docs(autonomy): ignition Phase 1 status note (billing blocker, bound supersession)

[kyle-sexton]

Dated status note on the ignition plan: Phase 1 built and PR'd (autonomy-demo-scratch#5), two deviations recorded — --max-turns absent from the installed CLI (mechanical bound is --max-budget-usd + timeout wrapper), and the deterministic-gate job startup-fails on personal-account GitHub Actions billing (operator action required).

No linked issue (tracks #778 without closing it).

Related

• autonomy: ignite the 2-to-3 loop — first standing routine (hourly C2 drain) + C2-promotion evidence drain on the proving ground #778 (ignition), docs/topics/autonomy-ignition/PLAN.md (merged via docs(autonomy): commit autonomy-ignition sub-topic plan (stress-tested) #796)
• feat: ignition phase 1 — hourly drain, PR-flow dispatch, gated evidence join kyle-sexton/autonomy-demo-scratch#5 (the Phase 1 build PR this note tracks)

🤖 Generated with Claude Code

https://claude.ai/code/session_01RdeVMaDzwP26bNiDjqM59Q

[claude]

Claude finished @kyle-sexton's task in 52s —— View job

---

Review: docs(autonomy): ignition Phase 1 status note

• Read changed file and analyzed diff
• Checked internal consistency across PLAN.md

---

The status note itself is well-structured: it's dated, records two concrete deviations with verifiable evidence (claude --help), classifies the billing failure correctly as a BLOCKER, and calls out the two open smoke-test unknowns. The [TODO] → [DOING] tag advance and the operator-decisions-table update are correct. The sanity-check grep list at line 117–119 was updated to match.

One consistency issue: two stale --max-turns references were not updated.

Finding 1 — Work item #3 body still names --max-turns (stale)

Line 94:

`--model sonnet --max-turns <N>` + wrapper timeout kill (per-run mechanical bound)

The status note and operator-decisions table both record that --max-turns is absent from the installed CLI and --max-budget-usd is the actual flag. This work-item description is now internally inconsistent with those updates. Fix this →

Finding 2 — Stress-test summary names max-turns (stale)

Line 191:

Sonnet + max-turns on the inner invocation (kills the Opus hardcode leak + unbounded runs)

This is the stress-test reversal record; it now lags the operator-decisions update. Should read max-budget-usd (or max-budget-usd + timeout wrapper) to stay consistent. Fix this →

---

Both findings are doc-internal consistency issues, not blockers on their own, but since this PR's explicit purpose is to record the deviation accurately, leaving two instances of the old flag name feels like incomplete coverage. Everything else — the plan conventions around dated deviations, the billing-blocker call-out, the open smoke items — looks correct.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 485f107e41

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[claude]

Claude finished @kyle-sexton's task in 1m 12s —— View job

---

Review: docs(autonomy): ignition Phase 1 status note

• Read PR diff and changed file
• Verify stale --max-turns references from prior review were addressed
• Review full diff for correctness and internal consistency

---

The latest commit (2986bd5 — "align stale --max-turns refs with recorded deviation") closes both findings from the prior review:

Finding 1 resolved — PLAN.md:92: work item #3 now reads --max-budget-usd <USD> + wrapper timeout kill. No longer prescribes the non-existent --max-turns <N>.

Finding 2 resolved — PLAN.md:190: stress-test summary now reads max-budget-usd on the inner invocation. Consistent with the operator decisions table and status note.

Remaining --max-turns mentions in the file are in the deviation-documentation context ("does not exist in the installed CLI", "absent from the installed CLI, see status note") — those are intentional and correct; they describe the old flag for traceability, not prescribe its use.

Full diff review:

Change	Assessment
Operator decisions table — budget bound row	Correct; names both --max-budget-usd and the timeout 3300 wrapper, records verification date
[TODO] → [DOING] on Phase 1 heading	Correct tag advance
Status note block	Well-structured: dated, itemised deviations numbered, BLOCKER explicitly labelled, two open smoke unknowns named
Work item #3 body (--max-budget-usd <USD>)	Now matches the verified CLI surface
Sanity-check grep list	Updated to --max-budget-usd with explanatory parenthetical; complete and accurate
Stress-test summary	Updated to max-budget-usd on the inner invocation

No remaining consistency issues. The PR achieves its stated purpose and the Codex/prior-review findings are both resolved. Ready to merge once the operator-action BLOCKER (GitHub Actions billing on the personal account) is cleared and autonomy-demo-scratch#5's gate job goes green.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 2986bd5b0f

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[claude]

Claude finished @kyle-sexton's task in 1m 40s —— View job

---

Review: docs(autonomy): ignition Phase 1 status note

• Read PR context and prior review history
• Read current state of PLAN.md
• Assess open Codex P2 finding (Phase 3 smoke criteria)
• Post final review

---

The latest commit (e3c68dc) is a re-trigger with no content change — the PR diff is unchanged from the state reviewed in 2986bd5. Prior findings (work-item #3 flag name, stress-test summary) are still resolved. One open finding from the second Codex review remains.

Finding — Phase 3 smoke step still says "within the turn bound" (stale)

Line 156:

2. Verify the run executed in an isolated worktree (live checkout untouched — `git status`
   clean before/after), on Sonnet (session telemetry model check), within the turn bound.

The status note on lines 65–76 documents that --max-turns does not exist in the installed CLI, and the operator decisions table (line 31) records that the per-run mechanical bound is --max-budget-usd + timeout wrapper kill. "Turn bound" as a verifiable smoke target no longer maps to anything in the implementation — a verifier executing Phase 3 would have no concrete thing to check.

The fix is narrow: replace "within the turn bound" with the actual bound mechanism, e.g. within the budget/timeout bound (--max-budget-usd + timeout 3300) to match what line 31 and the status note record.

Fix this →

---

Everything else looks correct and consistent:

Change	Assessment
Status note block (lines 65–76)	Well-formed: dated, two numbered deviations, BLOCKER labelled, two open smoke unknowns named
Operator decisions table budget-bound row (line 31)	Correct — names both --max-budget-usd and the timeout 3300 wrapper, records verification date
[TODO] → [DOING] (line 63)	Correct tag advance
Work item #3 (--max-budget-usd <USD>, line 92)	Matches verified CLI surface
Sanity-check grep list (lines 117–119)	Updated to --max-budget-usd with explanatory parenthetical
Stress-test summary (line 190)	Updated to max-budget-usd on the inner invocation
Remaining --max-turns mentions (lines 31, 68, 117–118)	Intentional — they document the absent flag for traceability

The Phase 3 "turn bound" phrase is the only instance of the old terminology that wasn't caught in the prior two rounds. It's a minor consistency issue (Phase 3 is [TODO], so no one is executing it yet), but since this PR's explicit purpose is accurate deviation documentation, patching it keeps the plan internally coherent before merge.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e3c68dc9ed

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Verify max-turns with a runtime probe

Fresh evidence from the current official Claude Code CLI reference says claude --help is incomplete and documents --max-turns as the print-mode agentic-turn limiter, while --max-budget-usd only limits API spend. In environments where the installed binary supports the documented flag but omits it from help, this note tells Phase 1 to drop the deterministic turn cap and rely on a cost cap plus wall-clock timeout; please probe claude -p --max-turns 1 ... or pin the unsupported CLI version before superseding the turn limit.

Useful? React with 👍 / 👎.