Verifying pi extensions

The runnable recipe — target, launch, doctor, feature map, cleanup — is .pi/verify/SKILL.md. This file explains what each check covers and where its evidence stops.

Run npm run check for repository conventions, types, lint, formatting, unused code and tests. npm run verify loads every root-manifest extension through a registration stub and checks the local skill library. The stub captures registrations; it does not dispatch events or execute tools. Missing TypeSafe credentials are informational offline. npm run verify -- --live explicitly opts into an external TypeSafe request using your configured credentials.

Run npm run smoke with pi 1.4.0 or later on PATH (or set PI_SMOKE_BIN to its executable path). This uses the actual pi RPC host, loads all manifest extensions, the owned workflow skills and prompt templates. It checks discovery and expansion of those resources from an unrelated working directory using the local provider, including /workflow. It checks /skill-search and /usage --json in RPC without provider calls, then enters plan mode, checks that plan_mode_question, plan_mode_complete, and plan_fetch_url are offered to the model, attempts a harmless touch smoke-blocked command through real tool dispatch, verifies it was blocked and that the file does not exist, exits plan mode, checks those three tools are no longer offered, explicitly enables subagents, and completes a scout child through the actual child-session implementation. It then has the fixture model write a file and stop, and checks that done-gate buys exactly one extra turn with its "Done check" message and keeps the reminder out of the transcript (display: false); checks that a write outside the workspace buys no turn; and runs /until test -f smoke-done, checking the agent is sent back until the file exists. The isolated host also loads built-in codemode explicitly. It runs controlled measurement trials and rejects empty work, then fails the same check after two distinct edits. It checks that native writes, shell writes and both nested variants are blocked with files absent, executes a real diagnostic observation, and checks that a subsequent write succeeds. It starts a real failed background job through deferred terminal_* tools, checks the actual exit status/output, proves one completion follow-up resumes the agent without steering, and drives an optional Jev attention watch against the redacted local fixture. The watch is inference, not verification; notify: false must not start another model turn. It also runs deferred find_code from codemode against a local TypeSafe fixture, checks URL-userinfo redaction at that HTTP boundary, and uses a structured MCP-shaped tool fixture (not an actual MCP transport) to check warning propagation to the outer codemode result without changing its schema. It fails on missing capabilities, failed assertions, or a 30-second RPC timeout.

The smoke creates a temporary working directory and agent configuration, uses an allowlisted environment, disables discovery/MCP/startup networking, and serves scripted OpenAI-compatible responses on loopback only. It needs no provider credentials and incurs no external model cost. Child sessions use the same temporary agent directory. TypeSafe requests point at the local fixture: semantic-search and injection-screen phases receive scripted typed answers; other phases exercise the existing unavailable-judge fallback. An explicit base URL overrides the classifier's catalog host, so a registered TypeSafe model still does not call api.typesafe.ai during smoke. scripts/ classifiers have no Pi registry and keep the direct HTTP transport against that same base URL. Normal user settings and sessions are not changed. The host, server and fixtures are cleaned up in finally on normal success or failure. Forced process termination can leave a pi-smoke-* directory in the system temporary directory.

npm run coverage runs the unit suite, offline registration verifier and real-Pi smoke under Node's built-in LCOV reporter, so it requires pi on PATH too. It works on Node 24+ without --test-coverage-include-all (which is Node 26-only). Generated coverage is ignored by Git. Run from a clean state on the CI runtime too; a local green result is not evidence of hosted CI success.

This smoke checks host integration and resource/extension wiring, not whether a model follows a skill well, model judgment, interactive rendering, external providers, or successful TypeSafe judgments. For an explicitly authorized live/UI check, start pi in a disposable directory with a separate PI_CODING_AGENT_DIR, provision only the intended provider credential there, and repeat /plan start, the harmless blocked bash command, /plan exit, /subagents on, and a child request. Record the actual tool result and child completion. Do not count a provider response or the registration stub as evidence that the blocked action or child execution worked.

For the local discovery/accounting commands alone, run node .pi/verify/local-commands.mjs. It loads all extensions in an isolated real Pi RPC host and verifies skill-search results, empty-branch JSON accounting and /workflow discovery without a listener or model calls. It can run where loopback listeners are prohibited. It does not cover prompt expansion, numerical model usage (covered by unit tests), or interactive skill-picker rendering.

Project recipe

Target: record the requested base/head and git rev-parse HEAD before running. For uncommitted work include staged, unstaged and untracked content identity; subagent results provide assessment.revision.fingerprint when Git is available. Any intervening edit invalidates earlier evidence for the final state.

The current deadwood build does not scan this repository's .mjs modules and reports unresolved sibling imports plus the intentionally host-provided Pi SDK as missing dependencies. Its failure is not a useful dead-code verdict here; record that limitation and use the repository's Knip/Fallow gates rather than adding runtime dependencies or suppressions to satisfy it.

Choose additional acceptance-specific checks for changes outside that smoke's coverage. A green registration stub alone is insufficient. For UI/native work, use an explicitly configured project verifier with those tools or report the capability gap. No universal extra tool grant is implied by this recipe.