pitchfork

name

pitchfork

description

Manage Gram's local development services (server, worker, dashboard, streams, etc.) with pitchfork — start/stop/restart daemons, check health, and query their logs. Activate whenever you need to bring up the local dev stack, restart a service after a code change, check whether a service is running or why it failed, or read/search service logs — even if the user just says "start the server", "is the worker running?", or "check the logs".

Pitchfork — Local Dev Services

Pitchfork is a daemon manager. All Gram local services are declared in pitchfork.toml at the repo root and run under a pitchfork supervisor. Each daemon's run command is a mise task (mise run start:<name>) and its readiness is polled via mise run check:daemon --name <name>, so start only returns once the service is actually healthy.

How to interact with pitchfork (in order of preference)

1. MCP server — if the pitchfork MCP tools are available (pitchfork_status, pitchfork_logs, pitchfork_start, pitchfork_stop, pitchfork_restart), use them. They return structured output and skip pager/formatting concerns entirely.
2. CLI — otherwise use the pitchfork CLI directly (commands below).
3. Mise tasks — mise run start / mise run stop bring the whole stack up/down. Use these for whole-stack operations even when the MCP/CLI are available, because mise run start also starts the supervisor and cleans up stale daemon entries. If the pitchfork binary itself is missing, run mise install first (mise provides it).

Daemons

Declared in pitchfork.toml. Names are shown namespaced by project (e.g. gram/server) in pitchfork list, but commands accept the short name (server).

Daemon	What it is
server	Main Gram API server
worker	Temporal worker
dashboard	Dashboard frontend dev server (depends on server)
streams	Go Pub/Sub consumers (gram streams)
pystreams-multi	Python Pub/Sub consumers
assistant-runtime	Assistant runtime
dev-idp, dev-idp-dashboard, mock-oidc	Local identity providers (login is credential-less)
admin	Admin app (depends on mock-oidc)
tunnel-gateway	Tunnel gateway

Whole-stack lifecycle

mise run start   # supervisor start + stop + clean + start --all-local
mise run stop    # pitchfork stop --all-local

mise run start cleanly stops and restarts everything. If services fail because Docker infra (Postgres, ClickHouse, etc.) isn't up, run ./zero --agent for full environment setup instead.

Individual daemons

pitchfork start is idempotent — it does nothing if the daemon is already running. After changing Go/backend code you must restart the affected daemon to pick up the change:

pitchfork start server worker      # start if not running (waits for ready)
pitchfork restart server           # stop + start (use after code changes)
pitchfork start server --force     # same as restart
pitchfork stop server

If commands fail because the supervisor isn't running, start it with pitchfork supervisor start.

Status

pitchfork list                     # all daemons + status (running/available/errored/...)
pitchfork list --json              # machine-readable
pitchfork list --status errored --status failed   # only broken daemons
pitchfork status server            # detail for one daemon, includes error/exit info

available means defined in config but not started. errored includes the exit reason — follow up with logs.

The MCP pitchfork_status tool returns a JSON array covering all daemons (equivalent to pitchfork list --json); there is no per-daemon MCP detail call, so use pitchfork status <name> on the CLI when you need one daemon's full detail.

Logs

This is the primary debugging tool. Logs persist across restarts and are timestamped.

pitchfork logs server -n 100               # last 100 lines
pitchfork logs server worker -n 50         # multiple daemons interleaved
pitchfork logs server --since 5min         # relative time window
pitchfork logs server --since '10:30' --until '11:00'
pitchfork logs server --grep error --grep panic   # OR'd, case-insensitive
pitchfork logs server --regex 'level=(warn|error)'

Tips for agent use:

• Add --no-pager --raw when piping or capturing output.
• Do not use --tail/--follow — it blocks forever. Prefer bounded queries (-n, --since) and poll if needed.
• After a restart, pitchfork logs <name> --since 1min shows just the fresh boot output.
• The MCP pitchfork_logs tool only supports "last N lines" (default 50); drop to the CLI when you need time windows or grep filters.
• Lines like [start:server] ERROR ... exited with non-zero status appear on every normal stop/restart — they are supervisor noise, not crashes. Look at the surrounding lines (clean shutdown sequence vs. panic/fatal) before concluding a service crashed.