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

[rickylabs]

Summary

Control PR for the 0.0.6 runtime / public-surface lane. It carries orchestration evidence only —
the run dir under .llm/runs/release-0.0.6-features--orchestration/ and nothing else. It is not
an implementation umbrella: both owned issues land through their own fresh leaf branches with draft
PRs opened directly against main.

Scope

• Archetype / area: harness run artifacts (.llm/runs/**)
• Owns fix(streams): durable producer settles two inaccurate rejection reasons — close-drain reports producer-failed, single-attempt refusal reports retry-exhausted #1405 and fix(triggers/streams): job executions are never published to the durable job stream, so no execution trace can span producer → stream → consumer #1398 — referenced without closing keywords; the implementation PRs carry those.
• Sibling lanes run in parallel checkouts (docs / fixes / internals) and are not touched here.
• PR docs(harness): land the DevTools RFC run's closing artifacts #1522 is unrelated and has been removed from the milestone.

Two-PR dependency plan

Order	Issue	PR	Shape	Depends on
1	#1405	fix/1405-durable-producer-rejection-taxonomy	small, specified, deterministic; reasons-only	nothing — lands first
2	#1398	fix/1398-publish-job-executions-to-durable-stream	substantive public/runtime behaviour + live E2E	soft-depends on #1405

Why this order, and what the dependency actually is. #1398 publishes job executions through
DurableStreamProducer and must prove a live subscription observes the record within a bounded
time. When that proof fails, the settled write reason is the first diagnostic anyone reads — and
today two of those reasons misdescribe the state that produced them (#1405). Landing #1405 first
means #1398's failure diagnostics are trustworthy while its runtime evidence is being gathered. The
dependency is diagnostic, not structural: #1398 does not import anything #1405 adds, so if #1405
were blocked, #1398 could still proceed with the drift recorded.

#1398 builds on merged #1395 (versioned SSE/OTEL envelope, merged 2026-08-09T01:25:15Z) and #1402
(durable producer reconnect supervisor, merged 2026-08-09T05:11:32Z) — both ancestors of this
branch's baseline — but is not satisfied by them. It stays its own reviewable PR.

Slices

• S1 Bootstrap: run dir, supervisor identity, lane bindings, drift log, cut-trace baseline — a4964f9af
• S1 Research fix(streams): durable producer settles two inaccurate rejection reasons — close-drain reports producer-failed, single-attempt refusal reports retry-exhausted #1405 (in-session; small and fully specified) — a4964f9af
• S2 Research fix(triggers/streams): job executions are never published to the durable job stream, so no execution trace can span producer → stream → consumer #1398 → plan → separate PLAN-EVAL (MiniMax M3, PASS; findings F1/F2 folded in)
• S3 fix(streams): durable producer settles two inaccurate rejection reasons — close-drain reports producer-failed, single-attempt refusal reports retry-exhausted #1405 implementation slice + IMPL-EVAL (PASS) + merged 8ff1bcb8f
• S4 fix(triggers/streams): job executions are never published to the durable job stream, so no execution trace can span producer → stream → consumer #1398 implementation slice + live runtime/OTEL evidence + IMPL-EVAL (Qwen 3.8 Max via the automatic dispatcher, PASS) + merged d7e2b67b2
• S5 Lane closeout: cut-trace.md complete (2 merges, 2 re-planning events, 8 time-costing failures), retrospective.md written, context-pack.md written

Validation

Bootstrap commit touches .llm/runs/** only — no publishable source, so no package gate applies to
this PR. Gate evidence for the owned issues lands on their own PRs, not here.

Environment proof recorded in worklog.md:

• git fetch origin main → origin/main@01aa12b67; lane opened at tip, clean tree.
• deno task agentic:runtime doctor → no_change (schema 1.0), components 18, sessions 0.

Harness

• Run dir: .llm/runs/release-0.0.6-features--orchestration/
• Profile: .llm/harness/workflow/milestone-run.md (topical lane), role skill
  agent-milestone-orchestrator.
• Phase: RESEARCH → PLAN.
• PLAN-EVAL: N/A for fix(streams): durable producer settles two inaccurate rejection reasons — close-drain reports producer-failed, single-attempt refusal reports retry-exhausted #1405 (complete contract, deterministic, reasons-only — per the
  lane-policy owner decision of 2026-08-08). Required for fix(triggers/streams): job executions are never published to the durable job stream, so no execution trace can span producer → stream → consumer #1398, separate session, before
  implementation. No separate PLAN-EVAL for the two-PR sequencing plan itself; that decision is
  recorded in supervisor.md rather than assumed.
• Generator and evaluator never share a session; merge authority sits with the orchestrator and runs
  the milestone-run.md pre-merge gate per PR.

Drift / Debt

• D-1 (minor) — research/plan sub-agents run on Claude Opus medium/high rather than the
  canonical deep_analysis Fable 5 medium, on explicit owner instruction. Opposite-family review of
  Codex work is unaffected: Fable 5 low remains the review_codex reviewer for the fix(triggers/streams): job executions are never published to the durable job stream, so no execution trace can span producer → stream → consumer #1398 slice.
• D-2 (minor) — the brief routes evaluation through OpenHands after feat(agentic): automate formal evaluator phases #1524 passes/lands.
  Observed 2026-08-12: PR feat(agentic): automate formal evaluator phases #1524 is OPEN, mergedAt: null, with its bounded-live-smoke and
  repository-default-variable DoD boxes unticked. Per the brief's own condition this lane uses fresh
  local evaluator sessions; re-checked before each dispatch.

Definition of Done

• fix(streams): durable producer settles two inaccurate rejection reasons — close-drain reports producer-failed, single-attempt refusal reports retry-exhausted #1405 merged to main (8ff1bcb8f) through the seven-check pre-merge gate; all 5 acceptance boxes ticked by the evidence mirror, not by hand — slices/pre-merge-gate-1528.md
• fix(triggers/streams): job executions are never published to the durable job stream, so no execution trace can span producer → stream → consumer #1398 merged to main (d7e2b67b2) through the seven-check pre-merge gate; both formerly-deferred OTEL gates pass by name on both CI tiers at the merging head (postgres 94073971396 passed=88 failed=0 skipped=0; sqlite 94073971501 passed=83 failed=0 skipped=0); all 4 acceptance boxes mirrored — slices/pre-merge-gate-1536.md
• cut-trace.md records both merges captured from git log origin/main --first-parent after each merge, plus 2 re-planning events and 8 time-costing failures with mitigations
• retrospective.md written: what the lane produced, where each review layer caught what the layer above missed, four mistakes this orchestrator made, and four candidate rules explicitly not promoted from a single occurrence

[rickylabs]

[PHASE: REVIEW] — lane closeout

Both owned issues are on main. This control PR carries orchestration evidence only and closes nothing.

Issue	PR	Merge	Closed
#1405 durable producer rejection taxonomy	#1528	8ff1bcb8f	auto, COMPLETED
#1398 job executions → durable job stream	#1536	d7e2b67b2	auto, COMPLETED

Both merged through the seven-check pre-merge gate (slices/pre-merge-gate-1528.md, slices/pre-merge-gate-1536.md), each with review-threads clean.

#1398's acceptance was made mechanical rather than asserted. Its "live subscription observes the record within a bounded time" criterion was bound to two gates the repo had already deferred against this very issue. Done meant removing the deferral and having the gates pass — which they now do, by name, on both CI runtime tiers (failed=0 skipped=0).

Filed from inside the run: #1542 (quality:gate roots omit published packages — both merge records in this lane explicitly rest on explicit target scans rather than the repo gate) and #1543 (undeclared plugin-streams-core imports, filed as unverified with publish:dry-run evidence as its first acceptance box).

Retrospective: .llm/runs/release-0.0.6-features--orchestration/retrospective.md. It records eight time-costing failures with mitigations, four mistakes this orchestrator made, and four candidate rules explicitly not promoted — each appeared once and is left for a future run to confirm or falsify.

Canary and stable cut remain root's; nothing here was published.