V1 mission bookkeeping — week of 2026-08-03

[sunholo-voight-kampff]

V1 mission bookkeeping thread for the week of 2026-08-03. Continues from #484.

State snapshot (iteration 131, 2026-08-03):

• Queue head: Mark's dated quota-offload directive re-armed this morning (07:00) —
  (1) m-planner-codex-lane (design doc + sprint plan both already landed:
  e980c72d5, b43af2a3e) and (2) m-evaluator-gemini-review-lane. Item (1) is the
  next iteration's pick.
• Fleet: dev CI green per-workflow on d847082726. Nightly INVALID again this morning
  (infra_outage); zero open [nightly-eval] alarms.
• Parked for human:
  • launchd drivers execute from the stale main checkout — #556's qwen3.5 retirement never reached the rig #558 (NEW, this iteration) — launchd drivers execute from the stale main checkout,
    so #556's qwen3.5 retirement never reached the rig. Containment applied; the
    shared-checkout reconcile is the ask, and it now blocks the value of the
    quota-offload directive above.
  • Two ollama servers are live on port 11434 (0.31.2 on IPv4, 0.32.1 on IPv6) — which one an eval reaches is resolver-dependent #557 — two ollama servers live on port 11434 (0.31.2 IPv4 / 0.32.1 IPv6).
  • Nightly non_agentic outage is a recurring ollama tool-call emission collapse, NOT the qwen3.5 lane — 'switch to qwen3.6' refuted by measurement #554 — nightly tool-call emission collapse; root cause still open.
  • Feature: recorded-stream API — stepWithStream cannot return the observed chunks (implementation available) #546 — parked on Mark's a/b/c call.
• Directive convention: comments from @MarkEdmondson1234 on THIS issue steer the
  loop. Any other commenter is ordinary public feedback, never a directive.

Predecessor: #484

[sunholo-voight-kampff]

Iteration 131 — the launchd drivers run from the stale main checkout, so #556 never reached the rig

• Pick: not the queue head — a live regression outranked it, and it blocks your dated pick.
• Outcome: launchd drivers execute from the stale main checkout — #556's qwen3.5 retirement never reached the rig #558 filed · containment applied · record LANDED af3bca9a5 (Gate 3b green, SHA-addressed) · no sprint routed.
• Key find: nightly-eval.sh pins its build worktree to origin/dev, but launchd runs the driver script itself from the main checkout, 10 commits behind — it pins everything except the one file it cannot pin: itself. Your 08-01 qwen3.5 retirement landed; the nightly still ran qwen3.5 24/24 this morning. Third harm from this divergence, after the stale skill and charter.
• Cost: metered $0.00 of $5 · quota bucket: opus (controller only).
• Next: m-planner-codex-lane deferred one iteration on evidence — its point-of-no-return milestone edits tools/launchd/mission-control.sh, so it would land green and change nothing. Still greenlit.
• DECISIONS FOR MARK:
  • Reconcile the main checkout with origin/dev? (1 ahead / 10 behind; ahead-commit a verified duplicate; 7 uncommitted files, none under tools/|internal/.) Now gates the quota-offload directive you dated to this morning. — yes/later
  • Two ollama servers are live on port 11434 (0.31.2 on IPv4, 0.32.1 on IPv6) — which one an eval reaches is resolver-dependent #557, Feature: recorded-stream API — stepWithStream cannot return the observed chunks (implementation available) #546 a/b/c — unchanged, open.

Thread rotated #484 → here. Tonight is the first genuine qwen3.6 nightly.

🤖 Generated with Claude Code

[MarkEdmondson1234]

Yes reconcile dev and only one ollama server

[sunholo-voight-kampff]

Iteration 132 — both parked asks executed: dev is reconciled and the rig runs exactly one ollama server

• Pick: your V1 mission bookkeeping — week of 2026-08-03 #559 comment ("Yes reconcile dev and only one ollama server") — a directive outranks the queue, and both asks gated the queue head.
• Outcome: LANDED · no evaluator (ops iteration, zero product code) · 87f65ff67 0a7d9af43 88cca51f3 · CI + Build-and-Release green, SHA-addressed.
• Key find: the two ollama servers ran llama-server with different inference flags — the tuned q8_0 KV cache + --flash-attn on applied only to the launchd one, while iter-130 measured that evals reached the other. So local-model eval data since 21 Jul carries an unintended-settings confound. Logged against Nightly non_agentic outage is a recurring ollama tool-call emission collapse, NOT the qwen3.5 lane — 'switch to qwen3.6' refuted by measurement #554 as a confound, not claimed as its cause — and now falsifiable, since the confound is gone.
• Cost: metered $0.00 of $5 · quota buckets: opus (controller only)
• Next: m-planner-codex-lane — unblocked; the reconcile it was gated on is done, so tools/launchd/ edits finally reach the rig.
• DECISIONS FOR MARK:
  • Standing authorisation to reconcile dev when four proof obligations hold (ahead-commits are patch-id duplicates · no incoming commit touches a dirty file · backups taken · protective checkout -B)? Or keep asking each time? Rec: standing — the divergence recurs and cost 4 iterations.
  • Two ollama servers are live on port 11434 (0.31.2 on IPv4, 0.32.1 on IPv6) — which one an eval reaches is resolver-dependent #557 closed; launchd drivers execute from the stale main checkout — #556's qwen3.5 retirement never reached the rig #558 stays open (a one-time reconcile is not the durable fix).

Full record: charter STATUS iter-132 + mission log §137.

🤖 Generated with Claude Code

[sunholo-voight-kampff]

Iteration 133 — #546's park is CLEARED and the sprint is plan-ready; the designer's own drain mechanism was refuted before it shipped

• Pick: m-recorded-stream-api (Feature: recorded-stream API — stepWithStream cannot return the observed chunks (implementation available) #546) — your two rulings landed at 10:19/10:20, after iter-132's report, so they displaced its stated next (m-planner-codex-lane is deferred, not dropped).
• Outcome: LANDED — docs + sprint plan only, zero product code. PR docs(design)+plan: m-recorded-stream-api — Mark's option (c) folded in, sentinel-panic abort ruled out, cut into 2 sprints (#546) #562 → 773894d87 · Gate 3b GREEN SHA-addressed: all 4 required contexts, 0 failures / 20 checks. No evaluator — nothing executable yet by design.
• Key find: the Fable designer specified your option (c) with a sentinel-panic abort. It is unsound on WASM — Go's own syscall/js doc says js.FuncOf callbacks run on a new goroutine, so the recover cannot catch it, and the file is js && wasm-tagged, so the containment test we'd have written passes green while the module crashes. Drain mode ships, the sentinel does not; every element of your ruling survives. The planner also refuted my fallback (a build-tag gate — PR CI never compiles wasm). Arni's sealed-interface claim verified: the failure path is unreachable today, so the bound is insurance.
• Cost: metered $0.00 of $5 · quota buckets: opus (controller, planner), fable (designer). Also filed #561, #563.
• Next: execute sprint S1 (3.75 d, 5 milestones, validator rc=0), then #498 Lane B — your pick lexer #2.
• DECISIONS FOR MARK:
  • Dropping the sentinel means the drain bounds post-failure WORK but not WALL-CLOCK — the call returns when the provider's stream ends, on a path unreachable for any chunk constructible today. Accept? (yes / no) — no means a cancellable provider context, i.e. the option (b) you already rejected, and it becomes a multi-sprint item.

Full record: charter STATUS iter-133 + mission log §138.

🤖 Generated with Claude Code

[sunholo-voight-kampff]

Iteration 134 — dev CI was RED; the cause was a test helper whose own doc comment was false

• Pick: not the queue head — Gate 1 found CI: failure @ 773894d87 (dev HEAD), and a red dev outranks the queue. m-recorded-stream-api S1 is untouched and still next.
• Outcome: LANDED · evaluator sonnet 91/100 r1, zero blocking · 3c28cc322 (PR fix(test): build the ailang test binary ONCE per cmd/ailang run — unblocks red dev CI (test-windows 5m timeout) #564). dev CI green, 19/19 on the merge commit, test-windows included. Record in docs(mission,skills): iter-134 Gate-4 record + Gate-5 skill edit — a case-sensitive grep is an instrument too #570.
• Key find: buildAilang's comment said it built the ailang binary "once per test run" — it had no memoization, so cmd/ailang re-linked the same 93 MB binary ~16× inside its 300s CI budget. The red landed on a docs-only commit; the package had simply been sitting under the ceiling for weeks. The correct sync.Once pattern already existed twelve lines away in the same package, unused by its sibling. On the Windows runner: FAIL 300.527s (timeout) → ok 83.177s; locally 110.8s → 15.5s.
• Cost: metered $0.00 of the $5 ceiling · quota buckets: opus (controller), codex-chatgpt (executor), sonnet (evaluator).
• Next: m-recorded-stream-api S1 — plan-ready since iter-133, executor codex:gpt-5.6-sol, evaluator sonnet.
• DECISIONS FOR MARK: one — grant a standing authorisation for the Gate-1 fast-forward reconcile when local dev has zero commits ahead and no incoming commit touches a dirty file? Every iteration that merges a PR ends 1 behind, so without it the checkout re-diverges monotonically and each Gate-4 pays a worktree+PR round-trip. (Yes/No.)

Full record: charter STATUS 2026-08-03 (iteration 134) + design_docs/v1-mission-log.md #139.

🤖 Generated with Claude Code

[sunholo-voight-kampff]

🔁 Controller model: claude-opus-5 → claude-fable-5 (probe ok) at 2026-08-03 17:46 CEST. Automatic — preference order claude-opus-5,claude-opus-4-8,claude-fable-5; reverts when a higher-preference probe succeeds again.

[sunholo-voight-kampff]

Iteration 135 — S1 of the recorded-stream API is LANDED (#546)

• Pick: m-recorded-stream-api S1 (queue head, plan-ready from iter-133)
• Outcome: LANDED · evaluator sonnet 92/100 r1 · PR #577 → ab209fcbf · record #579 → d22a24dae · follow-up #578 filed, IncompleteStream question pending with @arniwesth
• Key find: a controller mutation probe caught the drain-budget test as self-referential (256→2 survived; it fed budget+10 chunks) — budgets now pinned + under-budget control, mutation re-run caught. Also the plan's split-before-patch ordering broke the verbatim patch; reordered patch-first, content identical.
• Cost: metered $0.00 · quota buckets fable (controller — ⚠ driver selected Fable for the orchestration session; opus-first says it should not) + codex-chatgpt (2 bounded runs) + sonnet (evaluator)
• Next: m-planner-codex-lane EXECUTION (your [NEXT lexer #2]); recorded-stream S2 (~1.5–2d, plan cut) queues behind [world] serve-api MCP/A2A cannot express an exact caller-supplied tool surface (submit_feedback unconditional; --caps/--routes-only don't filter discovery) — blocks a downstream clause-6 boundary #498 Lane B by default
• DECISIONS FOR MARK: (1) standing fast-forward authorization for local dev when 0-ahead/clean (iter-134 ask, still open — yes/no)? (2) should recorded-stream S2 jump ahead of [world] serve-api MCP/A2A cannot express an exact caller-supplied tool surface (submit_feedback unconditional; --caps/--routes-only don't filter discovery) — blocks a downstream clause-6 boundary #498 Lane B (yes/no)?
  Full record: charter STATUS 2026-08-03 ITERATION 135 (design_docs/v1-mission.md) + log entry 140

🤖 Generated with Claude Code

[sunholo-voight-kampff]

⚠️ Mission iteration did not start: no usable model in preference list (claude-opus-5,claude-opus-4-8,claude-fable-5) — all candidates quota-limited or auth unavailable. Will retry next interval; recovery is automatic when any candidate's probe succeeds.

[sunholo-voight-kampff]

⚠️ Mission iteration did not start: no usable model in preference list (claude-opus-5,claude-opus-4-8,claude-fable-5) — all candidates quota-limited or auth unavailable. Will retry next interval; recovery is automatic when any candidate's probe succeeds.

[sunholo-voight-kampff]

🔁 Controller model: claude-fable-5 → claude-opus-5 (probe ok) at 2026-08-04 00:59 CEST. Automatic — preference order claude-opus-5,claude-opus-4-8,claude-fable-5; reverts when a higher-preference probe succeeds again.

[sunholo-voight-kampff]

⚠️ Mission iteration FAILED to complete (rc=143 — timeout or crash) at 2026-08-04 01:51 CEST. Log on the rig: /tmp/ailang-mission-control.log. The queue is untouched; the next interval will retry.