Repository navigation
Codex use receipt through its SDK surfaces, layer-scoped, with retained native outputs (GPT-6 reviewed) - #312
Conversation
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
f8cfe0f to
6dbd019
Compare
6dbd019 to
6a5ae96
Compare
|
GPT-6 cross-family review of generation DEFECTS FOUND
|
|
GPT-6 verification pass on head DEFECTS FOUND
Repaired in generation |
…ed native outputs Generations -1..-6 of the codex use receipt. -6 retains the sanitized native Codex outputs with a SHA256SUMS, requires the same resume thread, rejects five negative controls, and is scoped with layer_refs to foundation/native-clients and foundation/agent-sdks (GPT-6 cross-family findings). Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
7327c3c to
cba7b52
Compare
Summary
This PR adds a use receipt for the Codex CLI (0.155.1) on the workstation
nativestack-5975wx-20260925. It drivesgpt-6-astraat effort max through Codex's non-interactive SDK surfaces:codex exec --json --sandbox read-only --output-schema.codex exec resumeon the same thread under a second schema.Both answers are checked against independent facts from
catalogs/landscape/foundation.json. Only the latest generation,-6, counts. The rows below come from the receipt JSON:-1codex-cli 0.155.1 model gpt-6-astra effort max sandbox read-only | codex answer LAYER_COUNT=20 | independent python count 20-2turn 1 event types: ['item.completed', 'item.started', 'thread.started', 'turn.completed', 'turn.started'] | items: ['agent_message', 'command_executi-3CODEX_SDK_SURFACES_VERIFIED | codex 0.155.1 | same thread resumed: True-4CODEX_SDK_SURFACES_VERIFIED | codex-cli 0.155.1 | negative controls rejected 3/3 | retained 8 native files, SHA256SUMS sha256=b4ccf4232b344b4ca0fc7001-5CODEX_SDK_SURFACES_VERIFIED | codex-cli 0.155.1 | negative controls rejected 3/3 | retained 8 native files, SHA256SUMS sha256=53551a67e891ae810f390adb-6CODEX_SDK_SURFACES_VERIFIED | codex-cli 0.155.1 | negative controls rejected 5/5 | retained 8 native files, SHA256SUMS sha256=2b74370f62c46a985847e0dbWhat
-6retainsevidence/artifacts/codex-use-20260926-6/: both--jsonevent streams, both schema outputs, both schemas, and both stderr streams. They are sanitized withhost_receipts.sanitize(), and theirSHA256SUMSsha256 is printed on the receipt's first output line..gitignoreexceptions scoped to the-5and-6directories. The repository ignores*.jsonl, and the exception follows the mover-v3 precedent.caton local client configuration (~/.codex,~/.claude, or plugin caches), that item's output is replaced by its byte count and sha256 before hashing. This machine-specific config is never published.thread.startedto carry turn 1'sthread_id. It rejects five negative controls: a wrong count, a wrong first layer, a removedturn.completedevent, a removed resumethread.startedevent, and a mismatched resumethread_id. The command exits 0 only when the positive check passes and all five controls are rejected.foundation/native-clientsandfoundation/agent-sdks, the two layers codex wins whose roles this run exercises.Status changes (generated from the component matrices)
The workers-layer alternative "Codex native workers" stays
receipts_recorded. This run says nothing about owned writing children or worktrees.Review history
-2was held at needs_changes on backed claims.-3fixed that, and an independent session agreed.GPT-6 cross-family review of
-3(Codexgpt-6-astra, effort max; posted below as a comment) found three problems:Fixes: Layer-scoped host-receipt verification; version-aware supersession (GPT-6 + Opus verified) #321 added layer scoping and version-aware supersession, then
-4was recorded.-4: the independent review held it at needs_changes because the repository's*.jsonlrule kept its event streams out of the commit, and one stream contained local client configuration.-4's streams remain unpublished and unregistered.-5: an independent review agreed on all four points. It re-ran the command in a throwaway copy, checkedSHA256SUMSon a clean HEAD, compared item structure to confirm only client-config output was withheld, and simulated the agree to confirm only the two named layers change.-6: a GPT-6 verification pass (posted below) found that-5's checker accepted a missing resumethread.startedevent,tid2 in ([tid], []), a regression from-3.-6requirestid2 == [tid]and adds two thread controls. An independent session agreed on all four points and reproduced both new controls.Residuals
-6's limitations still says "three negative controls" (left over from-4); its claim and output say five. Receipts are immutable, so this stands as a known wording error.Verification
scripts/validate.py,host_receipts.py validate,component_matrix.py --check,new_host_grand_list.py --checkandevidence_manifest.py --check: all pass.verdict_review_gate.py: passed.git archive HEADcopy passesvalidate.pyandevidence_manifest.py --check, and itsSHA256SUMSverify.SOTA sources
codex exec,--json,--output-schema,exec resume): https://developers.openai.com/codex/noninteractive and https://github.com/openai/codex/blob/rust-v0.155.1/docs/exec.mdsandbox_mode,model_reasoning_effort): https://developers.openai.com/codex/config-reference🤖 Generated with Claude Code