Repository navigation
Traceable, instant, abuse-safe briefs (audit P0s + result-first UI) - #10
Conversation
- Parser: invalid/missing evidence kind and unquoted evidence become inference; negative, non-finite or absurd hour figures are dropped. - Client grounds every fact against the user's own words and downgrades unquoted ones. - Rules ignore negated and already-automated clauses and abstain with a question instead of forcing a recommendation; evidence now quotes the user's sentence. - Rule-based replies ship instantly (removed ~11s artificial typing delay). - Analytics: removed the duplicate direct-capture tracker; every property value is enum or boolean, so report wording can no longer leave the browser. - Live mode: per-client rate limit, concurrency cap, daily budget, 45s timeout, client cancellation, bounded history, PDFs parsed once. Limits fall back to the rule brief. - Brief leads with the recommended move, a source quote, a traceability count and a pilot plan; engine badge always visible; AA text contrast; 44px touch targets; 16px input. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
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: 21c71dbcf8
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".
- quoteAppearsIn now requires each ellipsis-separated segment to appear verbatim, on word boundaries and in order; the unordered word-overlap fallback is gone. - The server returns the attachment text it read (decoded files, extracted PDFs) as a data-source part; the client grounds against typed text, decoded text files and that part. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: da3730a3e6
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".
- Quote normalisation keeps Unicode letters and numbers (NFKC, case-folded). Scripts written without spaces (Han, Kana, Thai) match on substrings; spaced scripts keep word boundaries. - The already-automated matcher no longer treats every auto* word (automobile, autograph) as automation; it matches automatic/automated and named auto-actions only. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: ecb55e9350
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".
The transport strips data-* parts before sending, so the attachment text returned for grounding stays in browser state and is not resent every turn (which could push valid attachment conversations past the request-size limit). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 71c1ba873c
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".
When older turns are dropped, the collected user material (workflow, constraints, attachment text) is prepended as one bounded user turn, framed as data, so refinements keep the original evidence. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 089980a322
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".
The already-automated and negation filters describe current state, so they now apply only when analysing an existing process. Create mode reads the full request again. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…r-audit-design-a2577a # Conflicts: # app/globals.css
What
Implements the P0 items from the 30 Oct audit plus the highest-value UI changes from the "Signal / Evidence" direction.
Trust
parseReportJson: a claim is afactonly when explicitly labelled and quoted; invalid/missing kinds and unquoted evidence becomeinference. Negative, non-finite, string or >52,000 h/yr hour values are dropped.groundReport/quoteAppearsIn: the client checks each fact against what the person typed and downgrades unsourced ones.affirmedText), so negated or already-automated steps no longer trigger fixes. With no supported signal it abstains ("No supported improvement yet") and asks one question. Evidence now quotes the user's own sentence.Speed: rule-based replies ship at once (the ~11 s artificial typing delay is gone).
Privacy: removed
WorkflowPostHog.tsx(a second tracker with no prod guard that fired "completed" on submit).primary_opportunityis nownamed | none, and the sanitizer drops any property value outside an allowlisted enum.Cost (live mode only):
lib/live-guard.tsadds 8 req/10 min per client, 4 concurrent, a daily cap (LIVE_DAILY_REQUEST_LIMIT, default 300), a 45 s timeout plus client-disconnect abort, the last 12 messages only, and PDFs parsed once. A tripped limit returns the rule brief with a notice, not an error.UI: the brief opens on the recommended move, a source quote, a proof strip (workload labelled "not savings", steps, "traceable to your words N of M") and a "How to test it" pilot. The engine badge is always visible. Faint/muted tokens meet AA (calculated). 44 px targets on touch, 16 px composer text, shell min-height 420 px.
Why
The audit showed briefs could look authoritative without support, free results were wrong on negated input, report wording could reach analytics, and the paid endpoint had no spend controls.
Risk
affirmedTextdrops whole clauses that contain negation words. It errs toward abstaining. Samples and fixtures still produce recommendations.How it was tested
npm test: tsc, next build, 61/61 node tests, including the newtests/trust-guards.test.mjs.npm run lintis clean.Not verified: production deploy, VoiceOver, 200% zoom, the iOS Safari keyboard, live model quality, and n8n export (out of scope).
🤖 Generated with Claude Code