savant-code 0.0.17 → 0.0.19

package/dev/session-summaries/2026-08-01-sidebar-folded-startup.md

@@ -0,0 +1,159 @@
+# Session Summary: 2026-08-01 Sidebar Folded Startup
+
+**Session ID:** 2026-0801-sidebar-folded-startup
+**Duration:** 10:24 — 10:31 (local time)
+**Status:** completed
+
+---
+
+## Initial State
+
+### Environment
+
+- **OS:** Windows (`win32`)
+- **Language/Runtime:** TypeScript, React, OpenTUI, Bun 1.3.14
+- **Branch:** `main`
+- **Last Commit:** Existing repository history; working-tree changes preserved
+
+### Known Issues
+
+- Right-sidebar sections and active FID cards could render expanded on startup,
+  harming visual density.
+- Existing working-tree modifications were present before this session and were
+  not overwritten.
+
+### Dependencies
+
+- No external service or package dependency required.
+- FreeBuff ECHO Protocol v0.1.2 and Savant ECHO v0.2.0 active.
+
+---
+
+## Planned Work
+
+1. [x] Create and converge `FID-2026-0801-001` through RED/GREEN/AUDIT.
+2. [x] Implement folded startup defaults and focused regression coverage.
+3. [x] Verify, review, close, and archive the FID.
+
+---
+
+## Work Completed
+
+### Task 1: FID Perfection Loop
+
+- **Status:** completed
+- **FIDs Created:** `FID-2026-0801-001`
+- **Changes Made:**
+  - `dev/fids/FID-2026-0801-001-sidebar-folded-on-startup.md`: Created,
+    converged, updated with implementation evidence, closed, and archived.
+- **Verification:** RED identified both expanded defaults; GREEN selected the
+  primitive-boundary fix; independent AUDIT passed.
+
+### Task 2: Sidebar Startup Defaults
+
+- **Status:** completed
+- **FIDs Created:** `FID-2026-0801-001`
+- **Changes Made:**
+  - `cli/src/components/savant-ui/primitives/sidebar-section.tsx`: implicit
+    `defaultExpanded` changed from `true` to `false`.
+  - `cli/src/components/savant-ui/echo/fid-card.tsx`: implicit
+    `initialExpanded` changed from `true` to `false`.
+  - `cli/src/components/savant-ui/echo/__tests__/sidebar-collapse.test.tsx`:
+    added 5 startup-collapse regression tests.
+- **Verification:** All focused tests and static checks passed.
+
+---
+
+## Issues Discovered
+
+### Issue 1: Sidebar and FID cards expanded at first render
+
+- **Severity:** medium
+- **FID:** `FID-2026-0801-001`
+- **Status:** resolved
+
+### Issue 2: Initial test import-order warning
+
+- **Severity:** low
+- **FID:** `FID-2026-0801-001`
+- **Status:** resolved by reordering imports; final ESLint run has zero warnings.
+
+### Issue 3: Interactive visual smoke unavailable
+
+- **Severity:** low
+- **FID:** `FID-2026-0801-001`
+- **Status:** deferred; source/test/static evidence is complete, but no
+  unsupported interactive pass is claimed.
+
+---
+
+## Perfection Loop Summary
+
+| Loop | Target                        | RED                                                   | GREEN                                          | AUDIT                                 | Delta                           |
+| ---- | ----------------------------- | ----------------------------------------------------- | ---------------------------------------------- | ------------------------------------- | ------------------------------- |
+| 1    | Sidebar/FID startup expansion | Found two expanded defaults and missing focused tests | Changed two defaults; added 5 regression tests | Independent design/code audits passed | 2 source defaults + 1 test file |
+
+---
+
+## Validation Results
+
+- [x] Focused test: `bun test cli/src/components/savant-ui/echo/__tests__/sidebar-collapse.test.tsx` — **5 passed / 0 failed**
+- [x] CLI typecheck: `bun run --cwd=cli typecheck` — **passed**
+- [x] Focused lint: `bunx eslint --max-warnings 0 ...` — **passed, zero warnings**
+- [x] Focused format: `bunx prettier --check ...` — **passed**
+- [x] `git diff --check` — **passed**
+- [x] Production call graph — **confirmed**
+- [~] Interactive CLI smoke — **deferred**
+
+---
+
+## Final State
+
+### Code Changes
+
+- **Files Modified:** 2 existing source files
+- **Files Added:** 1 focused test file
+- **FID Records:** 1 closed and archived
+- **Changelog:** 1 closeout entry added
+
+### Git Status
+
+- **Branch:** `main`
+- **Uncommitted Changes:** yes; pre-existing and session changes remain uncommitted
+- **New Commits:** none; no commit or push authorized
+
+---
+
+## Open Questions
+
+- Whether to collect an interactive CLI visual capture showing the closed
+  chevrons on first paint remains a separate deferred evidence task.
+
+---
+
+## Lessons Learned
+
+- Reusable component defaults control startup visual density across every caller.
+- Nested expansion boundaries must be fixed together; folding only the parent
+  leaves expanded detail when reopened.
+- Server-render tests provide deterministic first-paint coverage without adding
+  a new UI-test dependency.
+
+---
+
+## Next Session
+
+### Priority Tasks
+
+1. [ ] Optionally collect an interactive CLI smoke capture for visual evidence.
+2. [ ] Continue with the next requested FID or implementation task.
+
+### Blockers
+
+- None for the implemented code path.
+
+### Notes for Next Agent
+
+- `FID-2026-0801-001` is closed and archived.
+- Interactive smoke is explicitly deferred, not passed.
+- Preserve unrelated working-tree changes.

package/dev/session-summaries/2026-08-02-0.0.15-release-closeout.md

@@ -0,0 +1,32 @@
+# 2026-08-02 — v0.0.15 Release and FID Closeout
+
+## Scope
+
+Completed the local-only v0.0.15 documentation and lifecycle closeout after verifying
+that FID-2026-0802-001, FID-2026-0802-002, and FID-2026-0802-003 each had a converged
+ECHO Perfection Loop, implemented changes, and verification evidence.
+
+## Completed
+
+- Added the `v0.0.15` CHANGELOG section covering lint/recovery, CommandCode provider
+  support, and first-run provider onboarding.
+- Updated the Savant-Code release boundary to `0.0.15` in `VERSION`, root/CLI/SDK
+  manifests, the CLI release manifest, `bun.lock`, and `protocol.config.yaml`.
+- Updated current README, privacy, landing-page, and launch-post version references;
+  added CommandCode to the root README gateway list.
+- Changed the three verified FIDs to `closed` and moved them to `dev/fids/archive/`
+  only after the changelog entry existed.
+
+## Verification
+
+- Archive move verified: source active paths absent, archive paths present, all three
+  statuses equal `closed`.
+- Historical release references remain historical and were not rewritten.
+- Independent version, FID lifecycle, Markdownlint, typecheck, and documentation
+  validation is recorded by the closing verification commands for this session.
+
+## Safety
+
+No GitHub, remote, commit, reset, checkout, clean, stash, or history operation was
+performed. Independent internal package versions and the separate Savant-Free version
+scheme were preserved.

package/dev/session-summaries/2026-08-02-repository-hygiene.md

@@ -0,0 +1,51 @@
+# 2026-08-02 — 0.0.15 Documentation and Repository Hygiene Closeout
+
+## Scope
+
+Reviewed current documentation for stale 0.0.15, provider, architecture, onboarding, Windows, privacy, and versioning
+claims. No GitHub, remote, history, reset, checkout, clean, commit, or publish operation was performed.
+
+## Documentation updates
+
+- Removed the obsolete pre-rebrand checkpoint from `ARCHITECTURE.md`.
+- Updated root and Chinese README guidance for the current local/BYOK release and provider onboarding.
+- Corrected CommandCode's source-of-truth variable to `COMMAND_CODE_API_KEY`.
+- Clarified that CLI `/provider` credentials and SDK backend credentials are separate contracts.
+- Updated SDK, npm release, Windows, privacy, versioning, and historical-modes documentation.
+- Preserved historical bodies while clearly labeling outdated mode analysis as historical.
+
+## Reversible archive
+
+User approved archive-only organization. Exactly 19 untracked artifacts moved to:
+
+`dev/archive/2026-08-02-repository-hygiene/`
+
+Categories:
+
+- `recovery-artifacts/`
+- `scratchpad-scripts/`
+- `release-drafts/`
+- `nova-drafts/`
+- `router-research/`
+
+`MANIFEST.before.json` and `MANIFEST.after.json` record original paths, destinations, byte counts, and SHA-256 hashes.
+All 19 files re-verified byte-for-byte. No content was deleted. The distinct router research files were preserved
+separately because their hashes differ.
+
+## FID lifecycle
+
+- FID: `FID-2026-0802-004-docs-and-repository-hygiene.md`
+- Status: closed
+- Archived: `dev/fids/archive/`
+- Changelog: `CHANGELOG.md` v0.0.15 entry
+
+## Verification
+
+- `bun run lint:md` passed.
+- Scoped `git diff --check` passed; only line-ending normalization warnings were reported.
+- Current-doc stale-reference scan passed for old release versions, pre-rebrand checkpoint text, legacy account URLs, and
+  the old CommandCode variable spelling.
+- Archive hash and source-absence checks passed for 19 of 19 files.
+- No active FIDs remain.
+
+The broader pre-existing worktree diff was not modified or normalized as part of this documentation/hygiene scope.

package/dev/session-summaries/2026-08-03-agent-roster-fix-and-ready-check-closeout.md

@@ -0,0 +1,73 @@
+# Session Summary — 2026-08-03 Agent Roster Fix + Release Ready-Check (FID-2026-0803-013)
+
+**FID:** FID-2026-0803-013 (agent roster over-reporting — 1 LOW finding)
+**Status:** Verified + archived
+**Signed:** Savant
+
+## What was done
+
+### Agent roster over-reporting fix (FID-2026-0803-013)
+
+- **Root cause:** The Savant orchestrator reported 13 spawnable agents when asked
+  "what's the roster?". The main-agent instructions prompt auto-appends a
+  functional spawn list (`strings.ts` — "You can spawn the following agents:")
+  built from the 13-entry `spawnableAgents` allowlist (9 core + `basher`,
+  `tmux-cli`, `browser-use`, `context-pruner` infrastructure + `researcher-web`/
+  `researcher-docs` as two entries), and the system prompt contained no
+  canonical roster definition — so the model parroted the spawn list back.
+- **Fix:** Added an explicit `# Agent Roster` section to the Savant default
+  system prompt (`agents/savant/savant.ts`, `buildDefaultSystemPrompt`):
+  - The 9 canonical ECHO roles with `# / Agent / Phase / Responsibility`
+    columns matching ARCHITECTURE.md (Savant/Orchestrator, Detective, Forge,
+    Verifier, Thinker, Scout, Researcher, Recorder, Scribe).
+  - An "Important distinction" subsection naming `researcher-web`/`researcher-docs`
+    as the single Researcher role's tool libraries and `basher`/`tmux-cli`/
+    `browser-use`/`context-pruner` as infrastructure helpers (spawnable but NOT
+    roster members).
+  - Explicit instruction: "When asked about the agent roster, report only the 9
+    roles listed above."
+- **Deliberately unchanged:** `spawnableAgents` allowlist (13 — functionally
+  correct; removing infrastructure would break spawning) and the `strings.ts`
+  addendum (the functional "what can I spawn" list). Operator later supplied the
+  exact canonical roster content (ARCHITECTURE.md table format + helpers
+  subsection), which replaced the initial simpler table; FID amendment recorded.
+
+### Sibling agent audit (thinker, detective, forge, verifier + scout/scribe/recorder/researcher)
+
+- **Clean — zero findings.** All 8 specialists have `spawnableAgents: []`, so
+  the "You can spawn the following agents:" addendum never renders for them
+  (structurally immune to roster over-reporting). Their prompts correctly
+  forbid spawning ("Do NOT spawn other agents", "you are the editor agent").
+- `base-chat.ts` (2 spawnables: `researcher-web`, `thinker-gemini`) and
+  `editor-multi-prompt.ts` (3 spawnables: best-of-N editor helpers) are
+  exactly-scoped and self-describing; all spawnable IDs verified to exist on
+  disk (Law 4). No roster framing, no confusion.
+- `thinker.ts` embeds `ECHO_PROTOCOL_INSTRUCTIONS` but the constant contains no
+  roster enumeration — only role references in passing. No action needed.
+
+### Release ready-check (full-session QC sweep)
+
+- **Typecheck ×4** (sdk/common/agent-runtime/cli): all exit 0.
+- **Test suites:** sdk 431 pass / 0 fail, common 521 pass / 0 fail,
+  agent-runtime 583 pass / 0 fail.
+- **ESLint** `--max-warnings 0`: clean. **`lint:md`**: exit 0.
+- **Version:** 0.0.16 consistent across VERSION + root/sdk/cli manifests.
+- **Open FIDs:** zero. **Eval:** baseline 4/4 PASS tracked in
+  `docs/reports/savant-code-benchmark-v2-2026-08-03.md`.
+- **Loose ends closed this pass:** FID-013 CHANGELOG Verification bullet added
+  (the only FID missing one); FID-008 identified as a benign numbering gap
+  (never committed — CHANGELOG correctly skips 007→009); savant-free
+  `0.0.123` npm-track version confirmed intentional (separate release track).
+
+## Verification
+
+- agents typecheck exit 0; CLI typecheck exit 0 (bundle regenerated via
+  `prebuild:agents`, embeds the roster); agent-runtime strings template suite
+  11/11 pass; ESLint `--max-warnings 0` on changed files; `bun run lint:md`
+  exit 0; forbidden-name sweep clean (Savant only). Independent AUDIT via
+  code-reviewer: clean.
+
+## Lifecycle
+
+FID-2026-0803-013 → verified + archived. CHANGELOG v0.0.16 Added + Verification
+entries present. Ready to ship as part of the v0.0.16 full release.

package/dev/session-summaries/2026-08-03-build-artifact-hygiene-closeout.md

@@ -0,0 +1,40 @@
+# Session Summary — 2026-08-03: Build Artifact Hygiene (FID-2026-0803-011)
+
+**Status:** Closed · **FID:** `dev/fids/archive/FID-2026-0803-011-build-artifact-hygiene.md`
+**Author:** Savant
+
+## What was done
+
+FID-2026-0803-011 (approved for direct implementation) went through the full
+Perfection Loop COMPLETE → IMPLEMENT → AUDIT and was archived. It grew from
+the 0803-010 amendment's follow-up note and **corrected that note's premise**
+with git-aware evidence before any code changed:
+
+- **BH-1** — corrected the FID-0803-010 amendment: `cli/bin/` is gitignored
+  (root `.gitignore:42-43` + `cli/.gitignore` `bin`); `git ls-files cli/bin/`
+  is empty — the "committed .exe binaries" wording was wrong.
+- **BH-2** — purged ~360 MB of stale local build artifacts (Jul 28-31) that
+  consumers (e2e/tmux) existence-check only; verified regenerable via
+  `build:binary` / `savant-free/cli/build.ts` / root `ci`.
+- **BH-3** — added `cli/scripts/clean.ts` + `"clean"` script; added a
+  post-compile removal of the 21 MB `index.js.map` in `build-binary.ts`
+  (bun 1.3.11 emits it despite `--sourcemap=none`; the release tarball ships
+  only binary + wasm + env.json; nothing references it).
+
+## Gates (all green)
+
+- `git ls-files cli/bin/` — 0 tracked (nothing lost in the purge).
+- Regeneration proof: `bun savant-free/cli/build.ts 0.0.0-dev` — exit 0;
+  `cli/bin/` now contains exactly `env.json`, `savant-free.exe`,
+  `tree-sitter.wasm` (no map).
+- `cd cli && bun run typecheck` — exit 0 (covers `clean.ts` + `build-binary.ts`).
+- ESLint `--max-warnings 0` on changed scripts — clean.
+- `bun run lint:md` — exit 0. Forbidden-name sweep — clean.
+- Independent code review: clean — no issues; one non-actionable nit.
+
+## Lifecycle
+
+- CHANGELOG.md — Added + Verification bullets under v0.0.16.
+- LEARNINGS.md — 4 lessons prepended (CRLF preserved).
+- FID — `verified` with full Resolution, archived.
+- Signing: Savant only.

package/dev/session-summaries/2026-08-03-code-map-audit-closeout.md

@@ -0,0 +1,44 @@
+# 2026-08-03 — Code-Map Package Audit Closeout (FID-2026-0803-006)
+
+## Scope
+
+Continuation of the 2026-08-03 quality session. Audited `packages/code-map` (tree-sitter indexing +
+language detection; ~900 lines read 0-EOF) for correctness, performance, and type safety. Produced 9
+findings (2 medium / 7 low). Opened FID-2026-0803-006, ran the Perfection Loop (RED → GREEN → AUDIT →
+SELF-CORRECT → COMPLETE), presented for operator approval, implemented, verified, and archived.
+No git commit, push, tag, or publish operation was performed.
+
+## Findings and fixes (all implemented)
+
+- **CM-1 (MEDIUM):** dead `call in {}` guard in `buildTokenCallers` (parse.ts) — a `toString`/`valueOf`
+  token collision crashed the code map. Fixed to `call in Object.prototype` + regression test.
+- **CM-2 (MEDIUM):** `UnifiedLanguageLoader` cached a `Parser.init` rejection forever (silent permanent
+  disablement). Fixed: lazy init, shared in-flight promise, clear-on-rejection retry, one-time warn.
+- **CM-3 (LOW):** dead wasm-path fallback loop → collapsed to explicit first-path return.
+- **CM-4 (LOW):** no-op `catch (err) { throw err }` removed.
+- **CM-5 (LOW):** `LanguageConfig.initPromise` dedupes concurrent lazy inits.
+- **CM-6 (LOW):** `.mjs`/`.cjs`/`.mts`/`.cts` added to the language table (same grammars).
+- **CM-7 (LOW):** read failures classify as `skipped`; stat+read single-window try/catch (TOCTOU).
+- **CM-8 (LOW):** `as { delete?: … }` cast dropped; `tree.delete?.()` retained (mock/runtime compat).
+- **CM-9 (LOW):** `getDirnameDynamically` hardened with an outer try/catch.
+
+## Verification (all green)
+
+- code-map suite: 51 pass / 0 fail (incl. 1 new CM-1 regression test)
+- code-map typecheck: clean · full-repo ESLint `--max-warnings 0`: exit 0 · `bun run lint:md`: exit 0
+- Prettier clean on all changed files
+- Independent AUDIT via code-reviewer: clean (no CRITICAL/HIGH/MEDIUM; CM-7 statSync window completed
+  in response, CM-8 `?.` deviation documented)
+- SDK smoke caveat: `smoke-test-dist` requires a built dist absent from the tree; the CJS tree-sitter
+  harness failure is pre-existing/environmental — `getFileTokenScores` export verified at source level
+  via typecheck
+
+## Files changed
+
+`packages/code-map/src/{parse,languages,utils}.ts` ·
+`packages/code-map/__tests__/{integration,languages,parse}.test.ts` (parse.test.ts unchanged — 8 mock
+tests were the CM-8 guard evidence) · `CHANGELOG.md` · `dev/LEARNINGS.md` · this summary.
+
+## Status
+
+Open: none (all FIDs archived). Working tree contains the v0.0.16 in-flight release work.

package/dev/session-summaries/2026-08-03-database-llm-providers-low-fixes-closeout.md

@@ -0,0 +1,47 @@
+# Session Summary — 2026-08-03: Database + LLM-Providers LOW Fixes (FID-2026-0803-010)
+
+**Status:** Closed · **FID:** `dev/fids/archive/FID-2026-0803-010-database-llm-providers-low-fixes.md`
+**Author:** Savant
+
+## What was done
+
+FID-2026-0803-010 (approved by the operator) was implemented through the full
+Perfection Loop COMPLETE → IMPLEMENT → AUDIT and archived. All 7 findings
+landed with zero behavioral change:
+
+- **DB-A** — `createMessage` read-back uses `requireRow(...)` instead of the
+  file's last `!` assertion (`database/src/service.ts`).
+- **DB-B** — dead pre-rebrand `agent_configs` table removed from the schema
+  (`database/src/index.ts`) plus its test-teardown consumer in
+  `service.test.ts` (scope correction documented in the Resolution).
+- **DB-C** — lazy `prepare()` statement cache replaced 20 per-call
+  `db.prepare()` sites (`database/src/service.ts`).
+- **LLM-A** — inline chat streaming transform extracted into shared
+  `chat/stream-transform.ts`; the model calls the factory and
+  `stream-transform.test.ts` now drives the REAL logic (3 tests). A
+  TransformStream backpressure hang in the rewritten test was diagnosed
+  (readable-side HWM = 1) and fixed with a concurrent-drain helper.
+- **LLM-B** — `getArgs` parses provider options once when the base key and the
+  configured provider name coincide.
+- **LLM-C** — byte-identical completion/ helper copies deleted; completion
+  model imports retargeted to `../chat/`.
+- **LLM-D** — dead `internal/index.ts` barrel deleted.
+
+## Gates (all green)
+
+- Typecheck: `packages/database` + `packages/llm-providers` — exit 0 ×2.
+- Tests: database 11/11; llm-providers 58/0 (112 expect; baseline 57 — 2
+  simulated transform tests became 3 real-transform tests).
+- ESLint `--max-warnings 0` on both packages (import/order auto-fixed).
+- Static double-audit: no `!` read-back; no `db.prepare(` in service functions;
+  `agent_configs` + `internal/index` zero hits repo-wide; completion imports
+  resolve via `../chat/`.
+- Independent code review: clean — no correctness issues; one nit applied
+  (internal token-usage type no longer exported).
+
+## Lifecycle
+
+- CHANGELOG.md — Added + Verification bullets under v0.0.16.
+- LEARNINGS.md — 4 lessons prepended (CRLF preserved).
+- FID — `verified` with full Resolution, archived.
+- Signing: Savant only.

package/dev/session-summaries/2026-08-03-echo-enforcement-doc-drift-closeout.md

@@ -0,0 +1,43 @@
+# Session Summary — 2026-08-03 — ECHO Enforcement Layer Doc Drift Closeout
+
+**Session:** Quality session continuation · **FID:** FID-2026-0803-009 · **Author:** Savant
+
+## Work Completed
+
+1. **ECHO enforcement layer audit (scan phase, prior turn):** Verified the runtime enforcement is fully
+   ECHO-compliant — FSM gating (`tool-executor.ts:386-467`), SoD (Forge no bash / Detective no writes / Verifier
+   zero-tool), all 13 Orchestrator tool names resolve to runtime handlers, protocol.config.yaml 9-workspace
+   `type_check` chain resolves, ARCHITECTURE.md 9-agent table matches definitions. Found 4 LOW doc-drift items
+   (EC-1..EC-4) in `ECHO.md` + `protocol.config.yaml`.
+2. **Database + llm-providers audit (scan phase, prior turn):** Both packages typecheck clean; db 11/11 tests,
+   llm 57/57 tests; zero `any` casts. Reported 6 LOW findings (DB-A/B/C, LLM-A/B/C) — no FID opened yet (awaiting
+   operator decision on whether to fold into a FID).
+3. **FID-2026-0803-009 perfection loop + implementation (this turn):**
+   - Re-read the single-agent ECHO adaptation spec (v0.1.2) 0-EOF (Law 1) — signing rule confirmed: **Savant only**.
+   - FID converged (RED evidence → GREEN minimal fixes → AUDIT plan).
+   - **IMPLEMENTED all 4 fixes:** ECHO.md:69 Researcher row `web_search, read_url (web); read_docs (docs)`;
+     ECHO.md:55-60 roster-intro footnote (9 canonical vs 4 infra spawnables, browser-use vs agent-definition
+     nuance); ECHO.md:64 Forge restricted cell `spawn_agents, ask_user`; protocol.config.yaml:28 build comment.
+   - **Double audit passed:** static grep of each edit + `bun run lint:md` exit 0 + YAML parse + pipe-count
+     integrity check.
+   - **Independent AUDIT (code-reviewer):** clean — one doc-precision nit on the EC-2 footnote (basher/tmux-cli/
+     context-pruner are agent definitions, not helper-lib dirs), addressed in response.
+   - **Lifecycle:** CHANGELOG Added + Verification bullets; FID → verified; LEARNINGS prepended (CRLF-preserved);
+     archived to `dev/fids/archive/`.
+
+## Gate Status
+
+| Gate | Result |
+|---|---|
+| `bun run lint:md` | exit 0 |
+| `protocol.config.yaml` YAML parse | OK |
+| ECHO.md table pipe counts | uniform (8 cols) |
+| Independent AUDIT | clean (1 nit addressed) |
+| Typecheck | N/A — doc/config only, no code touched |
+
+## Open Items / Next
+
+- **6 LOW findings from the database + llm-providers audit (DB-A..C, LLM-A..C)** — no FID yet. Operator to decide:
+  fold into one FID, or defer (they are cheap: 2 cleanup, 1 perf, 1 test-of-simulation drift, 2 micro).
+- Operators' deferred item remains untouched: pre-rebrand leftovers (`file-picker-max`/`file-lister`/scout max-mode)
+  — no action per operator instruction.

package/dev/session-summaries/2026-08-03-evals-benchmark-audit-closeout.md

@@ -0,0 +1,50 @@
+# Session Summary — 2026-08-03: Evals Benchmark Audit (FID-2026-0803-007)
+
+**Date:** 2026-08-03
+**FID:** FID-2026-0803-007 (evals benchmark runner audit) — Status: verified, archived
+**Author:** Savant
+
+## Scope
+
+Fourth quality-session surface: the evals benchmark harness
+(`evals/benchmark/agent-runner.ts`, `judge.ts`, `trace-analyzer.ts`,
+`trace-utils.ts`, `meta-analyzer.ts`, `logger.ts`, `run-benchmark.ts`,
+`eval-task-generator.ts`, `lessons-extractor.ts`, `runners/savant.ts`,
+`setup-test-repo.ts`, `gen-repo-eval.ts`, `analyze-task-scores.ts`,
+`filter-supplemental-files.ts`).
+
+## Findings (12, approved in full)
+
+- **2 HIGH** — EV-1a/1b: the evals package failed typecheck (exit 2) since v0.0.15 —
+  two botched "fixes" from FID-0802-006 that never compiled, invisible to the CI gate.
+- **4 MEDIUM** — EV-2 (unused `JudgingResultSchema`), EV-3 (timeouts don't abort the LLM run),
+  EV-4 (dead mislabeled judge), EV-5 (median picks the higher scorer).
+- **6 LOW** — EV-6 (`catch (error: any)`), EV-7 (9× `any[]`), EV-8 (untyped truncation callbacks),
+  EV-9 (defensive cast confirmed), EV-10 (bare catches), EV-11 (partial trace lost on abort).
+
+## Implementation highlights
+
+- **Headline:** `bun run typecheck` in evals now exits 0 (was 2).
+- All `withTimeout(client.run(...))` sites replaced with `signal: AbortSignal.timeout(...)`
+  (SDK already supports abort); `SavantCodeRunner` threads signal + `traceSink` so partial
+  traces survive an abort.
+- `JudgingResultSchema` and new `TraceAnalyzerResultSchema` now `safeParse` before use.
+- `judge-sonnet` deleted; median corrected to `floor((n-1)/2)`.
+- Recorded nuances: EV-10 probe-site deviation (intent comments, not warn noise),
+  EV-5 behavior change, EV-3 lessons-extractor extension.
+
+## Verification (all green)
+
+- evals typecheck exit 0 · evals 67 tests pass · sdk/common typechecks 0 errors
+- full-repo ESLint zero-warnings · markdownlint exit 0 · Prettier clean
+- Independent code review: clean (no CRITICAL/HIGH/MEDIUM); 3 notes recorded in Resolution
+
+## Lifecycle
+
+- FID-2026-0803-007 → `verified`, full Resolution recorded, archived to `dev/fids/archive/`.
+- CHANGELOG.md — Added + Verification bullets under v0.0.16.
+- dev/LEARNINGS.md — 5 lessons prepended (CRLF-preserved, lint-clean).
+- Signing: Savant only; no forbidden harness names in any new file.
+
+**Net effect:** the benchmark harness compiles again, its LLM runs actually stop when they time out,
+judge/analyzer output is validated instead of cast, and the dead/mislabeled/`any`-typed surface is gone.

package/dev/session-summaries/2026-08-03-quality-scan-hygiene-fixes-closeout.md

@@ -0,0 +1,57 @@
+# 2026-08-03 — Quality Scan Hygiene Fixes Closeout (FID-2026-0803-005)
+
+## Scope
+
+Continuation of the 2026-08-03 quality session. A fresh four-pass codebase scan (ECHO compliance /
+performance / code quality / robustness) over `packages/agent-runtime/src`, `cli/src`,
+`protocol.config.yaml`, and FID hygiene produced 6 findings (1 medium / 4 low / 1 info-by-design).
+A single FID (FID-2026-0803-005) was opened, run through the Perfection Loop (RED → GREEN → AUDIT →
+SELF-CORRECT → COMPLETE), presented for operator approval, implemented, verified, and archived.
+No git commit, push, tag, or publish operation was performed.
+
+## The scan
+
+- Four parallel passes: any/as cast inventory, silent catches + non-null assertions, hot-path sync IO +
+  timers/effects, and protocol-config/FID hygiene.
+- All `any`/`as` hits in production code resolved clean except one removable cast (C2) and test-only
+  mocks (accepted convention).
+
+## Findings and fixes (all implemented in FID-2026-0803-005)
+
+- **E1 (LOW):** `protocol.config.yaml` version drift `0.0.15` → `0.0.16` (VERSION + release manifests
+  were already 0.0.16).
+- **P1a (MEDIUM):** `checkpoint-store.ts` `captureSnapshot` conflated `ENOENT` with every read failure —
+  a non-ENOENT error (EACCES/EISDIR/EMFILE) was recorded as the delete-on-restore `null` marker, so a
+  rewind could DELETE an existing file it merely failed to read. Fixed with `paths.ts`-idiom errno
+  narrowing + a per-turn `skippedPaths` set (never serialized, never restored). 2 regression tests added.
+- **P1b (INFO):** checkpoint sync IO verified correct by design (capture-before-write ordering + per-path
+  dedup); documented, no change.
+- **C1 (LOW):** three redundant `agentTemplate!` assertions removed (`run-agent-step.ts`).
+- **C2 (LOW):** `as string[]` cast dropped in `executeCustomToolCall` (`tool-executor.ts`) — the native
+  gate already used the cast-free form.
+- **C3 (LOW):** `generator!` replaced with an explicit definite-assignment guard
+  (`run-programmatic-step.ts`) — fails diagnosably if an eval'd handleSteps function returns undefined.
+- **C4 (LOW):** `tool-stream-parser.ts` logs the tool-call input JSON.parse failure at debug level
+  (previously an invisible `catch {}`).
+
+## Verification (all green)
+
+- agent-runtime suite: 583 pass / 0 fail (incl. 2 new P1a tests)
+- 4-way typecheck (sdk / common / agent-runtime / cli): 0 errors
+- `bun x eslint . --max-warnings 0`: exit 0
+- `bun run lint:md`: exit 0
+- Prettier clean on all changed files
+- Independent AUDIT via code-reviewer: clean (no CRITICAL/HIGH/MEDIUM; one LOW note on skipped-capture
+  observability, accepted as documented debt)
+
+## Files changed
+
+`protocol.config.yaml` · `packages/agent-runtime/src/tools/handlers/tool/checkpoint-store.ts` ·
+`packages/agent-runtime/src/run-agent-step.ts` · `packages/agent-runtime/src/tools/tool-executor.ts` ·
+`packages/agent-runtime/src/run-programmatic-step.ts` · `packages/agent-runtime/src/tool-stream-parser.ts` ·
+`packages/agent-runtime/src/tools/handlers/tool/__tests__/checkpoint-store.test.ts` · `CHANGELOG.md` ·
+`dev/LEARNINGS.md` · this summary.
+
+## Status
+
+Open: none (all FIDs archived). Working tree contains the v0.0.16 in-flight release work.

package/dev/session-summaries/2026-08-03-quality-session-checkpoint-rewind-closeout.md

@@ -0,0 +1,56 @@
+# 2026-08-03 — Quality Session + Checkpoint & Rewind Closeout
+
+## Scope
+
+Multi-part session: (1) repo-wide quality audit tracks, (2) an external resource review that produced the
+Checkpoint & Rewind feature, and (3) full implementation, verification, and documentation closeout. No git
+commit, push, tag, or publish operation was performed.
+
+## Quality session tracks
+
+Four audit FIDs were opened, implemented, verified, and archived under v0.0.16 (all entries already in
+`CHANGELOG.md`):
+
+- **FID-2026-0802-008** — SDK package audit (client / run / run-state).
+- **FID-2026-0803-001** — ECHO enforcement layer drift (savant.ts + protocol.config.yaml + ARCHITECTURE.md).
+- **FID-2026-0803-002** — llm-providers + database package audit (16 findings, 1 critical).
+- **FID-2026-0803-003** — SDK impl layer + common util audit (18 findings).
+
+## Feature selection: Reasonix review
+
+Reviewed the DeepSeek-Reasonix Go codebase (resources/) for adoptable features. Shortlist: checkpoint/rewind,
+durable memory with BM25 recall, host checks in project memory, hierarchical instructions, capability
+diagnostics, profiles-as-commands. User selected **Checkpoint & Rewind** — noting Savant already has session
+persistence/restore (DB + session files + `cloneSessionState`), which correctly scoped the design.
+
+## FID-2026-0803-004 — Checkpoint & Rewind (implemented)
+
+- **Store:** `packages/agent-runtime/src/tools/handlers/tool/checkpoint-store.ts` — persistent per-turn
+  checkpoints, promoted from the zero-caller `file-snapshot-store.ts` primitive (deleted after migration).
+  `openTurn`/`captureSnapshot`/`closeTurn`/`listTurns`/`restoreTurn`/`forkFrom`; first-capture-wins dedup;
+  `content: null` ⇒ delete-on-restore; `resolveAndContain` re-validation; retention 20;
+  `path.basename` turnId guard; deterministic `listTurns` tiebreaker (same-millisecond opens).
+- **Capture seam:** `executeToolCall` write-gate capture before `write_file`/`str_replace`/`apply_patch`
+  dispatch; `checkpointDir`/`checkpointTurnId` threaded through `AgentRuntimeScopedDeps` → SDK `RunOptions`
+  → subagent spawn context (subagent writes inherit the parent turn).
+- **CLI UX:** `/rewind` command + OpenTUI `RewindPicker` + `rewind-picker-store`; `executeRewind` with
+  code / conversation / both / fork modes; transcript truncation + session fork via `rewind.ts`;
+  turn lifecycle (openTurn at aiMessageId, closeTurn in `finally`) in `use-send-message`.
+- **Tests:** 15 checkpoint-store + 10 rewind-command. Independent AUDIT via code-reviewer (clean; turnId
+  guard, crash-recovery docs, and 2 tests added in response).
+
+## Documentation closeout (this entry)
+
+- `CHANGELOG.md` — v0.0.16 Added + Verification entries for FID-2026-0803-004 (all five session FIDs present).
+- Version bumped 0.0.15 → 0.0.16: `VERSION`, root/`sdk`/`cli`/`cli-release` package.json manifests.
+- `README.md` — badge + release header updated; Checkpoint & Rewind feature bullets under CLI, SDK, and
+  Agent Runtime; `/rewind` added to the slash-commands list.
+- `sdk/README.md` — new "Checkpoint & Rewind" API Reference section with usage example.
+- `docs/agents-and-tools.md` — checkpointing note under the File Operations tool catalog.
+- `dev/LEARNINGS.md` — session entry appended.
+- FID-2026-0803-004 archived at `dev/fids/archive/`; no open FIDs remain.
+
+## Gates (all green)
+
+4-way typecheck (sdk/common/agent-runtime/cli), zero-warning ESLint, `bun run lint:md`, Prettier clean on all
+changed files, and full suites at 0 fail (agent-runtime 581, CLI 2740, SDK 430).