chore(harness): orchestrate the 0.0.6 runtime and public-surface lane

.llm/runs/release-0.0.6-features--orchestration/context-pack.md

@@ -0,0 +1,76 @@
+# Context pack — 0.0.6 runtime / public-surface lane
+
+Closing summary. The lane is **complete**: both owned issues landed on `main`. Read this first if
+resuming or auditing; everything below is traceable to a named artifact.
+
+| Field | Value |
+| --- | --- |
+| Run id | `release-0.0.6-features--orchestration` |
+| Profile | `.llm/harness/workflow/milestone-run.md` (topical lane) |
+| Supervisor | Claude · Opus 5 · high |
+| Control branch / PR | `chore/release-0.0.6-features-orchestration` / **#1525** (evidence only, closes nothing) |
+| Baseline | `origin/main@01aa12b67` |
+| Status | **Complete** — 2/2 issues merged, 2 follow-ups filed |
+
+## Outcome
+
+| Issue | PR | Merge | Gate record |
+| --- | --- | --- | --- |
+| #1405 durable producer rejection taxonomy | #1528 | `8ff1bcb8f` | `slices/pre-merge-gate-1528.md` |
+| #1398 job executions → durable job stream | #1536 | `d7e2b67b2` | `slices/pre-merge-gate-1536.md` |
+
+Both auto-closed `COMPLETED` via closing keywords; `status:shipped` on each issue and PR.
+
+## What changed, in one paragraph each
+
+**#1405** — two settled write **reason strings** misdescribed the state that produced them: a write
+rejected during the graceful close-drain reported `producer-failed` (the producer was healthy and
+closing), and a non-retryable append on attempt 1 reported `retry-exhausted` (nothing was exhausted).
+Fixed by making the closing intent observable to the rejection selector and by branching `#failActive`
+on `isRetryable`, adding exactly one public member, `transport-refused`. The façade's duplicate
+selector was deleted so the two cannot drift apart again — that drift was the defect's origin. No
+change to which writes are accepted, rejected, or delivered.
+
+**#1398** — job executions were never published to the durable job stream because the workers **API
+service** installed the execution-state mutation hook while the **background** entrypoints that
+generated projects actually run never did. Fixed by installing the hook on the worker and combined
+runtimes and publishing inside `context.with(extractContext({traceparent, tracestate}), …)` so the
+publish span joins the `job.execute` trace — including the pre-span `create()` record, which is the
+trap that would otherwise fail TC-14 silently. The two E2E gates deferred against this issue were
+un-deferred and now pass live.
+
+## Evidence chain
+
+- **#1398 acceptance was made mechanical.** Its observational criterion was bound to two gates the
+  repo had already deferred against this issue. Both now pass **by name** on both CI runtime tiers at
+  the merging head: postgres `94073971396` (`passed=88 failed=0 skipped=0`), sqlite `94073971501`
+  (`passed=83 failed=0 skipped=0`).
+- **PLAN-EVAL** #1398: PASS, MiniMax M3, separate session — `plan-eval.md`. Findings F1/F2 folded in.
+- **IMPL-EVAL** #1405: PASS, DeepSeek V4 Flash 0731 — `slices/evaluate-1405.md`. Superseded as
+  *policy* by D-3 after the fact, not retracted.
+- **IMPL-EVAL** #1398: PASS, Qwen 3.8 Max via the **automatic dispatcher** —
+  `slices/evaluate-1398.md`.
+- Merge order and every time-costing failure: `cut-trace.md`. Lessons and mistakes:
+  `retrospective.md`.
+
+## Open items leaving this lane
+
+| Item | Where |
+| --- | --- |
+| `quality:gate` roots omit published packages — **both merge records here rest on explicit target scans, not the repo gate** | **#1542** (0.0.7) |
+| Undeclared `@netscript/plugin-streams-core` imports — filed **unverified**, `publish:dry-run` evidence is its first acceptance box | **#1543** (0.0.7) |
+| Canary and stable cut | **root's**, not this lane — nothing was published here |
+
+## Drift index
+
+D-1 research sub-agent lane override · D-2 evaluator transport fallback · **D-3** IMPL-EVAL waived
+for the small deterministic class · **D-4** phase evaluation moved to the automatic dispatcher ·
+**D-5** label-driven eval trigger contract. Full text in `drift.md`.
+
+## Artifact hygiene note
+
+The two raw evaluator JSONL streams (2.4 MB combined) were **untracked** and moved to
+`.llm/tmp/` scratch, which `.gitignore` excludes. They were 2.4 MB of a 2.5 MB run dir, against a
+96 K largest-artifact precedent in the 0.0.5 run. Their substance is preserved verbatim in
+`plan-eval.md` and `slices/evaluate-1405.md`, each carrying run id, duration, event count, and
+`is_error`. Nothing was deleted — the files remain on disk locally.

.llm/runs/release-0.0.6-features--orchestration/cut-trace.md

@@ -0,0 +1,47 @@
+# Cut trace — 0.0.6 runtime / public-surface lane
+
+Merge-derived record. Every row is captured **from `git log origin/main` after the merge**, never
+from recollection or from the dispatch plan. This lane does not cut or publish; root orchestration
+owns the canary and the stable cut.
+
+## Baseline
+
+| Field | Value |
+| --- | --- |
+| Lane opened | 2026-08-12 |
+| `origin/main` at open | `01aa12b67` — `docs(harness): record FILING-LOG -- board migration executed once (#1523)` |
+| Owned issues | #1405, #1398 |
+
+## Merges
+
+| # | UTC | Merge commit on `origin/main` | PR | Issues closed | Pre-merge gate record |
+| --- | --- | --- | --- | --- | --- |
+| 1 | 2026-08-12T08:20:29Z | `8ff1bcb8fc741db17a765a1965861828e0ae6171` (`8ff1bcb8f`) — `fix(streams): distinguish producer refusal reasons (#1528)` | #1528 | **#1405** (auto-closed `COMPLETED` by the body's `Closes #1405`) | `slices/pre-merge-gate-1528.md` — all 7 checks PASS |
+
+| 2 | 2026-08-12T10:27:26Z | `d7e2b67b2be535c9ca13449f97f8f4585344030a` (`d7e2b67b2`) — `fix(workers): publish job executions to the durable stream on the job.execute trace (#1536)` | #1536 | **#1398** (auto-closed `COMPLETED` by the body's `Closes #1398`) | `slices/pre-merge-gate-1536.md` — all 7 checks PASS at head `f7d503fee` |
+
+Both rows captured from `git log origin/main --first-parent -1` **after** each merge, per the
+profile's merge-derived rule. Issue states re-read live: both `CLOSED` / `COMPLETED`, with
+`status:shipped` applied to each issue and PR.
+
+**Lane complete: both owned issues landed on `main`.**
+
+## Re-planning events
+
+| # | Date | Event | Decision |
+| --- | --- | --- | --- |
+| 1 | 2026-08-12 | #1536's head changed mid-flight, `e4319c685` → `f7d503fee` | The phase-eval dispatcher (#1524) merged **after** the branch's last `main` sync, so `openhands-phase-eval.yml` was absent from the PR's merge ref and no label cycling could trigger it. Owner approved syncing the branch, accepting a full CI re-run. All gate evidence was then re-read against the new head; no pre-sync evidence was carried forward. |
+| 2 | 2026-08-12 | Evaluation route changed twice mid-run | D-3 removed formal IMPL-EVAL for the small deterministic class (after #1405 had already merged); D-4/D-5 moved phase evaluation to the automatic label-driven dispatcher. #1398's IMPL-EVAL consequently ran on the automatic route rather than a manual launch. |
+
+## Failure modes that cost real time
+
+| # | Date | Failure | Cost | Mitigation recorded |
+| --- | --- | --- | --- | --- |
+| 1 | 2026-08-12 | **Dispatcher absent from the PR's merge ref.** #1524 merged 38 min after #1536's branch last synced with `main`, so `openhands-phase-eval.yml` did not exist for that PR. Labels were cycled correctly and produced **no run at all**; the only runs were the older `openhands-agent.yml`, all `skipped`. | One wasted label re-entry, plus a full CI re-run (~20 min, both `scaffold-runtime` tiers) after the branch sync. | For `pull_request` events GitHub resolves workflows from the **merge ref**. A newly merged workflow cannot fire on a PR whose head predates it. **Check the workflow exists in the PR head before concluding a trigger failed.** Diagnosed by `git cat-file -e <head>:<workflow>` → ABSENT, with the post-sync run succeeding as the control. |
+| 2 | 2026-08-12 | **Draft PRs report every check as `skipping`.** #1528 looked "clean" while nothing substantive had run. | None — caught by pre-merge check 4 before merge. | This is the #778/#775 class, alive and current. Absence of red is not green. The blocking tier only runs after draft→ready. |
+| 3 | 2026-08-12 | **Two live `scaffold.runtime` runs died before reaching the gates under test** — run 1 on a transient `generate plugins: fetch failed`, run 2 on a `triggers-api` health timeout at 120 s. | ~25 min of local runtime, no verdict produced. | Local WSL was not a usable arbiter for this suite. CI ran the same suite with the same change to a clean finish on both tiers, which is the control that established the local failures as environmental. **Do not label a local red a "flake" without that control.** |
+| 4 | 2026-08-12 | **A counting watcher would never have fired.** OpenHands updates its summary comment **in place** (`openhands-agent-summary` marker, `"conclusion"` field), so a watcher keyed on comment count polls to timeout while the verdict sits in an edited comment. | None — caught before arming, on inspecting the posted comment's markers. | Watch the **run status** and the comment's `conclusion` marker, never comment count. |
+| 5 | 2026-08-12 | **Stale check summary read as current.** `gh pr checks` reported `close-gate` red from a job that ran 40 min earlier, before the label and body changes it was complaining about. | None — caught by reading the job log, which carried the mirror's own "skipped because labels do not include status:ready-merge" notice. | Compare a check's run time to the change it is judging. The gate-integrity rule already says merge-history audits must take the latest run per check name; the same applies pre-merge. |
+| 6 | 2026-08-12 | **Automation added a second `status:` label.** `status:augment-review` was applied 1 s after the orchestrator moved off `status:impl-eval`, breaching the exactly-one-status invariant. | None — caught while verifying labels before merge. | Re-verify the `status:` set after any automated phase transition, not only after manual edits. |
+| 7 | 2026-08-12 | **Orchestrator's own slice brief named a broken gate command.** `deno test packages/plugin-streams-core` exits 1 with 19 `NotCapable` errors for want of `--allow-env`. | Minor; the implementer reported the red with its cause rather than hiding or working around it. | Use the package-declared `deno task --cwd <pkg> test`. Corrected in the #1398 brief rather than repeated. |
+| 8 | 2026-08-12 | **One unnecessary evaluator dispatch.** The lane brief's IMPL-EVAL waiver for the #1405 class was read as a blocked-transport fallback rather than the class default. | One DeepSeek IMPL-EVAL run (~643 s) that the owner did not want. | Recorded as D-3. "A waiver is available" and "the waiver is the default" are different instructions; resolve the ambiguity before spending. |

.llm/runs/release-0.0.6-features--orchestration/drift.md

@@ -0,0 +1,124 @@
+# Drift — 0.0.6 runtime / public-surface lane
+
+Append-only. Severity: `minor` | `significant` | `architectural`.
+
+## D-1 — research/plan sub-agent lane overridden to Opus (minor)
+
+**Date** 2026-08-12. **Canonical route** `deep_analysis` = Claude · Fable 5 · medium. **Used**
+Claude · Opus 5 medium/high. **Reason** the owner brief for this lane explicitly directs
+"Delegate plan/research to Claude Opus medium/high sub-agents when useful". Owner instruction is
+more specific than the default lane binding. Invariants preserved: generator ≠ evaluator, no lane
+self-certifies, no paid escalation. Fable 5 · low remains the `review_codex` reviewer for the
+#1398 slice, so opposite-family review of Codex work is untouched.
+
+## D-2 — evaluator transport falls back to local fresh sessions (minor)
+
+**Date** 2026-08-12. The brief routes PLAN-EVAL/IMPL-EVAL through OpenHands *after #1524
+passes/lands*. Observed: PR #1524 is **OPEN**, `mergedAt: null`, with unticked DoD boxes for the
+bounded live DeepSeek smoke and the repository default variable. Per the brief's own condition, this
+run uses fresh local Claude/OpenCode OpenRouter evaluator sessions through the toolchain. Re-checked
+before each dispatch; state recorded at the point of use.
+
+## D-3 — IMPL-EVAL not required for small deterministic taxonomy fixes (significant)
+
+**Date** 2026-08-12. **Owner ruling.** For the #1405 class — small, deterministic, fully specified
+fixes with exact negative tests — a separate formal IMPL-EVAL is **not** to be dispatched. The
+sufficient evidence set is: focused negative tests, CI, close-gate, and the orchestrator's own
+independent diff review.
+
+**My error.** The lane brief said an IMPL-EVAL owner waiver was "acceptable" for #1405 with exact
+negative tests. I read that as a fallback to use only if the evaluator transport was blocked, and
+recorded that reading in `supervisor.md`. The owner's intent was that the waiver is the **default**
+for this class. One evaluator dispatch (DeepSeek V4 Flash 0731 max, 642,836 ms) was spent that should
+not have been.
+
+**Timing, stated plainly.** The ruling arrived after #1405 had already been evaluated **and merged**
+(`8ff1bcb8f`, 2026-08-12T08:20:29Z). It therefore does not retract that merge or its evidence — the
+IMPL-EVAL returned PASS and its per-fix revert isolation is real evidence that remains in the record.
+The ruling governs this class going forward.
+
+**Not weakened for #1398.** Formal PLAN-EVAL and IMPL-EVAL remain **mandatory** for #1398: it changes
+public runtime behaviour, publishes a new record to a durable stream, and requires live evidence.
+Its PLAN-EVAL has already run (PASS, MiniMax M3) and its IMPL-EVAL is still required before merge.
+
+**Cleanup.** The detached #1405 evaluator worktree `/home/codex/repos/ns006-1405-impleval` was
+verified clean (empty `git status --porcelain`) and removed. The #1398 PLAN-EVAL worktree
+`/home/codex/repos/ns006-1398-planeval` is retained pending that issue's IMPL-EVAL. Pre-existing
+global stashes belonging to other lanes' branches were left untouched.
+
+## D-4 — phase evaluation moves to the automatic status dispatcher (significant)
+
+**Date** 2026-08-12. **Owner ruling.** PR #1524 (automatic phase dispatcher) is about to merge. Once
+it lands, **all future phase evaluations use the automatic status workflow** unless the owner selects
+a documented local route or an explicit skip. Manual IMPL-EVAL launches and manual
+`@openhands-agent` PR comments are not to be used for #1536.
+
+**Standing instruction for #1536:** keep it on its **current head and status** — head `e4319c685`,
+`status:impl-eval`, milestone `0.0.6`. Root will **deliberately re-enter** `status:impl-eval` with
+the Qwen override *after* #1524 lands, which is what triggers the automatic dispatcher. This
+orchestrator must not re-enter the label for the same head, must not trigger OpenHands, and must not
+launch a local evaluator. Its remaining job is to **watch the automatic verdict and then finish the
+merge gate**.
+
+**Timing, recorded factually.** The steer anticipated that a local #1536 evaluator had already been
+launched. It had **not**. For #1398 I wrote the evaluator prompt
+(`slices/impl-eval-1398-prompt.md`, 10:57) and pre-created the detached worktree
+`/home/codex/repos/ns006-1398-impleval` at `e4319c685`, then **stopped and raised the decision**
+rather than dispatching. Verified at the time of this entry: no `openrouter-run`/`claude-openrouter`
+process, and no `impl-eval-1398-raw.md` output file — only the prompt. **No duplicate spend occurred
+and none was in flight.**
+
+The two evaluator sessions this run did spend are unaffected and both completed before this ruling:
+PLAN-EVAL #1398 (MiniMax M3, PASS) and IMPL-EVAL #1405 (DeepSeek V4 Flash 0731, PASS — itself the
+subject of D-3).
+
+**Unused artifacts retained, not removed:** `slices/impl-eval-1398-prompt.md` and the
+`ns006-1398-impleval` worktree are left in place, clean and unused, in case the owner later selects
+the documented local route. They are inert; nothing reads them.
+
+**Interaction with D-3.** D-3 waives formal evaluation for the small deterministic class. D-4 does
+not widen that waiver — #1398 still gets a formal IMPL-EVAL; it now arrives through the automatic
+dispatcher rather than a manual launch.
+
+## D-5 — automation trigger contract for formal evaluation (significant)
+
+**Date** 2026-08-12. **Owner policy.** Formal PLAN/IMPL evaluation is triggered by labels, never by a
+manual OpenHands dispatch:
+
+| Phase | Initial trigger | Rerun |
+| --- | --- | --- |
+| PLAN-EVAL | the `openhands` + `status:plan-eval` label **pair**, exactly once | move away from `status:plan-eval`, then re-add |
+| IMPL-EVAL | automatically on **draft → ready**, unless `impl-eval:skip` | move away from `status:impl-eval`, then re-add |
+
+`eval:model:minimax|deepseek|qwen` is an optional **one-shot** override. A local eval already running
+may finish, but must never be duplicated. Manual `@openhands-agent` dispatch for formal PLAN/IMPL
+eval is prohibited. Merge continues to go through normal harness authority.
+
+**This lane complies with no change required:** no manual OpenHands dispatch was made for #1536 (or
+for anything else this run), and no local evaluator was launched for #1536 — the prompt and worktree
+were prepared and the dispatch deliberately withheld (D-4). The two local evaluator sessions this run
+did spend — PLAN-EVAL #1398 (MiniMax M3) and IMPL-EVAL #1405 (DeepSeek) — both **completed** well
+before this policy and are not duplicated.
+
+### Timing finding: #1536's automatic IMPL-EVAL could not have fired
+
+Measured from the issue timeline rather than assumed:
+
+| Event | UTC |
+| --- | --- |
+| #1536 `ready_for_review` | **2026-08-12T08:53:43Z** |
+| #1536 `status:impl` removed, `status:impl-eval` applied | 2026-08-12T08:53:45Z |
+| #1524 (the dispatcher) merged | **2026-08-12T09:24:15Z** |
+
+Both candidate triggers precede the dispatcher's existence by ~30½ minutes. The draft → ready
+transition therefore had nothing to fire, and the initial automatic IMPL-EVAL for #1536 **did not
+run and will not run on its own**.
+
+Under this policy the rerun path is the only one left: move #1536 away from `status:impl-eval`, then
+re-add it. That matches what root already stated it would do with the Qwen override, so the
+conclusion is unchanged — but it is now a **requirement** rather than a preference, and a watcher
+waiting for a spontaneous verdict would wait forever. Recorded so that is visible rather than
+discovered by timeout.
+
+**Not actioned by this lane.** The label re-entry is root's, per D-4. This orchestrator has not
+touched #1536's head, labels, or body.