progress.txt

## Progress Log

### 2026-02-26 — skill-opencode-integration plan (Tapestry)

Completed all 11 tasks in `.weave/plans/skill-opencode-integration.md` — replaced Weave's custom filesystem skill scanning with OpenCode's `GET /skill` HTTP API.

**Philosophy**: Users manage skills in one place (OpenCode's standard skill locations: `~/.config/opencode/skills/` and `.opencode/skills/`). Weave's specialized agents consume them via explicit config assignment in `weave-opencode.jsonc`. No more `weave-opencode/skills/` directory.

**New files:**
- `src/features/skill-loader/opencode-client.ts` — `fetchSkillsFromOpenCode(serverUrl, directory)` calls `GET /skill`, maps response to `LoadedSkill[]`, graceful fallback on error
- `src/features/skill-loader/opencode-client.test.ts` — 10 tests covering mapping, scope derivation, error handling, URL encoding

**Modified files:**
- `src/features/skill-loader/loader.ts` — `loadSkills()` now async, uses `fetchSkillsFromOpenCode` instead of filesystem scanning; `LoadSkillsOptions` gains `serverUrl` field
- `src/features/skill-loader/loader.test.ts` — fully rewritten with mocked HTTP responses (7 tests)
- `src/features/skill-loader/index.ts` — updated barrel: removed deleted exports, added `fetchSkillsFromOpenCode`
- `src/features/skill-loader/types.ts` — `LoadedSkill.scope` made optional; `SkillMetadata` removed (no longer needed)
- `src/create-tools.ts` — passes `ctx.serverUrl` to `loadSkills()`
- `src/config/schema.ts` — removed dead `SkillsConfigSchema`, `SkillsConfig`, and `skills` field from `WeaveConfigSchema`
- `src/config/merge.ts` — removed dead `skills` field merge
- `src/config/loader.ts` — added optional `homeDir` override parameter for test isolation
- `src/config/loader.test.ts` — fixed pre-existing test isolation bug (user's real config was leaking into tests)

**Deleted files:**
- `src/features/skill-loader/discovery.ts` — filesystem scanning replaced by SDK
- `src/features/skill-loader/discovery.test.ts`
- `src/features/skill-loader/merger.ts` — priority merging now handled by OpenCode
- `src/features/skill-loader/builtin-skills.ts` — vestigial empty array

**Key design decisions:**
- Raw `fetch()` to `GET /skill` rather than v2 SDK (v1 client has no `skills()` method; raw fetch avoids SDK version mismatch)
- `scope` derived from `location` field: contains `.opencode` → project, else → user
- Agent-specific skill assignment unchanged: `agents.tapestry.skills: ["skill-name"]` in `weave-opencode.jsonc`
- Override skill resolution in `builtin-agents.ts` untouched — source-agnostic, works with any `resolveSkills` function

536 tests pass. Build and typecheck clean.

### 2026-02-26 — tapestry-revert-v052: Full revert to v0.5.2 + cherry-pick Tapestry Verification

Completed all 5 tasks in `.weave/plans/tapestry-revert-v052.md` — reverted all 14 files changed since v0.5.2 back to their tagged state, then re-applied only the Tapestry `<Verification>` section and step 3c reference.

**Files changed:**
- `src/agents/tapestry/default.ts` — step 3c now references `<Verification>` protocol; `<Verification>` section added between `</PlanExecution>` and `<Execution>`; all other content is byte-identical to v0.5.2 (including step 5 post-execution review wording)
- `src/agents/tapestry/index.test.ts` — 7 Verification-related tests added; 7 original v0.5.2 tests preserved; total 14 tests

**Reverted to v0.5.2 (no post-v0.5.2 content):**
- `src/agents/loom/default.ts` — restored mandatory Step 5 post-execution review in PlanWorkflow
- `src/agents/loom/index.test.ts` — restored 3 deleted tests, reverted renamed test
- `src/hooks/start-work-hook.ts` — reverted "Loom" references back to "Pattern"
- `src/hooks/start-work-hook.test.ts` — reverted test assertions
- `src/hooks/verification-reminder.ts` — reverted to Weft/Warp delegation with git diff and automated checks
- `src/hooks/verification-reminder.test.ts` — reverted test assertions
- `src/hooks/work-continuation.ts` — removed switchAgent/clearWorkState on completion
- `src/hooks/work-continuation.test.ts` — removed 7 new tests
- `src/plugin/plugin-interface.ts` — removed agent passing to promptAsync
- `src/plugin/plugin-interface.test.ts` — removed 7 new tests
- `src/workflow.test.ts` — reverted integration tests
- `progress.txt` — reverted to v0.5.2 state (this entry is the only post-v0.5.2 addition)

**Result:** 519 tests pass, 0 failures. Non-Tapestry files are byte-identical to v0.5.2.

### 2026-02-22 — plan-validation plan: Add plan validation before /start-work (Tapestry)

Completed all 11 tasks in `.weave/plans/plan-validation.md` — added `validatePlan()` function that structurally validates plan files before Tapestry begins execution.

**New files:**
- `src/features/work-state/validation-types.ts` — `ValidationSeverity`, `ValidationIssue`, `ValidationResult`, `ValidationCategory` types
- `src/features/work-state/validation.ts` — `validatePlan(planPath, projectDir)` implementing 6 validation categories: structure, checkboxes, file-references, numbering, effort-estimate, verification
- `src/features/work-state/validation.test.ts` — 26 unit tests covering all scenarios

**Modified files:**
- `src/features/work-state/index.ts` — exports `validatePlan` and all new types
- `src/hooks/start-work-hook.ts` — integrates `validatePlan()` into all 3 paths (explicit, discovery, resume); errors block state creation; warnings are appended to context injection; added `formatValidationResults()` export
- `src/hooks/start-work-hook.test.ts` — fully rewritten with valid plan helper and new validation tests (19 tests)
- `src/workflow.test.ts` — added `makeValidPlanContent()` helper; updated 9 tests to use valid plan content and corrected count assertions (plans now have TODOs + Verification checkbox)

**Key design decisions:**
- Errors (missing required sections, no checkboxes, duplicate numbers, empty verification) are **blocking** — no work state created
- Warnings (missing optional sections, missing sub-fields, file references, number gaps) are **non-blocking** — appended to context injection
- Resume path also validates — malformed resumed plans clear state and return error
- `extractSection()` uses line-by-line parsing (not regex) to avoid `$` lookahead bug with multiline flag

493 tests pass. `bun run typecheck` and `bun run build` clean.

**GitHub**: Implements issue #3

---

### 2026-02-22 — Fix #1: Wire context window monitor to real token data (Tapestry)

Fixed the context window monitor never firing because it was called with hardcoded `{ usedTokens: 0, maxTokens: 0 }`.

**Root cause**: `plugin-interface.ts` called `checkContextWindow({ usedTokens: 0, maxTokens: 0 })` from the `chat.message` hook, where token data is not available.

**Changes**:
- `src/hooks/session-token-state.ts` (NEW): In-memory `Map<sessionId, {maxTokens, usedTokens}>` with `setContextLimit`, `updateUsage`, `getState`, `clearSession`, `clear`
- `src/hooks/session-token-state.test.ts` (NEW): 9 unit tests covering all state operations
- `src/hooks/create-hooks.ts`: Reads `experimental.context_window_warning_threshold` and `experimental.context_window_critical_threshold` from config (defaults 0.8/0.95)
- `src/hooks/create-hooks.test.ts`: 3 new threshold tests
- `src/hooks/index.ts`: Exports session token state functions
- `src/plugin/plugin-interface.ts`: `chat.params` captures `model.limit.context` per session; `event` handler processes `message.updated` (assistant role, `tokens.input > 0`) to call `checkContextWindow` with real data; `session.deleted` clears session state; hardcoded zero call removed from `chat.message`
- `src/plugin/plugin-interface.test.ts`: 7 new context window monitoring tests

460 tests pass. Build and typecheck clean.

**GitHub**: Fixes issue #1

---

### 2026-02-22 — Fix #15: PlanWorkflow/ReviewWorkflow contradiction (Tapestry)

Fixed a contradiction in `src/agents/loom/default.ts` where the PlanWorkflow marked Weft review as "(optional)" while ReviewWorkflow listed mandatory criteria that clearly applied to multi-step plans touching 3+ files.

**Root cause**: The `REVIEW (optional)` label on line 92 gave Loom an easy out to skip plan review, even when all ReviewWorkflow trigger conditions were met.

**Changes**:
- `src/agents/loom/default.ts`: Removed "(optional)", added explicit TRIGGER (3+ files OR 5+ tasks → mandatory) and SKIP ONLY IF (user says "skip review") conditions
- `src/agents/loom/index.test.ts`: Added 3 regression tests asserting the review step is not optional and contains trigger/skip conditions

440 tests pass. Build clean.

**GitHub**: Fixes issue #15

---

### 2026-02-22 — Add repository dispatch to notify weave-website on publish

Added a step to `publish.yml` that sends a `repository_dispatch` event to `pgermishuys/weave-website` after npm publish. This triggers the website to update its hardcoded version badge in `index.html` and redeploy, so users with JS disabled see the correct version.

**Requires**: `WEBSITE_DISPATCH_TOKEN` secret (PAT with `repo` scope) in this repo's GitHub settings.

**Files changed**: `.github/workflows/publish.yml`

---

### 2026-02-22 — Fix delegation logging field mismatch

Fixed broken delegation observability logging from v0.4.0 commit.

**Root cause**: `tool.execute.before` and `tool.execute.after` hooks read `args.agent` to identify the delegated agent, but the OpenCode task tool uses `subagent_type` (not `agent`). This meant:
- `tool.execute.after` always logged `agent=unknown` in production
- `tool.execute.before` fell back to `args.description` (logged description text instead of agent name)
- Tests used fake `{ agent: "thread" }` data, masking the bug

**Changes**:
- `src/plugin/plugin-interface.ts`: Changed both hooks to read `subagent_type` first, then `description` as fallback
- `src/plugin/plugin-interface.test.ts`: Updated tests to use real SDK data shape (`subagent_type`), added fallback-to-description test

All 437 tests pass. Build clean.

---

### 2026-02-21 — todowrite-sidebar-integration plan (Tapestry)

Completed all 6 tasks in the `todowrite-sidebar-integration` plan:

1. **Tapestry prompt** (`src/agents/tapestry/default.ts`): Added `<SidebarTodos>` section with concrete instructions for when/how to write sidebar todos during plan execution (STARTING, COMPLETING, BLOCKED, COMPLETES states). Max 35 chars, max 5 todos, `in_progress` for current task only.

2. **Loom prompt** (`src/agents/loom/default.ts`): Added `<SidebarTodos>` section with delegation-aware todowrite instructions (PLANNING, DELEGATING, QUICK TASKS states). Agent-prefixed delegations, clear completion signaling.

3. **start-work-hook** (`src/hooks/start-work-hook.ts`): Enhanced `buildFreshContext()` and `buildResumeContext()` to inject `**SIDEBAR TODOS**` bootstrapping instructions. Resume context includes remaining task count.

4. **context-window-monitor** (`src/hooks/context-window-monitor.ts`): Warning and recovery messages now include `todowrite` instructions to surface context alerts in the sidebar with `in_progress` status for visual prominence.

5. **work-continuation** (`src/hooks/work-continuation.ts`): Continuation prompt now instructs the agent to restore sidebar state via `todowrite` (summary + next task + 2-3 upcoming) before resuming work.

6. **Tests updated**: Added "SIDEBAR TODOS" assertions to start-work-hook tests, "todowrite" assertions to context-window-monitor and work-continuation tests.

All 404 tests pass. Build succeeds.

**Next**: This plan is complete. No follow-on work identified.

---

### 2026-02-21 — warp-agent plan (Tapestry)

Completed all 15 tasks in the `.weave/plans/warp-agent.md` plan — added **Warp** as the 8th built-in agent (security & specification compliance auditor).

**New files:**
- `src/agents/warp/default.ts` — full system prompt (Role, Triage, SecurityReview, SpecificationCompliance, Verdict, SkepticalBias, Constraints)
- `src/agents/warp/index.ts` — factory function following Weft pattern
- `src/agents/warp/index.test.ts` — 15 unit tests

**Modified source files:**
- `src/agents/types.ts` — added `"warp"` to `WeaveAgentName` union
- `src/agents/builtin-agents.ts` — imported and registered `createWarpAgent`
- `src/agents/model-resolution.ts` — Sonnet-class fallback chain for warp
- `src/agents/dynamic-prompt-builder.ts` — added `buildWarpSection` function
- `src/agents/loom/default.ts` — Warp added to Delegation and ReviewWorkflow sections
- `src/hooks/verification-reminder.ts` — Warp reference for security-relevant changes
- `src/shared/agent-display-names.ts` — added `warp: "warp"` entry

**Updated tests:** builtin-agents.test.ts, dynamic-prompt-builder.test.ts, agent-display-names.test.ts, workflow.test.ts, integration.test.ts (all counts updated from 7→8 agents).

**Updated docs:** README.md, docs/architecture.md, docs/agent-interactions.md, docs/configuration.md.

Key design decisions:
- Warp is **read-only** (write, edit, task, call_weave_agent all denied)
- **Skeptical bias** — REJECT by default when security patterns detected
- **Sonnet-class model** (needs reasoning power for spec analysis)
- **Self-triaging** — fast-exits with [APPROVE] if no security-relevant changes detected
- Three-layer spec knowledge: built-in LLM → webfetch RFC verification → optional `.weave/specs.json`

All 427 tests pass. Build and typecheck clean (zero warnings).

**Next**: Warp is fully integrated. Future work could add `.weave/specs.json` support to the config schema if project-level spec registries are needed.

---

### 2026-02-21 — Release v0.3.1

Released **v0.3.1** to GitHub.

- Optimized agent model assignments for role-appropriate reasoning:
  - **Weft** (reviewer): haiku-4.5 → sonnet-4.6 — better code review quality
  - **Warp** (security auditor): sonnet-4.6 → opus-4.6 — deeper security reasoning
  - **Spindle** (external researcher): sonnet-4.6 → haiku-4.5 — cost efficiency for fetch-and-format tasks
- All 427 tests passing, build clean

**Previous release**: v0.3.0 → **New release**: v0.3.1

---

### 2026-02-21 — Release v0.3.0

Released **v0.3.0** to GitHub: https://github.com/pgermishuys/weave/releases/tag/v0.3.0

- Bumped `package.json` version from `0.2.0` → `0.3.0`
- Created git tag `v0.3.0` and pushed to remote
- Created GitHub release with full changelog highlighting the Warp agent addition
- All 427 tests passing, build clean

**Previous release**: v0.2.0 → **New release**: v0.3.0