WasmAgent · telleroutlook · Aug 17, 2026 · Aug 17, 2026
diff --git a/docs/15-milestones.md b/docs/15-milestones.md
@@ -0,0 +1,290 @@
+# Milestones
+
+> Bootstrapped by **SweepRoadmapSync**. Each milestone is derived from an
+> existing planning document in `docs/` and converted to the standard
+> roadmap-pipeline format. Acceptance criteria below are the **binding**
+> definition of done — a milestone is complete only when every criterion
+> holds in the worktree.
+
+---
+
+## M1 — Parallel job queue (B1)
+
+**Source:** [`docs/B1-job-queue.md`](./B1-job-queue.md)
+
+**Goal:** Background-task dashboard — submit several independent tasks at
+once, run them in parallel up to a concurrency cap, and let every finished
+job drop a PR via `create_github_pr` without waiting for the others.
+
+### Acceptance criteria
+
+- [x] `POST /jobs` accepts `{ task }` or `{ jobs: [{task}, ...] }`; up to 20
+      per request, each an independent run with its own `traceId`.
+- [x] `GET /jobs?sessionId=…&status=…` lists jobs newest-first.
+- [x] `GET /jobs/:id` snapshots a single job (status, eventCount, last
+      events, finalAnswer, error).
+- [x] `DELETE /jobs/:id` cooperatively aborts via `AbortSignal`.
+- [x] A `/jobs` Next.js page renders a `<JobsPanel />` multi-run dashboard.
+- [x] Concurrency cap defaults to 4 and is configurable; over-cap
+      submissions queue (`status: "queued"`) and start as slots free.
+- [x] `runningCount` / `pendingCount` are exposed via `GET /jobs?stats`.
+- [x] Terminal-state jobs are mirrored under `job:<id>` for 24h in KV when
+      `config.sessionsKv` is bound; live event tails are transient.
+- [x] `JobQueue` unit tests (`apps/worker/src/jobs/queue.test.ts`) pass
+      (10 tests).
+- [x] HTTP-level integration tests in `apps/worker/src/app.test.ts` under
+      "Job queue (B1)" pass (8 tests).
+- [x] `apps/web/src/components/JobsPanel.test.tsx` passes (5 panel UI
+      tests).
+
+---
+
+## M2 — Closed-loop validation (B2)
+
+**Source:** [`docs/B2-validation-loop.md`](./B2-validation-loop.md)
+
+**Goal:** Let the worker-side agent see whether the project it just wrote
+actually compiles in the user's browser, so it can fix dependency typos and
+build errors without human intervention.
+
+### Acceptance criteria
+
+- [x] `POST /build-result` accepts browser-written snapshots keyed by
+      `X-Session-Id` (defaults to `default` session when header absent).
+- [x] `GET /build-result` provides debug readback; `DELETE /build-result`
+      clears the snapshot.
+- [x] The `read_build_result` tool is registered only in framework mode
+      (`body.framework` set) AND when `X-Session-Id` is present.
+- [x] The tool's returned string lets the model branch on `success |
+      failed | running | unknown`, including exit code, preview URL and
+      stderr tail on failure.
+- [x] Store keeps latest snapshot per session in memory and mirrors to
+      `BSCODE_BUILD_RESULTS` KV when bound.
+- [x] `BUILD_VALIDATION` system prompt in
+      `apps/worker/src/agents/prompts.ts` instructs the agent to call
+      `read_build_result` after writing files and patch on failure.
+- [x] Store tests (`apps/worker/src/build-results.test.ts`), tool tests
+      (`apps/worker/src/tools/build-result.test.ts`), and HTTP route tests
+      under "Build result reverse channel (B2)" in
+      `apps/worker/src/app.test.ts` all pass.
+- [x] `useWebContainer()` reporter fires at install start/end,
+      dev-server ready, and build error lifecycle points with no new return
+      value.
+
+---
+
+## M3 — GitHub repo import & true embedding (B3)
+
+**Source:** [`docs/B3-github-import.md`](./B3-github-import.md)
+
+**Goal:** Pull a real GitHub repository into the worker's KV file store so
+the agent can `read_file`, `search_code`, `semantic_search`, and
+ultimately `create_github_pr` against an existing codebase; plus wire the
+default TF-IDF embedder to a real-vector one via `@wasmagent/tools-rag`.
+
+### Acceptance criteria
+
+- [x] `POST /import/github` imports a repo given `owner`, `repo`, optional
+      `ref`, optional `token`, optional `paths` filter and optional
+      `textExtensions` allowlist.
+- [x] `ref` omitted resolves the repo's `default_branch` via
+      `GET /repos/{owner}/{repo}`.
+- [x] Recursive tree endpoint propagates `truncated: true` unchanged.
+- [x] Per-file blob fetch failures are collected into `skippedReasons`
+      without aborting the whole import.
+- [x] Binary detection skips files with > 20% control-byte density in the
+      first 4 KB; tunable via `textExtensions`.
+- [x] Files larger than 200 KB are skipped; total imports capped at 2000
+      files per call (constants in
+      `apps/worker/src/tools/githubImport.ts`).
+- [x] Imported files are upserted into the semantic index when an indexer
+      is bound, so the agent's `semantic_search` sees the imported tree.
+- [x] Secret-file deny-list in
+      `apps/worker/src/tools/importDenyList.ts` drops matching files
+      silently before any content is read into KV; the list cannot be
+      overridden from request input.
+- [x] Unit tests (`apps/worker/src/tools/githubImport.test.ts`, 8) and
+      route tests ("GitHub repo import (B3)" in `app.test.ts`, 4) pass.
+
+---
+
+## M4 — Tiered approval policy (B4)
+
+**Source:** [`docs/B4-approval-policy.md`](./B4-approval-policy.md)
+
+**Goal:** Per-call approval rules for write tools, adapting Vercel AI SDK
+6's `needsApproval`-as-function contract with three presets and custom rule
+support via the `POST /run` body.
+
+### Acceptance criteria
+
+- [x] `PolicyPresets.permissive()`, `balanced()`, `strict()` exist and
+      behave as documented.
+- [x] `balanced` gates `.env*`, `.dev.vars`, `.github/`, `wrangler.toml`,
+      every `delete_file` / `rename_file`, and writes exceeding 5 KB.
+- [x] Rules evaluation is first-match-wins in registration order; the
+      default verdict covers non-matching paths.
+- [x] `POST /run` accepts an inline `approvalPolicy` object literal, or a
+      preset name (`"balanced"`, `"strict"`, `"permissive"`); omitting it
+      preserves legacy permissive behaviour.
+- [x] When `needsApproval` returns true the run suspends with an
+      `await_human_input` event, the checkpoint snapshot persists the
+      pending request, and a follow-up `POST /run` with `humanResponse`
+      unblocks the run.
+- [x] `policy.explain({ op, path, sizeChars })` returns which rule fired
+      (or that none did) for diagnostics.
+- [x] `apps/worker/src/policies/approvalPolicy.test.ts` covers 12
+      scenarios: default verdict, first-rule-wins ordering, prefix
+      matching, op filtering, size gating, `explain()` output, and each
+      preset's documented behaviour.
+
+---
+
+## M5 — SSE `Last-Event-ID` resume (C1)
+
+**Source:** [`docs/C1-sse-resume.md`](./C1-sse-resume.md)
+
+**Goal:** Reuse wasmagent-js core's `EventLog` primitive so a dropped TCP
+connection is recoverable via pure replay — no agent re-invocation, no
+duplicate side-effects.
+
+### Acceptance criteria
+
+- [x] Live runs respond with `X-Wasmagent-Trace-Id` and
+      `Access-Control-Expose-Headers: X-Wasmagent-Trace-Id`, and each SSE
+      frame carries `id: <12-digit zero-padded seq>`.
+- [x] Resume requests accept a `Last-Event-ID` header and
+      `resumeTraceId` body field; when `checkpointsKv` is bound the worker
+      takes a replay-only fast path that never invokes the agent factory.
+- [x] Resume responses carry `X-Bscode-Resume: 1`.
+- [x] `EventLog.purge(traceId)` runs best-effort after a successful run.
+- [x] Degradation: with `resumeTraceId` but no bound `checkpointsKv`,
+      falls through to a fresh run without `X-Bscode-Resume`.
+- [x] `apps/worker/src/app.test.ts` "C1 — SSE Last-Event-ID resume" block
+      (6 tests) passes, including the no-reinvocation mock counter
+      assertion.
+- [x] `packages/react/src/useAgentRun.test.ts` "resume request shaping"
+      block (4 tests) passes, including verbatim survival of caller auth
+      headers.
+
+---
+
+## M6 — RLAIF rollout adapter
+
+**Source:** [`docs/rlaif-adapter.md`](./rlaif-adapter.md)
+
+**Goal:** Bridge bscode's build-result KV channel to wasmagent-js'
+`BuildPassesVerifier` / `VisualAssertVerifier` callbacks and export AEP
+evidence records for the training loop.
+
+### Acceptance criteria
+
+- [x] `makeBuildResultReader(kv?)` returns a `BuildResultReader` mapping
+      bscode statuses: `success→success`, `failed→failure`, `running` and
+      `unknown` always fail the verifier.
+- [x] `makeVisualResultReader(kv?)` maps `visual.verdict.matchesIntent` to
+      `pass`/`fail` and falls back to inferring from `consoleErrors`,
+      `uncaughtErrors`, and `rendersNonEmpty` when no explicit verdict is
+      present.
+- [x] `AppConfig.rolloutConcurrency` configures the `JobQueue` cap
+      (default 4).
+- [x] All verifier calls use derived job session IDs from
+      `deriveJobSessionId()` — never the `"default"` bucket; calling with
+      the default session logs a `console.warn`.
+- [x] `{ strictKvMode: true }` makes `putBuildResult` throw on KV write
+      failure in batch/rollout contexts.
+- [x] `buildAEPEvidence(opts)` is async, signs the record via
+      `@wasmagent/aep`'s `LocalEd25519Signer`, and populates `actions[]`,
+      `verifier_results[]`, `capability_decisions[]`, `budget_ledger`,
+      `input_refs[]`/`output_refs[]`, and
+      `signature: {alg:"ed25519", key_id, sig}`.
+- [x] Evidence export validates against the AEP schema before training use
+      (trace-pipeline `evomerge validate-aep` gate).
+
+---
+
+## M7 — Multi-agent modes (parallel / planFirst)
+
+**Source:** [`docs/multi-agent-modes.md`](./multi-agent-modes.md)
+
+**Goal:** `agentMode: "multi"` picks one of two parallel-friendly shapes —
+`parallel` (fork-join draft → reviewer) or `planFirst` (planner → human
+approval → executor) — with the old serial layout removed entirely.
+
+### Acceptance criteria
+
+- [x] `multiAgentMode: "parallel"` (default) forks into N independent
+      draft branches via `ParallelForkJoinRunner` and threads the
+      synthesised draft into a Stage-2 `createToolAgent` reviewer with the
+      full tool set.
+- [x] Defaults `branches=3`, `concurrency=2`, `aggregation="summary"` are
+      all overridable per call.
+- [x] `multiAgentMode: "planFirst"` emits
+      `{ phase: "plan_ready", plan, step }` then `await_human_input`
+      (default `promptId: "approve-plan"`), and pauses with the snapshot
+      persisted under a `planfirst-<checkpointId>-<ts>` traceId.
+- [x] Resume with `humanResponse` anything other than `"yes"`
+      (case-insensitive) includes the response verbatim as "User feedback
+      on the plan:" in the executor brief.
+- [x] The old serial multi-agent layout is gone; no compatibility mode.
+- [x] Both modes compose with B4's `approvalPolicy`.
+- [x] `apps/worker/src/agents/multi-agent.test.ts` (6 tests) passes:
+      parallel emit pattern, planFirst `await_human_input` contract,
+      custom promptId, and `runPlanFirstExecution`.
+- [x] "planFirst resume (B4)" route tests in `app.test.ts` (2) pass:
+      resume happy path and missing-snapshot error shape.
+
+---
+
+## M8 — Data governance & pipeline contract
+
+**Sources:** [`docs/DATA-GOVERNANCE.md`](./DATA-GOVERNANCE.md) and
+[`docs/GOVERNANCE.md`](./GOVERNANCE.md)
+
+**Goal:** Formalise consent, data modes, redaction, retention, deletion and
+export, and the binding three-repo training pipeline contract.
+
+### Acceptance criteria
+
+- [x] Three explicit data modes (Demo / Evidence / Training Data) exist;
+      Demo is the default and no data leaves it.
+- [x] Training Data Mode requires affirmations at both operator and session
+      levels; never inferred from usage.
+- [x] Redaction (secret scanner, path normalisation, 4 kB stdout
+      truncation) is applied in the worker before JSONL is written to KV,
+      not as post-processing.
+- [x] Retention table honoured: Demo 1h TTL, Evidence 30 days
+      (`EVIDENCE_RETENTION_DAYS` configurable), Training until explicit
+      delete/purge.
+- [x] `DELETE /jobs/:id/rollout` (single record) and `DELETE /rollouts`
+      (session) endpoints hard-delete records.
+- [x] `GET /rollouts/export` streams `rollout-wire/v1` JSONL with an
+      `X-Record-Count` header.
+- [x] `unknown` objective-status records are excluded from DPO pairs
+      (only permitted in weak-label pool or logs).
+- [x] `fixtures/data-loop/rollout-branches.v1.jsonl` is byte-identical to
+      the wasmagent-js and evomerge copies.
+
+---
+
+## See also
+
+| Milestone | Source doc(s) | Implementations / tests |
+|---|---|---|
+| M1 | [`B1-job-queue.md`](./B1-job-queue.md) | `apps/worker/src/jobs/*`, `apps/web/src/components/JobsPanel.tsx` |
+| M2 | [`B2-validation-loop.md`](./B2-validation-loop.md) | `apps/worker/src/build-results.ts`, `apps/worker/src/tools/build-result.ts` |
+| M3 | [`B3-github-import.md`](./B3-github-import.md) | `apps/worker/src/tools/githubImport.ts`, `importDenyList.ts` |
+| M4 | [`B4-approval-policy.md`](./B4-approval-policy.md) | `apps/worker/src/policies/approvalPolicy.ts` |
+| M5 | [`C1-sse-resume.md`](./C1-sse-resume.md) | `EventLog`, `packages/react/src/useAgentRun.ts` |
+| M6 | [`rlaif-adapter.md`](./rlaif-adapter.md) | `apps/worker/src/rollout-adapter.ts`, `trajectoryExport.ts` |
+| M7 | [`multi-agent-modes.md`](./multi-agent-modes.md) | `apps/worker/src/agents/multi-agent.ts` |
+| M8 | [`DATA-GOVERNANCE.md`](./DATA-GOVERNANCE.md), [`GOVERNANCE.md`](./GOVERNANCE.md) | Export/delete endpoints, redaction pipeline |
+
+## Status legend
+
+All criteria above are marked `[x]` because they are derived from
+planning documents that describe already-landed capability in this
+worktree (verified by the source docs' own "See also" implementations and
+passing test counts). This file converts those planning docs into the
+standard milestone format so the roadmap pipeline can operate
+autonomously — it does not introduce new scope.