Menu
Use when working with the local GPT/Gemini browser automation system, including service lifecycle, job creation, platform dispatch, SSE listening, Chrome extension validation, or agent integration docs.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
| name | auto-chat |
| description | Use when working with the local GPT/Gemini browser automation system, including service lifecycle, job creation, platform dispatch, SSE listening, Chrome extension validation, or agent integration docs. |
Use the real auto-chat CLI as the source of truth for local debugging and self-tests.
Do not validate user-facing workflows through npm run job:* or direct service imports. Keep npm run build, npm run check, and npm test for development verification only.
For production-style local use, install the package globally and initialize the agent integration:
npm install -g <package-or-tarball>
auto-chat --version
auto-chat init
auto-chat status
auto-chat --version (or auto-chat -v) prints the installed CLI version, read from the packaged package.json. Use it to confirm which build is actually on PATH after a reinstall.
auto-chat init installs this skill into common global agent skill directories, starts the background service, opens chrome://extensions, and always prints Chrome extension installation steps with the GitHub zip URL and local npm package zip path.
For repository development only:
npm install
npm run build
npm pack
npm install -g ./auto-chat-cli-*.tgz
auto-chat --help
Do not use npm link for production-flow testing. If the CLI output does not come from the intended installed package, rebuild, pack, and reinstall the tarball.
Start and stop the local task service only through the CLI:
auto-chat start
auto-chat status
auto-chat stop
The service must run in the background. Do not leave a foreground auto-chat server, npm run dev:server, or direct node apps/server/dist/index.js session running for normal debugging.
Use JOB_SERVER_URL=http://127.0.0.1:<port> only when intentionally isolating a test service on a non-default port.
Create and inspect jobs through the real CLI. Prefer platform-specific dispatch so a GPT task does not wake Gemini and a Gemini task does not wake GPT:
auto-chat add examples/text-job.json --replace
auto-chat add examples/job.json --replace
auto-chat add examples/gemini-job.json --replace
auto-chat add examples/gemini-text-job.json --replace
auto-chat list
auto-chat show <jobId>
auto-chat doctor <jobId>
auto-chat dispatch --platform gpt <jobId>
auto-chat dispatch --platform gemini <jobId>
auto-chat concurrency 3
auto-chat listen <jobId>
Use auto-chat listen <jobId> --json when raw SSE event shape matters. Prefer passing both --platform and <jobId> to dispatch; this asks the extension to claim that specific queued job and avoids older queued tasks on the same platform being picked first. Omitting <jobId> claims the oldest queued task for the selected platform. Omitting --platform wakes all platforms.
Use auto-chat concurrency to inspect the current plugin scheduler max concurrency, and auto-chat concurrency <1-8> to update it. The default is 1.
Text output lives at data/jobs/<jobId>/outputs/output-01.txt. Image output lives under data/jobs/<jobId>/outputs/; use image_order events to confirm image order.
Gemini image jobs should use prompts: string[] for multi-image tasks. Each array entry must describe exactly one image; the extension sends entries one by one because Gemini generates one image per conversation. GPT image jobs may still use a single multi-image prompt.
GPT and Gemini text jobs both support optional sourceImages; set mode: "text" for text output and mode: "image" for image output. Gemini source images are pasted directly into the composer, not uploaded through the file picker. The extension waits until Gemini's send control is no longer disabled before submitting.
For text outputs, the extension uses the platform copy action where available. If the clipboard still contains an old command beginning with auto-chat, treat it as stale automation text and keep waiting for a real copied response before marking the job failed.
Set an optional "outputDir": "<dir>" in the job JSON to have image outputs copied to that directory in addition to the normal data/jobs/<jobId>/outputs/ location.
outputDir only copies image (kind: "output") artifacts. Text-mode jobs (mode: "text") ignore outputDir entirely — text outputs are never copied anywhere else.data/jobs/<jobId>/outputs/ file is always written first and is unaffected by outputDir — the extra directory only receives a copy. If the copy fails (bad path, no permission), the job's own output is untouched and the job does not fail; only an output_copy_failed event is recorded.auto-chat add invocation, not the background service's directory.auto-chat add examples/output-dir-job.json --replace
auto-chat dispatch --platform gpt img_output_dir_test_001 && auto-chat listen img_output_dir_test_001
auto-chat show <jobId>, auto-chat doctor <jobId>, and auto-chat listen <jobId> all report the resolved outputDir and copy status (waiting / copied N files / copy failed — check events.jsonl for output_copy_failed when a copy fails, e.g. an unwritable path). The copy is best-effort: a failed copy does not fail the job or block the normal outputs/ artifact.
Set persistTab: true to keep the browser tab open after a job finishes. Set parentJobId on a follow-up job to reuse the parent's existing tab (or fall back to its conversationUrl if the tab is closed). Both GPT and Gemini support this.
Example flow:
# Step 1: run parent job with tab kept open
auto-chat add examples/persist-tab-job.json --replace
auto-chat dispatch --platform gpt persist_test_001 && auto-chat listen persist_test_001
# Step 2: run follow-up job in same conversation thread
auto-chat add examples/followup-job.json --replace
auto-chat dispatch --platform gpt followup_test_001 && auto-chat listen followup_test_001
The follow-up job sends into the existing conversation — the model retains prior context. Do not close the parent tab between steps if you want guaranteed tab reuse.
After extension or shared protocol changes:
npm run build
To create the committed zip that users download from GitHub:
npm run pack:extension
Reload Chrome from:
apps/extension/dist
The extension should automate only the user's own logged-in GPT/Gemini pages. Do not drive GPT/Gemini manually through browser tooling when the local service, CLI, SSE, or job files can provide the needed state.
Before reporting completion for code changes:
npm run build
npm run check
npm test
auto-chat --help
auto-chat init
auto-chat start
auto-chat status
auto-chat stop
For task-flow changes, also run a real CLI job flow against the background service and inspect the CLI output:
auto-chat start
auto-chat add examples/text-job.json --replace
auto-chat dispatch --platform gpt text_test_001
auto-chat list
auto-chat show text_test_001
auto-chat doctor text_test_001
auto-chat stop
For full HTTP/SSE/webhook details, use docs/agent-integration.md. The skill is the quick operating guide; the doc is the detailed protocol reference.