Conversation
- New examples/read.ts entrypoint: demo mode generates 3 Acme Corp fixtures (PDF/DOCX/PPTX), custom mode accepts user files + question - New skills/read-source/SKILL.md teaching the agent witan read usage - New examples/lib/acme-fixtures.ts with deterministic fixture generators and ACME_QUESTION cross-document question - Fix format.ts to pair parallel tool calls with their results instead of printing results orphaned from their commands - Update AGENTS.md, README.md, package.json with read example docs - Clean up cross-skill references in xlsx-code-mode and xlsx-verify Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
hershaw
changed the title
hershaw
left a comment
hershaw
left a comment
There was a problem hiding this comment.
tested and added an example, lgtm
![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: 2cf5e797a2
ℹ️ About Codex in GitHub
Codex has been enabled to automatically 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 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| const basename = path.basename(resolved); | ||
| fs.copyFileSync(resolved, path.join(workDir, basename)); |
There was a problem hiding this comment.
In custom mode, each input is copied into the temp workspace under path.basename(...), so two different paths like dirA/report.pdf and dirB/report.pdf will silently overwrite each other and only one document will actually be available to the agent. This can produce incorrect cross-document answers while appearing to process all files, so the staging step should ensure unique destination names.
Useful? React with 👍 / 👎.
| ext := extFromContentType(resp.Header.Get("Content-Type")) | ||
| if ext == "" { | ||
| ext = filepath.Ext(urlPath(input)) |
There was a problem hiding this comment.
The URL path suffix is only used when extFromContentType(...) returns empty, but extFromContentType returns .txt for any unrecognized text/* value; that means URLs like .../notes.md served as text/plain are forced to .txt, so downstream upload uses text/plain instead of markdown-specific handling (notably affecting --outline). Treating generic text types as "unknown" (or checking the URL suffix first in that case) avoids this misclassification.
Useful? React with 👍 / 👎.
2 participants
Summary
examples/read.tsentrypoint for multi-document QnA usingwitan readto extract text from PDF, DOCX, PPTX, HTML, and 10 other formatsskills/read-source/SKILL.mdskill teaching the agentwitan readusage (formats, flags, strategy)format.tsto pair parallel tool calls with their results — previously parallel results printed orphaned from their commandsTest plan
pnpm read --helpprints usagepnpm readruns demo mode end-to-end (generates fixtures, agent reads all 3 docs, synthesizes correct answer)pnpm read --verboseshows full message stream with correct tool call/result pairingpnpm read report.pdf minutes.docx "question"custom mode with user filespnpm qnaandpnpm verify(non-read examples)🤖 Generated with Claude Code