Repository navigation
Conversation
docs(dispatch): fork lanes when created threads lose permission
Every SessionStart/UserPromptSubmit/PreToolUse text now carries a current fact, one cue and one owner pointer. Decisions, markers, dedupe keys, trust delimiters and data caps are unchanged. Loop CLI mutations print a short receipt; show and validate keep the full plan. Measurements: devlog/_plan/261009_prompt_reduction/evidence.
Shorten always-on hook injections under a layering standard
…s as routers Adds structure/70_prompt_architecture.md and check-prompt-architecture.mjs (in gate.mjs) with an exact-size baseline for skills still over budget. dev, loop and pabcd keep one owner per rule and point elsewhere; hosted-CI evidence, state files and dated incident notes move to references or the devlog. The SessionStart dispatch card shrinks to a pointer and its resolver cell lives in delegation.md.
…nner, keep host bound checks Review found three obligations without an owner (skill-relative links, native vs external skill discovery, installed-CLI recovery), scanner false greens (inline-code headings, emphasized classes, folded YAML, reference links), four dropped host-bound comparisons, and an alias change outside the reduction. All fixed; aliases reverted.
Layering standard, prompt-architecture gate, and core skills as routers
…ences Sixteen dev-* and other routers move procedure sections verbatim into references in the same skill and keep condition -> reference rows. Stubs are pointers. All SKILL.md files now meet their budgets, so the gate baseline holds only the test-required reader-documents copies and no skill is eligible for an exception.
Catalog descriptions and routers follow the layering standard
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 417ba6f894
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".
| "Absent tools stay absent; silence grants no approval. Subagents send question candidates to main.", | ||
| "Details: $codexclaw:cxc-dev references/async-questions.md.", | ||
| ].join(" "); | ||
| return "[codexclaw] User questions: request_user_input_async when exposed; do not wait; silence is not approval. Owner: $codexclaw:cxc-dev async-questions.md."; |
There was a problem hiding this comment.
During a $codexclaw:cxc-interview session, this unconditional SessionStart/PostCompact guidance tells the model to use request_user_input_async, directly conflicting with skills/interview/SKILL.md:17-20, which requires synchronous request_user_input because only returned synchronous answers are captured for ledger-backed readiness. Following the newly shortened hint can therefore leave Interview answers uncaptured and the readiness state pending; retain the previous “Outside Interview” qualification in this always-on line.
Useful? React with 👍 / 👎.
1 participant
Promotes
devtomainfor codexclaw 0.2.42: the prompt-reduction release.cxc loopmutations print a short receipt.structure/70_prompt_architecture.md), thecheck-prompt-architecture.mjsgate,cxc-dev/cxc-loop/cxc-pabcdas routers (56 → 22 KB), dispatch card 1.2 → 0.4 KB (SessionStart now 1.7 KB plain / 2.1 KB managed).No hook, handler or guard decision changed. Each PR passed hosted CI on its exact head and an independent review;
npm teston the release tree: 3,772 tests, 0 failures.