pika-dev: implement issue agent service and deploy hardening

[justinmoon]

Attempt to make a little background bot to implement stuff. Should maybe move this to my personal infra.

Notes: codex resume 019cb713-68d5-7f30-b723-dc4dfa4cf784 in /Users/justin/code/pika/worktrees/pika-dev

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: f1ecb420-4fba-4413-808c-9b330f595319

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Post copyable unit tests in a comment
• Commit unit tests in branch pika-dev-main

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[justinmoon]

Agent review:

Issues found

1. Duplicate system prompt (runner.ts:133-135) — The initial prompt concatenates
   buildSystemPrompt + buildIssuePrompt, but buildSystemPrompt is already passed to
   createSession as systemPrompt at line 87. The agent gets the system prompt twice —
   once via the SDK's system prompt mechanism and once in the user message. This is
   wasteful but not broken (agents handle redundancy fine). Should probably just send
   buildIssuePrompt as the prompt.

2. No worktree cleanup on completion (runner.ts:132-176) — After a session
   completes or fails, the worktree stays on disk. The spec mentions "prune old
   worktrees after PR is merged or session is stale (>7 days)" as Phase 4, but
   there's no cleanup at all currently. Over time this will accumulate disk usage.

3. git add -A in worktree (git.ts:130) — The agent worktree could accumulate junk
   files (node_modules from test runs, build artifacts, etc.) and git add -A stages
   everything. A .gitignore in the repo handles most of this, but it's worth noting.

4. CSS served via readFileSync on every request (server.ts:37-41) — Not a real
   problem at this scale, but the CSS is re-read from disk on every hit. Could cache
   it at startup. Truly doesn't matter for now.

5. Dashboard event detail rendering (templates/session.ts:79) — payload.payload is
   a JSON string from the DB but rendered raw via textContent. Works, but the raw
   JSON blob isn't very human-readable in the UI. Future polish item.

6. No pagination — Dashboard lists are capped at 50/100 but there's no actual
   pagination UI. Fine for now with max 5 concurrent sessions.

Spec alignment

The implementation is ~98% faithful to the spec. The only material gap is the
NixOS deployment module (infra/nix/modules/pika-dev.nix) and justfile recipes,
which are deployment concerns rather than application code. Everything else from
the spec is implemented and working.

Lines: ~3,050 (spec target was <2,000). The overshoot is from thorough tests
(1,094 lines) and the DB layer being more verbose than estimated. Core application
logic is well within budget.

Bottom line

Ship it. The duplicate system prompt is the only thing I'd fix before deploying —
everything else is polish for later iterations.