Amp Bulk Executor

Run an Amp thread across many repositories in parallel and save structured results to JSONL.

Prerequisites

Node.js ≥ 18
Amp CLI installed (global or local)
- Global: npm i -g @sourcegraph/amp && amp --version
- Local: npm i -D @sourcegraph/amp && npx amp --version
.env with your keys (see .env.example)
- AMP_API_KEY — Amp API key
- SOURCEGRAPH_API_URL — your Sourcegraph code search URL
- SOURCEGRAPH_API_TOKEN — Sourcegraph access token
Optional: set AMP_BIN if amp is not on PATH

Where to get tokens

Amp API key: sign in at https://ampcode.com/settings and create/locate your API key, then export it as AMP_API_KEY.
Sourcegraph URL and token:
- Enterprise/self-hosted: your URL is your instance base; create a token at https://{yourinstance}.com/users/{yourusername}/settings/tokens (replace host/username).
- Trying OSS code: use the public code search at https://sourcegraph.com/search and point SOURCEGRAPH_API_URL=https://sourcegraph.com (create a personal access token in your Sourcegraph user settings).

Try the demo (best starting point)

Quick demo (5 repos):

npm run demo

Full demo (45 repos listed in input/repos.txt):

npm run demo:full

Results are written to a new file under runs/ named run_<timestamp>.jsonl.
The CLI prints the path at the end as: Output file: runs/run_<timestamp>.jsonl.

Customize and run yourself

Install dependencies

npm install

Add repositories to input/repos.txt

Use simple slugs like owner/name (e.g., vercel/next.js). The agent can resolve the full repo.

(Optional) Edit the prompt

The default prompt lives at prompts/authentication_analysis.md.
Edit this file to change the analysis and output schema, or pass a different prompt file with -p.

npm start
# or with options
npm start -- -i input/repos.txt -c 10 -p prompts/authentication_analysis.md -o runs

Results

Each run streams results to runs/run_<timestamp>.jsonl (one JSON per line).
The CLI prints the exact output path at the end (look for Output file: ...).

How it works (at a glance)

For each repo slug, the app invokes the Amp CLI with your prompt.
The prompt uses a toolbox backed by your Sourcegraph code search instance to find and read code across the repository.
Results stream to JSONL as each repo finishes.

CLI Options

-i, --input Path to file with repo slugs (default: input/repos.txt)
-c, --concurrency Number of concurrent sessions (default: 10)
-p, --prompt Path to prompt file (default: prompts/authentication_analysis.md)
-o, --output-dir Output directory (default: runs)
--amp-logs Write per-repo Amp debug logs (advanced)

Output

Results stream to runs/run_<timestamp>.jsonl, one JSON object per line with:
- slug, timestamp, status
- data (depends on your prompt)
- optional thread_url

Tip: Visit the thread_url to see the full Amp thread history for each result, including the agent's actions, the context it considered, and the tools it used. This is helpful when troubleshooting results and iterating on your prompt.

Example line:

{"slug":"vercel/next.js","timestamp":"2025-09-05T21:18:13.292Z","status":"SUCCESS","data":{"repo_name":"github.com/vercel/next.js","commit_hash":"de5a1b2","auth_type":"no_auth","reasons":"…","evidence_samples":["…"]},"thread_url":"https://ampcode.com/threads/T-011f58fe-f5c5-4a7a-8416-1f7849656c43"}

Security

This tool invokes Amp with -x --dangerously-allow-all. Only run prompts you trust and never commit secrets. Keep your .env out of version control.

Advanced: Logging & Failure

Logging
- --amp-logs writes a debug log per repo under runs/ (capped for safety).
- When not set, a temporary debug log is used only to capture the Amp thread URL.
Failures (simplified)
- If the agent outputs FAIL: ..., the run is recorded as status: "ERROR".
- Each repo can retry up to 3 times; persistent failures end as ERROR.
- Each repo run times out after ~5 minutes.

Development

npm run build  # Compile TypeScript to dist/
npm run demo   # Build and run with default input
# or run directly via ts-node
ts-node --esm src/index.ts

sourcegraph/bulk-repo-analyzer