first autonomous agent

Goal

Create an autonomous agent that monitors for new pull requests, runs /deep-review automatically, and posts the findings as a PR comment — no human session required.

Current Behavior

• /deep-review exists at plugins/deep-review/commands/deep-review.md and produces structured markdown output (verdict, severity-tagged findings, test gaps, bottom line).
• The skill is agent-capable (disable-model-invocation: false, allowed tools include Agent and Bash) but today only runs inside an interactive Claude Code session.
• There is no webhook listener, scheduled job, or .github/workflows/ directory in this repo to drive reviews without a human at the keyboard.
• The built-in /loop and /schedule skills are bound to an active Claude Code session, so they cannot survive the terminal closing or the machine sleeping.

Proposed Solution

Three deployment options — pick one before building:

Option A: GitHub Actions workflow (recommended for v1)

• Add .github/workflows/pr-deep-review.yml triggered on pull_request events.
• The workflow invokes Claude via the Anthropic SDK (Python or Node) — not Claude Code.
• Port the deep-review prompt from plugins/deep-review/commands/deep-review.md into a standalone script the workflow can run.
• Post findings back with gh pr comment inside the workflow.
• Runs on GitHub-hosted runners — no laptop/desktop dependency.
• Pay-per-token against the Anthropic API.

Option B: Self-hosted watcher (laptop / desktop / container)

• Background daemon polls gh api for new PRs or receives webhooks.
• Invokes Claude Code headless (if/when supported) or the Anthropic SDK directly.
• Posts findings via gh pr comment.
• Requires an always-on machine and network ingress if using webhooks.

Option C: Azure Function + webhook (aligns with #86)

• Deploy an Azure Function that receives GitHub PR webhook events.
• Orchestrates the review (optionally hybrid local LLM + Opus synthesis per #86).
• Hooks into the broader infrastructure proposed in #86.

Answers to the original questions

• Where does the agent live? Not on a personal laptop or desktop — availability drops when the machine sleeps. For v1, GitHub Actions (Option A) is the lowest-friction host. For long-term, an Azure Function (Option C) composes with #86.
• Can I use my Claude Max subscription? No. Claude Max authenticates via interactive OAuth inside the Claude Code client and cannot be used programmatically. An autonomous agent needs an Anthropic API key with pay-per-token billing.
• Do I need to pay for tokens? Yes, with any of the three options. See #85 (deep-review token optimization) and #86 (local LLM alternative) for cost-reduction strategies.

Acceptance Criteria

1. Opening a PR on this repo triggers a deep-review run without manual intervention.
2. The review result is posted as a single PR comment using the standard /deep-review output structure (verdict, findings, test gaps, bottom line).
3. The automation does not require a developer's machine to be online.
4. The chosen deployment option is documented, including how the Anthropic API key is stored (GitHub Actions secret, Azure Key Vault, etc.) and the expected per-review cost.
5. At least one prior merged PR is replayed through the pipeline to confirm the headless output matches what /deep-review produces interactively.
6. Regression safety: /deep-review continues to work interactively inside Claude Code — the automation path must not force a rewrite of the existing skill.

Open questions

• Which deployment option ships first? (Recommendation: Option A — lowest infra cost, fastest to validate.)
• Should automatic reviews be gated by a PR label (e.g., review:auto) so draft PRs don't burn tokens?
• Do we block this on #86 landing, or ship Option A standalone and migrate later?
• How do we prevent re-reviewing the same PR on every commit push? (Only review on PR open, or diff-since-last-review?)

Related

• #85 — Optimize deep-review skill for token usage (directly affects this agent's per-review cost)
• #86 — Local LLM Infrastructure for PR Code Review Pipeline (the full architecture that Option C plugs into)
• #93 — deep-review agent prompt improvements (quality of what this agent will be posting)