disinto/dev/AGENTS.md

42 lines
3.5 KiB
Markdown
Raw Normal View History

<!-- last-reviewed: f32707ba659de278a3af434e3549fb8a8dce9d3a -->
# Dev Agent
**Role**: Implement issues autonomously — write code, push branches, address
CI failures and review feedback.
**Trigger**: `dev-poll.sh` runs every 10 min via cron. Sources `lib/guard.sh` and
calls `check_active dev` first — skips if `$FACTORY_ROOT/state/.dev-active` is
absent. Then performs a direct-merge scan (approved + CI green PRs — including
chore/gardener PRs without issue numbers), then checks the agent lock and scans
for ready issues using a two-tier priority queue: (1) `priority`+`backlog` issues
first (FIFO within tier), then (2) plain `backlog` issues (FIFO). Orphaned
in-progress issues are also picked up. The direct-merge scan runs before the lock
check so approved PRs get merged even while a dev-agent session is active.
**Key files**:
- `dev/dev-poll.sh` — Cron scheduler: finds next ready issue, handles merge/rebase of approved PRs, tracks CI fix attempts. Formula guard skips issues labeled `formula`, `action`, `prediction/dismissed`, or `prediction/unreviewed` (replaced `prediction/backlog` — that label no longer exists)
- `dev/dev-agent.sh` — Orchestrator: claims issue, creates worktree + tmux session with interactive `claude`, monitors phase file, injects CI results and review feedback, merges on approval
- `dev/phase-handler.sh` — Phase callback functions: `post_refusal_comment()`, `_on_phase_change()`, `build_phase_protocol_prompt()`. `do_merge()` detects already-merged PRs on HTTP 405 (race with dev-poll's pre-lock scan) and returns success instead of escalating. Sources `lib/mirrors.sh` and calls `mirror_push()` after every successful merge.
- `dev/phase-test.sh` — Integration test for the phase protocol
**Environment variables consumed** (via `lib/env.sh` + project TOML):
- `FORGE_TOKEN` — Dev-agent token (push, PR creation, merge) — use the dedicated bot account
- `FORGE_REPO`, `FORGE_API`, `FORGE_URL` — Target repository (FORGE_URL used to auto-detect git remote)
- `PROJECT_NAME`, `PROJECT_REPO_ROOT` — Local checkout path
- `PRIMARY_BRANCH` — Branch to merge into (e.g. `main`, `master`)
- `WOODPECKER_REPO_ID` — CI pipeline lookups
- `CLAUDE_TIMEOUT` — Max seconds for a Claude session (default 7200)
**FORGE_REMOTE**: `dev-agent.sh` auto-detects which git remote corresponds to `FORGE_URL` by matching the remote's push URL hostname. This is exported as `FORGE_REMOTE` and used for all git push/pull/worktree operations. Defaults to `origin` if no match found. This ensures correct behaviour when the forge is local Forgejo (remote typically named `forgejo`) rather than Codeberg (`origin`).
**Session lock**: fd-based flock — released during idle phases (`awaiting_review`, `awaiting_ci`) so other agents can proceed; re-acquired before injecting the next prompt. This prevents the lock from blocking the whole factory while the dev session waits.
**Crash recovery**: on `PHASE:crashed` or non-zero exit, the worktree is **preserved** (not destroyed) for debugging. Location logged. Supervisor housekeeping removes stale crashed worktrees older than 24h.
**Lifecycle**: dev-poll.sh (`check_active dev`) → dev-agent.sh → tmux `dev-{project}-{issue}` → phase file
drives CI/review loop → merge + `mirror_push()` → close issue. On respawn after
`PHASE:escalate`, the stale phase file is cleared first so the session starts
clean; the reinject prompt tells Claude not to re-escalate for the same reason.
On respawn for any active PR, the prompt explicitly tells Claude the PR already
exists and not to create a new one via API.