feat: code-owned run engine, enforced verification, real topology registry (v0.3.0)

.github/workflows/ci.yml

@@ -10,6 +10,7 @@ jobs:
   build:
     runs-on: ${{ matrix.os }}
     strategy:
+      fail-fast: false
       matrix:
         os: [ubuntu-latest, macos-latest, windows-latest]
     steps:
@@ -25,11 +26,57 @@ jobs:
       - name: Typecheck
         run: bun run typecheck
 
+      - name: Unit tests (engine, event log, config injection, guards)
+        run: bun test
+
       - name: Build
         run: bun run build
 
-      - name: Smoke test
+      - name: Engine end-to-end (real git worktrees, real evidence capture)
+        run: bun run e2e
+
+      - name: Plugin integration (built bundle, real hook shapes)
+        run: bun run integration
+
+      - name: Installer smoke test
         run: bun run smoke
 
-      - name: Verify templates shipped
-        run: bun -e "import { existsSync } from 'node:fs'; const files = ['dist/cli/templates/orchestrator.md', 'dist/cli/templates/patterns/long-proof.md', 'dist/cli/templates/commands/teamwork.md', 'dist/cli/templates/prompts/orchestrator.txt']; for (const f of files) { if (!existsSync(f)) { console.error('Missing: ' + f); process.exit(1); } }"
+      - name: Verify templates and engine modules shipped
+        shell: bash
+        run: |
+          for f in \
+            dist/cli/templates/sentinel.md \
+            dist/cli/templates/patterns/long-proof.md \
+            dist/cli/templates/commands/teamwork.md \
+            dist/cli/templates/prompts/sentinel.txt \
+            dist/engine.js \
+            dist/events.js \
+            dist/policy.js ; do
+            if [ ! -f "$f" ]; then echo "Missing: $f"; exit 1; fi
+          done
+          echo "all expected build outputs present"
+
+  pack-install:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: oven-sh/setup-bun@v2
+        with:
+          bun-version: latest
+
+      - run: bun install --frozen-lockfile
+
+      - name: Build and pack
+        run: bun run build && npm pack
+
+      - name: Install the packed tarball into a clean project
+        shell: bash
+        run: |
+          mkdir -p /tmp/packtest && cd /tmp/packtest
+          npm init -y >/dev/null
+          npm install "$GITHUB_WORKSPACE"/opencode-teamwork-*.tgz >/dev/null
+          test -f node_modules/opencode-teamwork/dist/engine.js
+          test -f node_modules/opencode-teamwork/dist/cli/templates/patterns/long-proof.md
+          test -f node_modules/opencode-teamwork/SKILL.md
+          node -e "const p=require('./node_modules/opencode-teamwork/package.json');console.log('packed package ok:',p.version);if(!p.files.includes('SKILL.md'))process.exit(1)"

CHANGELOG.md

@@ -4,6 +4,83 @@ All notable changes to opencode-teamwork are documented here. The format
 follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/).
 
+## [0.3.0] — 2026-09-19
+
+**The run loop moved into code.** Until now the "DAG engine", worktree
+isolation, cost tracking and checkpoint/resume were instructions in a prompt:
+`src/worktree.ts`, `src/cost.ts`, `src/state.ts` and `src/artifacts.ts` were
+imported by nothing, so none of those guarantees were enforced by the runtime.
+
+### Added
+- **`src/engine.ts`** — a real scheduler: plan validation (cycles, dangling
+  dependencies, duplicate ids, unknown topologies), topological dispatch waves,
+  a concurrency cap, attempt accounting, per-round model escalation, dead-letter
+  after `maxRounds`, and budget enforcement that refuses dispatch at the cap.
+- **`src/events.ts`** — append-only `events.jsonl` with a sha256 hash chain.
+  `state.json` is now a *derived* snapshot; `deriveSession()` replays a run and
+  `verifyChain()` refuses a tampered log.
+- **Engine tools**: `teamwork_plan`, `teamwork_dispatch`, `teamwork_verify`,
+  `teamwork_status`, `teamwork_resume`.
+- **`src/guard.ts`** — runtime role enforcement. Read-only roles
+  (verifier, falsifier, scout) cannot call write tools even if a host build
+  ignores part of the permission block.
+- **`src/policy.ts`** — one source of truth for topology names, model ladders,
+  required checks and budget. A test asserts every topology resolves to a
+  pattern file.
+- **Verification evidence**: checks now carry `cmd`, `exitCode`, `stdoutSha256`
+  and `durationMs`. A PASS with no executed check, or with a check that
+  contradicts its exit code, is rejected and does not count as a round.
+- **Command flags parsed in code** — `--topology`, `--budget`, `--concurrency`,
+  `--session`, with an unknown topology reported instead of accepted.
+- **Long runs survive compaction**: the `experimental.session.compacting` hook
+  re-injects the plan, task states and budget from the event log.
+- **Tests**: 26 unit tests (`bun test`), a real end-to-end run against a git
+  repository (`bun run e2e`), and a built-bundle integration test that drives the
+  actual hook shapes OpenCode sends (`bun run integration`). CI runs all three on
+  Linux, macOS and Windows, plus a `npm pack` → clean-install check.
+- `docs/self-improvement.md` — the design for improving the *policy* (not the
+  models) from the event log, with the guard metrics and failure modes.
+
+### Fixed
+- **Agent frontmatter was inert.** The plugin passed each agent file's whole
+  markdown — frontmatter included — as the `prompt` string, so `mode` defaulted
+  to `all`, `permission: edit: deny` on the verifier did nothing, `temperature`
+  was ignored, and `hidden` is not a config key at all. Frontmatter is now
+  parsed and injected as real `AgentConfig` keys.
+- **The plugin clobbered per-role models.** `config.agent[name] = { prompt }`
+  replaced whatever the installer had written, discarding the model map. Config
+  is now spread-merged, and a test asserts the user's model survives.
+- **Three incompatible topology vocabularies.** The pattern files said
+  `long-proof` / `distributed-coding` / `document-review` while the sentinel
+  prompt and `state.ts` said `proof` / `large-swarm` / `doc-review` — four of
+  five names in the routing table matched no file.
+- **Pattern files never reached the model.** `getAllPatterns()` had no call
+  sites and no prompt referenced `patterns/`. Orchestrating agents now get a
+  generated index with absolute paths.
+- **A template language OpenCode cannot expand.** `{{if eq .sessionId ""}}` in
+  the `/teamwork` command was passed to the model as literal text (only
+  `$ARGUMENTS`, `$1..$N`, `` !`cmd` `` and `@file` are supported), and
+  `$ARGUMENTS` appeared twice, duplicating every request.
+- **Shell injection surface in the worktree manager.** Git was invoked by
+  building command strings from model-produced names. All git calls now use
+  argv arrays with `shell: false`, and agent/session names are validated.
+- **`cleanupSession` never cleaned up.** It ran `git worktree remove` against
+  the worktrees' *parent* directory (never a worktree), swallowed the failure,
+  then deleted directories git still had registered. It now removes real
+  worktrees, deletes the session's branches, prunes, and verifies.
+- **`dist/` layout did not match `package.json` exports.** `build:support`
+  emitted `dist/src/*.js` while `exports` pointed at `dist/*.js`.
+- **`SKILL.md` now ships.** It was in the repo but not in `files`.
+- Leaf agents get `task: deny`; only the sentinel and v1 orchestrator can spawn
+  subagents, so a worker can no longer fan out its own swarm.
+
+### Changed
+- `src/state.ts` is a compatibility shim over the event log;
+  `saveSessionState()` now throws with guidance rather than accepting a
+  model-authored snapshot.
+- `bun run verify` is the single gate: typecheck → unit → build → smoke → e2e →
+  integration. `prepublishOnly` runs it.
+
 ## [0.2.1] — 2026-08-28
 
 CLI UX fix.