Score

One daemon runs the whole issue → PR → green → merged pipeline:

# from the repo root: cleanup + dispatch every tick, landing every 2 ticks, repair every tick
bun run start

bun run start --once --dry-run --verbose --no-merge

--once runs one full pass of every phase and exits. --dry-run reports what each phase would do without touching a session, a worktree, or a branch. --no-merge lets landing gate and soak but never commit the merge.

Score has exactly one contract with every project it cranks: the repo exposes make verify at its root, running that project's full verification (check, tests, build — whatever the repo deems necessary) and exiting nonzero on failure. Landing's merged-tree gate runs it before any merge, and agent briefings point at it; the recipe lives in the repo's own Makefile, versioned with the code, so Score carries zero language assumptions — Bun, Go, Rust, all the same to it. A repo without the target never merges.

Phases run strictly in order within a pass — cleanup → dispatch → landing → repair — which keeps the primary checkout single-writer and preserves the legacy authority split: dispatch never merges, landing never edits code, repair never merges. A phase that throws is logged and the pass continues.

Landing stages the exact head SHA it vetted (never the mutable branch name) and soaks per commit: a new push mid-soak restarts the count from zero, and an observation without a head SHA is refused outright. Every command the daemon runs is bounded by a deadline (120s default, 30min for make verify) and killed as a whole process tree on timeout — a hung gh or wedged verify can't stall a phase.

repair also stays a manual one-shot, for when a specific PR needs a nudge now:

bun run start repair --only 12,14 --dry-run --include-clean --no-spawn

Under the daemon, repair leaves a PR alone while the agent it already pinged is still working on it — session alive, nothing pushed, defects unchanged, and fewer than REPAIR_STALE_TICKS ticks since the ping. The manual subcommand has no such ledger and always acts.

Cleanup runs the same clock over issue worktrees with no PR at all: after REPAIR_STALE_TICKS ticks with no new commits the agent is pinged, and after a second silent window — or immediately once its session is gone — a worktree that is clean and has no commits ahead of base is reclaimed so the issue redispatches. A worktree holding real work is never removed: while its agent is alive it is reported loudly every tick, and once the agent is gone a fresh agent is respawned in place to finish the preserved work.

Environment

| Variable | Default | What it does | |---|---|---| | TICK_INTERVAL_MS | 60000 | The daemon's only clock; phases declare tick multiples. | | SOAK_TICKS | 2 | Consecutive green landing ticks before a merge. | | REPAIR_STALE_TICKS | 10 | Ticks before a silent agent is re-pinged; also each window of cleanup's stranded ladder (ping, then reclaim a clean no-PR worktree). | | MAX_PARALLEL | 1 | Issues in flight at once. | | MAX_MERGES | 5 | Merges per landing tick. | | SKIP_LABELS | hold,wip,do-not-merge | Labels landing refuses to merge. | | WORKTREE_ROOT | ~/wt | Parent of the per-repo worktree directory. | | EPIC_LABEL_PREFIX | epic: | Label prefix marking dispatchable issues. | | AGENT_CMD | claude | Command repair spawns in a worktree. | | SCORE_SERVER_URL | http://127.0.0.1:3000 | Read-only API used by score tui. |

Others keep their legacy names and defaults (GH_REPO, AUTO_PULL_MAIN, ONLY_ISSUE_BRANCHES, SESSION_SUFFIX); see apps/daemon/src/daemon/daemon.run.ts and apps/daemon/src/repair/repair.run.ts.

Managed mode

score config init writes ~/.score/config.jsonc; score up <project> ensures the fleet's loopback API is running and runs the project daemon supervised (launchd/systemd), reading only that project's resolved.json — no env tuning. Each project's agent.harness is either "claude" (tmux sessions, unchanged) or "opencode" (a durable HTTP session per issue against a locally-owned opencode serve child):

"score": {
  // ...
  "config": {
    "agent": { "harness": "opencode", "model": "anthropic/claude-sonnet-5" },
    // ...
  },
},

One child per daemon

A managed harness: "opencode" project starts exactly one opencode serve child before the poll loop begins, and every phase (cleanup, dispatch, repair) shares that same OpencodeService instance — sessions are addressed by exact title, never by a locally-cached ID, so a restart resolves the same conversation instead of creating a new one. Bootstrap preflights opencode --version (in place of tmux -V) and refuses to start if OPENCODE_SERVER_PASSWORD is set — the adapter has no HTTP auth support in v1, so a passworded child would spawn and then be unreachable. --dry-run still starts and stops the child (so the lifecycle is exercised end to end) while every create/prompt/abort/delete call is suppressed.

Unexpected exit is fatal, never mid-phase

If the child dies without the daemon having asked it to stop, the current phase is allowed to finish, the rest of that pass is skipped, and the daemon logs fatal: opencode child exited unexpectedly, writes it to status.last_error, and exits nonzero. The supervisor restarts the process; because opencode sessions are durable and title-addressed, the new daemon resolves the exact same session by title and resumes it with its prior context intact — the same shutdown path SIGTERM already used, just triggered by the child instead of a signal.

Real-binary smoke

Verified manually against opencode 1.17.15 (opencode serve --hostname 127.0.0.1 --port 0), driving the same HTTP calls OpencodeService makes:

1. Create: POST /session with title: "score-demo-smoke-1" → session ses_0101ed83cffe375BS6n4I31fgr.

2. Prompt: POST /session/{id}/prompt_async, "Reply with exactly one word and nothing else: pineapple" → assistant replies pineapple.

3. Kill child: kill -9 the opencode serve process — the next request against its port refuses the connection.

4. Restart: a fresh opencode serve starts on a new port.

5. Resolve same title: GET /api/session?search=score-demo-smoke-1 returns exactly one match — the same ses_0101ed83cffe375BS6n4I31fgr.

6. Resume: POST /session/{id}/prompt_async, "What single word did you reply with earlier in this conversation? Answer with just that word." → assistant replies pineapple, proving the resumed session retained the pre-restart context rather than starting a fresh conversation:

   user:      Reply with exactly one word and nothing else: pineapple
   assistant: pineapple
   [child killed, opencode serve restarted, same title resolved]
   user:      What single word did you reply with earlier in this conversation? Answer with just that word.
   assistant: pineapple

Supervisor platforms

score up / down / tui pick the supervisor by platform: launchd (~/Library/LaunchAgents) on macOS, systemd user units (~/.config/systemd/user/score-<key>.service) on Linux. Other platforms are unsupported and fail before touching anything.

The TUI reads fleet snapshots, dated log records, daemon telemetry, and GitHub merge history through apps/server; it does not open project state or log files or invoke gh itself. score up installs one fleet-level loopback server alongside the project daemons, so score tui needs no separate server command. bun run server:start remains the foreground development path, and SCORE_SERVER_URL can point the TUI at another Score server. Lifecycle keys still call the local supervisor adapter directly because the API remains read-only.

On Linux, systemd user units are killed at logout unless lingering is enabled. Run this once per operator account, or the daemons die with your SSH session:

loginctl enable-linger $USER

score doctor deliberately does not check this — it is documented, not enforced.

Layout

Bun workspaces + Turborepo. apps/{daemon,tui,server} are entry points (the score CLI is apps/daemon; tui is the Ink terminal viewer; server is the API-only Express app); packages/{shared,core,agents,tracker} are libraries — ports live in core, implementations in agents/tracker. Files are named <noun>.<role>.ts (.service, .policy, .render, .interface, .run); see AGENTS.md.

Verify

bun run check
bun run test
bun run build

License

Copyright (C) 2026 egoisutolabs.com

Score is free software: you can redistribute it and/or modify it under the terms of the GNU Affero General Public License as published by the Free Software Foundation, version 3 of the License only. See LICENSE for the full text.