Updated just now.
2026-10-01 — Finished on 1 October 2026 at 7:33 am, live on Nick's Mac Studio and Mac mini. Every Claude session on both Macs refuses a paid web search or page read through DeepAPI (the research service that bills per call) unless the agent says what the free search tool failed to do, and lets ordinary work through. OpenAI's Codex agent and every copy of the DeepAPI instructions carry the same free-tools-first rule, and the automatic update DeepAPI runs once a day cannot undo it. An hourly check compares DeepAPI's count of paid searches with the ones made by the content-ideas research job Nick approved on 30 September: an unexplained search fails the check, one with a stated reason shows as a warning. Nick approved the last correction to the written rule at 7:14 am and it is in place. Nothing is open; Chantelle's and the teammates' Macs apply the change on their own but cannot be checked from here.
2026-10-01 — Finished and signed off on 30 September 2026 at 10:21 pm, live on Nick's Mac Studio and Mac mini. Every Claude session on both Macs now refuses a paid web search or page read through DeepAPI (the research service that bills per call) unless the agent says what the free search tool failed to do, and lets ordinary work through. OpenAI's Codex agent and every copy of the DeepAPI instructions carry the same free-tools-first rule, and the automatic update DeepAPI runs once a day cannot undo it. An hourly check compares DeepAPI's count of paid searches with the ones made by the content-ideas research job Nick approved on 30 September: an unexplained search fails the check, one with a stated reason shows as a warning. Five brand-new agent sessions all used free search for an ordinary lookup. Still open: Nick's one-tap approval, already on his phone, to correct one out-of-date sentence in the rule about DeepAPI; and Chantelle's and the teammates' Macs, which apply the change on their own but cannot be checked from here.
Definition of done: selftest passes every case and fails against the allow-everything stub
Proof: the relevant item
Verified: 2026-09-30 (100%, checked by Astra round 13 — five routine commands allowed, three plain paid calls refused, no finding; 80 cases + 5 end-to-end pass, an allow-everything stub fails 51; every evasion and false refusal found by the Sonnet verifier and Astra rounds 1–12 is a case; live refusal seen in a real session at 10:18 pm)
Definition of done: --check all in place on the Studio; a second run writes nothing
Proof: the relevant item
Verified: 2026-09-30 (100%, checked by Sonnet verifier — "native-first: all in place (11 items)", ~/.codex starts with the block, ~/.codex2 untouched, hook count 1; sync selftest 8/8 incl. Codex2 by link)
Definition of done: red on excess, green on a Muse-only day, live run prints today's numbers
Proof: the relevant item
Verified: 2026-09-30 (100%, checked by Astra — selftest 25/25; the Muse race PASS (round 2), the in-flight wait PASS (round 4), an unused stated fallback now surfaces as WARN with its reason and an unexplained call as FAIL (round 5); live: green, every paid search/page call today the Muse job's)
⚠️ its last two VERIFIED lines disagree (67% vs 100%) — showing the lower figure, unconfirmed
Definition of done: no rule file says the vendor file cannot be edited
Proof: the relevant item
Verified: 2026-09-30 (67%, checked by Astra — BUILD §4 and the global rules corrected; the relevant record line 816 still says "cannot be durably edited here", ticket d683813c waits on Nick's one-tap approval) · 2026-10-01 (100%, checked by Sonnet verifier against origin/main af946d3310 — "does not hold" in BUILD 0, "cannot be durably edited" in MACHINE-RULES 0, the old vendor line in the global rules 0, the ENFORCED line in RULE 61 present once; Nick approved ticket a074aeb7 at 7:14 am)
Definition of done: four assertions pass with the real vendor updater in a throwaway home
Proof: the relevant item
Verified: 2026-09-30 (100%, checked by Sonnet verifier — ALL PASS (4/4, plus setup))
Definition of done: Astra returns SHIP on the closed steps' proofs
Proof: the relevant item
Verified: 2026-09-30 (100%, checked by Astra — rounds 1–12 NOT SHIP, each finding fixed, tested and landed; round 13 SHIP at 10:21 pm, the relevant record; postmortem below) · 2026-09-30 (100%, checked by Astra — rounds 1–12 NOT SHIP, each finding fixed, tested and landed; round 13 SHIP at 10:21 pm, the relevant record; postmortem in the plan)
Definition of done: every fresh session used the free tool
Proof: the relevant item
Verified: 2026-09-30 (100%, checked by Astra reading the transcripts in evidence/fresh-sessions-2026-09-30 — five of five used free web search or fetch, none attempted a paid route; note: the dispatched helper was briefed with the rule by its parent session, which is how every dispatched agent here receives its rules)
Definition of done: every reachable Mac all in place; every other Mac named NOT DONE
Proof: the relevant item
Verified: 2026-09-30 (100%, checked by Sonnet verifier — Studio all in place (11 items); Mac mini all in place (4 items), confirmed first-hand over ssh; Chantelle's Mac and the teammates' Macs NOT DONE, no route from the Studio)