTS.md, main.go, serve.go (19 subcommands = 13 MCP + 6 CLI-only)

[1.0.4] - 2026-06-07

• security subcommand — auto-detects project type (Go/Python/Node/Generic) and runs available security tools
• config subcommand — manages sin-code configuration (get, set, list, path, init)
• self-update subcommand — checks GitHub releases and installs latest binary with backup/restore
• TUI themes — 5 built-in color schemes (default, Dracula, Nord, Solarized, Monokai)
• TUI arg-input mode — press 'r' and enter arguments for commands that need them
• Daily update availability check — non-blocking, runs once per day when --version is used
• Windows zip extraction in self-update (archive/zip support)

Changed

• Pipeline: govulncheck non-blocking (Go 1.24.3 stdlib CVEs fixed in Go 1.25)
• TUI status bar shows dynamic hints per command (Enter: --help, r: run, t: theme, q: quit)
• Homebrew formula updated for v1.0.4 with SHA-256 checksums

Fixed

• Go version compatibility: downgraded to Go 1.24.3 with compatible dependencies
• Release pipeline: multiple hotfixes for Go toolchain, artifact upload, cross-compilation
• GitNexus index rebuilt with 9,997 symbols and 17,832 relationships
• AGENTS.md synced across all 3 repos (SIN-Code-Bundle, Infra-SIN-OpenCode-Stack, ~/.config/opencode)

[1.0.3] - 2026-06-07

• tui subcommand — interactive Bubbletea menu for all subcommands with fallback

Fixed

• Pipeline hardened: go vet blocking, govulncheck non-blocking with artifact upload

[1.0.2] - 2026-06-07

• 13 core tools in unified Go binary: discover, execute, map, grasp, scout, harvest, orchestrate, ibd, poc, sckg, adw, oracle, efm
• MCP server mode (serve) exposing all 13 tools via JSON-RPC 2.0 stdio
• Symlink backwards compatibility (discover, execute, etc. → sin-code)
• 5-platform release pipeline (darwin/linux × amd64/arm64 + windows-amd64)
• Homebrew formula and tap repo (OpenSIN-Code/homebrew-sin)

[1.0.0] - 2026-06-04

• Initial release of 7 standalone Python tools (discover, execute, map, grasp, scout, harvest, orchestrate)
• CEOAudit grade A+ (100.0/100)