Blast Radius

The agent that owns a bad deploy — where your team already works.

Blast Radius joins your workspace as a real member (Ambiguous AI), watches every deploy, and closes the risk loop end to end: it estimates the impact of a pull request before merge, verifies that estimate against live traffic after deploy, alerts the team with evidence — not opinion — and rolls back the offending commit only when a human says so. Every outcome is written back to the workspace: a Task, a postmortem Doc, a Ledger Sheet row. Nobody ever leaves the channel.

It's not a chatbot you talk to. It's a teammate that watches production while you keep doing your job.

A bad deploy looks like this: someone bumps gpt-4o-mini to gpt-4o on Friday and cost per request jumps 24×. A commit titled "handle special orders" quietly adds a 503 path. A new endpoint ships without a null check. Nobody notices until the bill arrives or the errors pile up — then the rollback is a panic, the blame is a meeting, and the postmortem is a doc nobody reads.

It's not only about LLM calls. A model swap or a longer prompt changes cost. A new endpoint or a removed null check changes the error rate. A sequential await or an added network call changes latency. Blast Radius reads the whole diff and reports on all three.

Monitoring tools show you dashboards. Blast Radius shows you which commit did it, what it will cost before you merge, and a one-command way back — in the channel where your team already talks.

From a rehearsal deploy (blast-radius-demo, PR #31 — 11 commits):

The prediction bridge, the diff-based blame, and the surgical commit rollback are the parts you won't find in a chatbox.

cd agent
cp .env.example .env        # fill in the keys (table below)
npm install
npm run bootstrap           # provisions the "Blast Radius" agent user + Ledger sheet
npm run dev                 # tsx watch — hot reload

GET /health → { "ok": true, "db": "up" } — the agent listens on port 3001.

docker build -t blast-radius-agent ./agent
docker run -d --init --name blast-radius -p 3001:3001 \
  -v blast-radius-data:/data --env-file ./agent/.env blast-radius-agent

Multi-stage build, base image pinned by digest to the latest patched Node 22, non-root user, healthcheck, lockfile-pinned deps, and a build-time npm audit gate (fails on high/critical). The image ships zero secrets.

1. Webhooks — expose the agent (cloudflared tunnel --url http://localhost:3001) and point GitHub webhooks of the demo repo at it (Pull requests + Pushes, secret = GITHUB_WEBHOOK_SECRET). Smoke test: POST /deploys with a SHA.

2. Baseline — deploy the current state of blast-radius-demo and send ≥ MIN_REQUESTS (5) requests so there is a previous release to compare against.

3. Big PR — open a PR with several commits (model bump, prompt change, new endpoints, error paths). INSIGHT comments within seconds.

4. Merge — Vercel deploys; the agent records the deploy (the demo app reports service.version from VERCEL_GIT_COMMIT_SHA automatically).

5. Seed — scripts/seed-traffic.js or any ≥5 clean requests → 🔴 cost regression alert with a commit table.

6. Trigger errors — any of:

   Trigger	Endpoint	Result
   prompt contains escalate	POST /api/chat	502
   order_id % 7 === 0	GET /api/orders?order_id=14	503 quarantined
   order_id % 13 === 0	GET /api/orders?order_id=13	503 payment hold
   missing order_id	GET /api/orders	500
   item without name	POST /api/report	500

   → 🚨 error spike alert: error log ↔ commit ↔ author table, the vague-message commit blamed from its diff.

7. Decide in the channel — status <deploy> (fresh numbers + evaluation history), then rollback <commit-id> from the table → "reverting to just before X" → the revert deploy ships, and the ledger + postmortems update.

8. Inspect the workspace — per-evaluation postmortem Docs, the Ledger Sheet, the Watch Task.

Guards: rollback requires an open approval window and the deploy must be the tip of the default branch — refusals always come with a reason, in the command's own thread.

PR ──► INSIGHT (per-hunk estimate, PR comment + channel)
 └─► merge ──► deploy recorded ──► traffic ──► evaluation vs previous release
      └─► breach ──► alert (evidence table) ──► human command ──► rollback
            └─► postmortem Doc + Ledger row + Watch Task  (per evaluation)

• Verdicts — cost_regression (≥ COST_REGRESSION_PCT, default 30%), latency_regression (≥ LATENCY_REGRESSION_PCT), error_spike (error rate up). ok/insufficient_data/skipped_revert stay silent. One verdict per deploy — but every evaluation is preserved: the same deploy can be documented first as cost regression, then as error spike, each with its own doc and ledger row.
• Evidence over opinion — error messages are matched to the commit that introduced them by message and by diff; deterministic matches outrank the LLM's ranking.
• Deterministic fallbacks — the price table computes cost impact; hunk ids anchor the INSIGHT table; if the LLM fails, the pipeline degrades to pattern inference instead of stopping.
• Crash-safe rollback queue — SQLite-backed intents with CAS claiming; a crash mid-rollback recovers on boot.

Channel (13 types): INSIGHT pre-merge post · cost/latency regression alerts · error spike alert (evidence table) · status reply (+ history) · keep reply · timeout reply · rollback queued ack · rollback result · rollback skipped · already-rolled-back note · revert-deploy note · unknown-command reply. GitHub: INSIGHT PR comment. Workspace objects: postmortem Doc · Ledger Sheet row · Watch Task.

npm run reset wipes DB, channel, tasks, docs and sheet for a clean rehearsal.

MIN_REQUESTS defaults to 5 to keep rehearsals fast; raise it to 20+ for production statistical confidence (the original rehearsal spec used 20).

GitHub webhooks verified with HMAC-SHA256 + timingSafeEqual; ingest/deploys token-gated; parameterized SQL throughout; no shell execution; outbound HTTP only to fixed hosts; .env never tracked (history scanned clean); Dependabot alerts enabled; npm audit clean.