Agent Dev Workflow

Agent-friendly dev workflow (bin/ + worktree isolation)

This skill scaffolds a bin/ task-runner that lets many copies of a project run concurrently on one machine — one per git worktree — each with its own database and port, sharing a single Postgres container. It's the contract AI agent orchestrators (Conductor and similar) want: register setup / run / cleanup once and every session gets an isolated environment for free. As a high-value side effect, tests run against a dedicated database and can never wipe your dev data.

It's modeled on GitHub's scripts-to-rule-them-all, but uses bin/ (the name script/ never caught on) and adds the worktree-isolation layer.

When this is the right tool

A project backed by Postgres where you want any of: concurrent worktree dev, orchestrator setup/run/cleanup hooks, an end to tests clobbering dev data, or a replacement for a docker-compose-just-for-local-postgres. If there's no database, most of this doesn't apply. If there's no concurrency need and tests already use a separate DB, the payoff is small.

The core idea: identity derives from the worktree

One shared Postgres container (started once, reused by name — path-independent, so it survives worktree deletion). Each context gets a database and a port per process it runs, derived from its worktree path:

A worktree isn't necessarily one process: a repo may run an HTTP API + a gRPC service + a Vite frontend per worktree. The pattern generalizes cleanly — the project's port band claims one range per process kind, _common.sh grows one picker per kind, and bin/setup emits every derived port plus the wiring vars that connect the processes (e.g. ORCHESTRATOR_URL). One note: gRPC ports aren't curl-checkable — readiness-probe them with a plain TCP check.

Because identity comes from the path, two sessions never collide and re-running setup is idempotent. Every derived value has an env override (APP_DATABASE, APP_PG_PORT, PORT, …) for explicit orchestrator control.

What's invariant vs. project-specific

The whole point of distilling this from several real Postgres-backed projects is to know which parts you copy verbatim and which you adapt.

Invariant (copy from references/bin/, just swap the app/APP_ prefix):

• worktree-aware DB naming (app_db_name, app_db_name_for_root, is_main_worktree, app_hash)
• shared-container management (ensure_postgres, wait_for_postgres, app_psql) and its machine-scoped counterpart bin/stop (below)
• DB create/recreate helpers; the setup/reset-db/db/cleanup/stop/test scripts (keep these names — e.g. db, not query — so the muscle memory transfers across repos)
• port picking (port_in_use, find_available_port, app_pick_port) — including the macOS lsof/ss split; do not "simplify" it away (see gotchas)
• stdout=env / stderr=status discipline
• the dev-session state machine (fullstack variant): the atomic noclobber claim of .dev/state.env with PHASE=booting → PHASE=running only after ports listen, snapshot-coherent reads, the ownership-checked EXIT/INT/TERM trap, per-child process groups via set -m (never kill 0), and app_dev_stop
• pid + start-time identity (app_pid_lstart, app_pid_is) consulted before every kill and every health verdict (see gotchas — pid recycling), and the TERM → KILL → verify escalation in app_dev_stop (see gotchas — a single SIGTERM is not a teardown)
• deriving the session's child list from state.env (app_dev_child_keys) rather than hardcoding it per stack (see gotchas)
• the bin/gc proof structure: mandatory gates + at least one containment proof, the long-lived-branch denylist, the canonical-DB guard, and the fd-3 record stream (see gotchas — stdin-slurping loops)

Project-specific (the knobs you set):

• the prefix, PG user/password/image, container/volume names
• the publish address (APP_PG_BIND, default 127.0.0.1) — keep it on loopback unless a container genuinely must be reachable off-box. Note this is frozen at container creation: changing it later does nothing until the container is re-created, so ensure_postgres warns on the drift (see gotchas)
• the base port band — the default PG port plus one worktree range per process kind (backend HTTP, gRPC, frontend, …). This is the project's claim on the machine: worktree isolation keeps a project's own copies apart, but the base band is what keeps different projects from colliding when several run at once. Pick a distinct, uncommon band per repo (see gotchas) and keep PG + every process range in one coherent band.
• the Postgres major — pin the same one production runs (references/snapshots.md)
• app_server_dir — repo root (single package) vs a subdir (monorepo)
• app_migrate — how migrations run → references/migrations-and-seeds.md
• single-service dev (frontend runs separately) vs fullstack dev (an attach-aware per-worktree singleton that starts both, supervises both, and kills both) → use references/bin/dev or references/bin/dev-fullstack
• bin/gc's integration branch — APP_GC_BASE_REF defaults to origin/main; set origin/develop in repos that integrate on develop
• the seed posture — synthetic SQL, snapshot-as-seed, or empty + per-suite fixtures (references/migrations-and-seeds.md), and whether it has prod snapshots at all (references/snapshots.md)
• whether the repo has auxiliary services and what shared endpoints to seed into worktree envs (next section)

Auxiliary services: shared per machine, never per worktree

What replicates per worktree is the app + its database — nothing more. Auxiliary services (a GTFS validator container, a local OIDC IdP, a fake object store) run once per machine and are shared by every worktree instance; each worktree's env just points at the shared endpoints. N worktrees hitting one validator is fine — one container instead of N is the difference between "docker is fine" and "my laptop is swapping".

This fixes the docker-compose question cleanly:

• No aux services (the dev flow is just app + DB) → no compose at all. bin/ owns the shared Postgres container directly, as prescribed.
• Aux services exist → compose covers only those: docker compose up -d once per machine. The compose file never contains the per-worktree app or Postgres — those stay with bin/.

Don't expand bin/ into orchestrating aux containers; hold the boundary instead: bin/ = per-worktree identity (DB, ports, app processes); compose = machine-wide singletons.

Env seeding is the connective tissue: bin/setup emits the shared endpoints alongside the per-worktree values (e.g. VALIDATOR_URL=http://localhost:9010, overridable via env like everything else), so one stdout capture threads into the run step and every worktree converges on the same shared services.

Worktree lifecycle: create under .claude/worktrees/, sweep with bin/gc

Agent worktrees should live under the repo's .claude/worktrees/ directory (gitignore **/.claude/worktrees/ so nested checkouts never show as untracked files). That location is what makes the end of the lifecycle automatic: a worktree is alive while its PR is in review and becomes garbage the moment it merges — but merges happen on the reviewer's schedule, with no local session listening. bin/gc closes that gap with a lazy, state-based sweep: it enumerates linked worktrees under any */.claude/worktrees/ path and removes only the ones it can prove merged (ancestry, git cherry patch containment, or a merged PR via gh — plus mandatory gates: clean, unlocked, not the main/current worktree, not a long-lived branch), stopping any recorded bin/dev session and dropping the derived database along the way. Everything unproven is skipped with a one-line reason; a --nudge mode is cheap enough to run from a SessionStart hook. Worktrees created outside .claude/worktrees/ escape the sweep entirely — that's the trade you make by putting them elsewhere.

Teardown has two scopes — worktree and machine

Every teardown script except one is worktree-scoped, and that is deliberate: bin/cleanup drops one derived database, bin/dev stop ends one session, bin/gc sweeps the worktrees it can prove merged. None of them stops the Postgres container, because no single worktree can know whether another still needs it.

That leaves a real gap the first version of this skill didn't fill: nothing was able to say "this container is idle machine-wide, stop it." On a laptop running several of these repos at once, the containers accumulate and the only recourse is docker stop by hand — which loses every guard the rest of the pattern provides.

bin/stop is that one machine-scoped script. It stops the shared container only after proving it idle: no live bin/dev session in any worktree of the repo (enumerated from git, so sessions in scratch dirs outside .claude/worktrees/ still count), and no client backends connected to any database in it. Unproven → one-line reason, container left running, exit 0. Same prove-then-act shape as bin/gc; --dry-run and --force are combinable.

One portability note: the session proof needs the fullstack dev's .dev/state.env machinery. In single-service repos those helpers don't exist, so the proof is skipped and the verdict line says so rather than claiming "no dev sessions" for a check that never ran. Containment survives, because a running single-service dev server holds a database connection and is caught by the connection proof.

Two things it deliberately does not do. It never removes the container or volume — stopping is non-destructive and bin/setup restarts it with every database intact. And it does not lock other sessions out: a concurrent bin/setup calling ensure_postgres will bring the container right back. That race is by design (availability beats shutdown), so a container that keeps reappearing is telling you another session is working in the repo — information, not a failure.

bin/gc vs bin/stop — complementary, not overlapping

They share a shape — prove, act only on what's proven, skip the rest with a one-line reason, exit 0 either way — which makes them easy to confuse. They prove different things about different scopes:

That last row is the real distinction. bin/gc finds a running session in a worktree it has proven merged and shuts it down — it earned the authority to kill by proving the work already landed on the integration branch. bin/stop has no such proof available, so a live session is a veto: someone is working, and pulling the database out from under them is never right. One treats a live process as garbage to collect; the other treats it as a stop sign.

Their verdicts also age differently. Merged stays merged, so a bin/gc verdict computed a minute ago is still true. Idleness is a snapshot that can be invalidated a second later — hence bin/stop's re-verify immediately before acting, and the documented race above.

The workflow is sequential: bin/gc, then bin/stop. gc is what makes stop succeed — sweeping merged worktrees stops their sessions and drops their databases, which is exactly what turns a busy container into a provably idle one. gc then leaves the container running because it cannot know whether the main checkout still needs it; stop is the only thing entitled to answer that. Running stop first usually just prints a refusal.

One caveat when reading bin/stop's first proof: it enumerates worktrees of this repo. A separate clone of the same project shares the container by name but is invisible to git worktree list. That's why the connection check exists alongside the session check — a second clone running its own stack shows up as client backends even though its worktrees are unreachable from here.

Build order

1. Read the relevant references first (below) — they encode hard-won bugs.

2. Copy references/bin/* into the project's bin/, rename the prefix, fill the constants in _common.sh. Claim a base port band well outside 5432 and distinct from your other projects (see gotchas — or derive it from a hash of the repo name). Pin the same Postgres major as production so snapshots restore cleanly (references/snapshots.md).

3. Set app_server_dir and app_migrate for this project (references/migrations-and-seeds.md).

4. Choose the dev variant (single-service vs dev-fullstack). The fullstack variant carries the whole session protocol (singleton claim, attach, status/stop/logs, supervision) — copy its mechanics verbatim and adapt only the two child-launch blocks. Copy bin/gc alongside it and set APP_GC_BASE_REF if the repo integrates on a branch other than main.

5. Wire test isolation — the preload that points tests at app_test (references/test-isolation.md). Do this when (or before) the project grows a DB-backed test suite: it's the step that stops tests wiping dev data. It's a clean add later, but if you know tests are coming, scaffolding bin/test + the preload up front means the first test is born-isolated.

6. chmod +x bin/* (not _common.sh). If a docker-compose.yml exists, evict its Postgres (and app) services — delete the file outright if that's all it had; keep it only as the once-per-machine aux-services runner (see "Auxiliary services" and gotchas). Update .env.example + the project's agent docs (CLAUDE.md or equivalent) to point at bin/.

7. Verify end to end (don't assume): bin/setup on a clean checkout; the sentinel-row test from test-isolation.md; bin/dev in two worktrees lands on two different ports; bin/dev twice in the same worktree → the second attaches (exit 0, KEY=VALUE block) instead of double-booting; kill one child process → the whole stack exits loudly (no half-dead session); bin/gc --dry-run reports sane verdicts; bin/cleanup in a throwaway worktree leaves the main DB intact. bash -n every script.

   Two teardown checks are worth doing explicitly, because both failures report success: launch a child that ignores SIGTERM (trap "" TERM), then confirm bin/dev stop escalates to SIGKILL and reaps it instead of exiting 0 with the process still alive; and run bin/stop --dry-run both with a live session up (must refuse) and with nothing running (must report idle).

Reference files

Where this meets ci-quality-gates

This skill and ci-quality-gates are complementary and meet at exactly one seam: the test script. ci-quality-gates owns the script contract (lint/format/format:check/typecheck/test) and the CI workflow that runs bun run test against an ephemeral service-container Postgres with an explicit DATABASE_URL. This skill owns what that same test script connects to locally: bin/test + the bunfig.toml preload force an isolated app_test DB, and the preload's ??= is the handshake that lets CI's explicit DATABASE_URL win. Adopt both together for a DB-backed app — the CI skill defines the gate, this skill keeps it from clobbering dev data. (SquadQuest is the worked example: a "test": "bun test" script, bin/test, and a bunfig.toml preload that defers via ??=.)

Some repos deliberately run CI with no database at all (live-DB suites self-skip). That's a legitimate posture too — but the preload must then skip when Postgres is unreachable instead of failing the run. See test-isolation.md §CI for both postures.

Guardrails

• Keep the scripts to database + process lifecycle. Rich demo/fixture data is the app's concern and churns fast — at most a thin idempotent seed hook (migrations-and-seeds.md). A bloated bin/ stops being portable.
• These are dev-only. CI keeps its own explicit DATABASE_URL against an ephemeral DB (a GitHub Actions service-container Postgres, owned by ci-quality-gates); the test preload's ??= defers to it. Verify CI stays green; you shouldn't need to touch the CI workflow.
• Don't drop the canonical dev DB casually — bin/cleanup guards it behind --force; bin/reset-db is the intended "start the main DB over" path.