relay-code-pi

A Pi coding-agent package for the relay-code automation framework — Modal.com + Trigger.dev. It lets a non-expert "automater" build and deploy production-ready automations end-to-end by mirroring the framework's three-phase lifecycle (research → plan → execute) and turning its conventions (lazy-env pattern, schema contract, ALLOWED_TASKS sync, deploy order, secret handling) from advisory prose into enforced, deterministic Pi tools.

Install

Project-local (recommended — installs to .pi/settings.json, shareable with your team, auto-installs on trust):

pi install npm:relay-code-pi@latest -l

Global (installs to ~/.pi/agent/settings.json):

pi install npm:relay-code-pi@latest

Use -l for project-local (writes .pi/settings.json, shareable with your team, auto-installs on trust). Drop -l for the global user install.

Requires Pi ≥ 0.80.6 (for @narumitw/pi-goal agent_settled).

What it ships

• 10 deterministic tools (extensions/relay-tools.ts) — action tools that wrap the framework's CLIs (relay_test, relay_deploy_trigger, relay_deploy_modal, relay_smoke_test, relay_dev_worker, relay_locate_automation) + convention-enforcing tools (relay_add_task, relay_add_env_var, relay_add_schema_field) + a conformance relay_lint. A /plan command toggles Plannotator plan mode.
• Skills (skills/) — setup + lifecycle orchestration (relay-system-setup / relay-research-automation / relay-plan-automation / relay-execute-or-resume-automation / relay-update-or-fix-automation / relay-sub-agent-builder) and vendored integration API docs.
• Constitution — the framework's security rules and deploy order, scaffolded by relay-system-setup as the project-root prompts/AGENTS.md, which Pi auto-discovers and bundles into the system prompt every turn. It carries no project state; per-project progress lives in docs/automations/<plan>/progress.md, read on demand via relay_locate_automation.
• 6 sub-agents (agents/) — gated fan-out via pi-subagents (airtable, apify, gohighlevel, modal, trigger-dev, unipile).

Safety gates

• relay_deploy_modal refuses until relay_smoke_test writes the deploy-gate marker — the deploy order is mechanically enforced, not advisory.
• Secret values live in .env / .env.production. The user installs the user-level env-storage skill at ~/.pi/agent/skills/env-storage/ (a separate Pi skill, not shipped with this package) and runs /skill:env-storage (Load) BEFORE invoking relay-system-setup — or pastes Modal + Trigger.dev keys directly into the project's .env / .env.production. The agent never types a secret value into a tool input, so secrets never enter the model context.
• File-mutating tools run inside withFileMutationQueue and throw on a missing anchor rather than silently corrupting a hand-edited file.

Develop

npm install
npm test          # vitest — pure-core unit tests
pi -e .           # package-load smoke gate (expect SMOKE_OK)

Installing env-storage

The env-storage skill is a separate Pi package (env-storage-user-skill), installed once per machine at the user level. It does NOT ship with this package. To install:

pi install npm:env-storage-user-skill@latest

The skill content is dropped into ~/.pi/agent/skills/env-storage/, where Pi auto-discovers it globally. From any project, run /skill:env-storage to Load the user's master .env + .env.production secrets into the project, or Update the masters from the project. Run Load before invoking /skill:relay-system-setup.

Upgrading from ≤ 0.4.1

Earlier versions of relay-code-pi shipped skills/env-storage/ inside the package. That folder is now removed — pi update / npm update of this package will no longer carry the skill. If you relied on the package-shipped env-storage skill, install it separately:

pi install npm:env-storage-user-skill@latest

Existing user-data masters at ~/.pi/agent/skills/env-storage/.env and .env.production are preserved across the upgrade — the new package install only adds the skill files (SKILL.md + assets/).

Version history

Current version: 0.5.7. The publish flow keeps this table in sync — every release appends a row (see CLAUDE.md → Publishing). Versions marked "(published; no matching git commit — pre-reconcile drift)" exist on npm but were never recorded as a git commit on main; do not create commits or tags for them retroactively.

Version	What changed
0.1.0	Initial release — deterministic relay-code tools, lifecycle skills, 6 sub-agents, and the constitution.
0.1.1	Bundled third-party pi-extensions into the tarball so they load after pi install; added the files whitelist + MIT LICENSE.
0.2.0	Restructured lifecycle skills (consultation / plan / execute / sub-agent-builder); added the apify-actor-development skill; dropped the Telegram extension; stopped tracking .pyc bytecode.
0.2.1	Bundled pi-context-usage for /context visualization.
0.3.0	Dropped before_agent_start constitution injection and the bundled pi-goal / plannotator / pi-web-access packages.
0.3.1	Fixed the published tarball to include patches/.
0.4.0	Removed pi-secret-mask; shipped the .env-storage skill in-package; moved the constitution into prompts/.
0.4.1	Scaffolded dotfiles now ship as *.template (npm drops bare dotfiles); .env-storage now manages .env.production.
0.5.0	Split env-storage into its own user-level npm package (env-storage-user-skill) and dropped it from this package.
0.5.1	(published; no matching git commit — pre-reconcile drift)
0.5.2	(published; no matching git commit — pre-reconcile drift)
0.5.3	Deleted the dependency-bundling field (deprecated-spelling cleanup); added the pi-claude-marketplace extension; added the health-check starter task and maxDuration defaults to the scaffold; added prompts/ to the publish whitelist.
0.5.4	Dropped the pi-claude-marketplace extension added in 0.5.3 (no consumer in the package).
0.5.5	(published; no matching git commit — pre-reconcile drift; the release in which @llblab/pi-telegram was re-added without bundleDependencies, so it hoisted and its tools stopped loading)
0.5.6	Restored bundleDependencies (correct spelling, all 6 external pi packages listed) and re-added @llblab/pi-telegram to the manifest, fixing the pi-telegram load failure that first appeared after 0.5.5.
0.5.7	Architecture deepening from the codebase review (no behavior change): deploy-gate + dev-worker state machines moved into pure cores (src/cores.ts), new mutateFile mutation adapter in the tool glue, spec/plan/roster/deploy-order prose anchored to their canonical constants, and the six-agent "Context — where to look" preamble template aligned to the shipped agents.

License

MIT