GPT reviewer now skips anything listed in a cumulative
"Rejected Recommendations" section at the bottom of plan files.
Claude appends rejections there during integration, building a
negative cache across iterations.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
BSD sed (macOS) and GNU sed (Linux) interpret `sed -i ''`
differently — GNU treats the empty string as the script
argument, not the backup suffix. Replace with portable
sed-to-tmpfile-and-mv pattern.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Collapse the entire multi-file structure (bin/, lib/, prompts/) into
one executable bash script at the repo root. Everything is inlined:
frontmatter helpers, prompt strings, and the Claude stream-json
progress display (now a function calling node --input-type=module).
- plan-status is now `plan-refine status` subcommand
- Remove dead Puppeteer/Brave CDP code (chatgpt-send.mjs,
plan-refine-brave) — superseded by Codex CLI approach
- Remove npm artifacts (package.json, package-lock.json) — sole
dependency was puppeteer-core for the dead browser automation
- Simplify install.sh to just symlink + prereq checks
- Update .gitignore: node_modules/ -> *.feedback-*.md
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Connects to user's running Brave via Chrome DevTools Protocol
to automate ChatGPT interaction. Uses puppeteer-core to open a
tab, send the prompt, wait for response, and extract the result.
No cookies, no separate profiles, no copy/paste. Just connects
to the browser where the user is already logged in.
One-time setup: relaunch Brave with --remote-debugging-port=9222
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Oracle browser automation has macOS Keychain issues that prevent
cookie-based auth. Replaced with a two-command clipboard workflow:
plan-refine <file> # copies ChatGPT prompt to clipboard
plan-refine <file> --integrate # Claude CLI integrates feedback
Step 1 uses Oracle --render --copy (no browser needed) to bundle
the evaluation prompt with plan content. Step 2 uses claude -p
for headless integration.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Scripts now resolve symlinks before computing SCRIPT_DIR, so
they work correctly when invoked via ~/.local/bin/ symlinks.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
After Oracle captures ChatGPT feedback, automatically invokes
claude -p to integrate the feedback back into the plan file.
Full iteration (ChatGPT + Claude) now runs with one command.
Use --no-integrate to skip the Claude step and get clipboard
fallback for manual integration.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Semi-automated plan refinement pipeline using Oracle for ChatGPT
browser automation. Includes frontmatter-based tracking and a
dashboard scanner for plan status across projects.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
