@firstpick/pi-extension-feature-system-prompt

Classifies Pi requests and adds a short, fail-closed routing bridge to the enabled feature-development-workflow skill only for feature work.

Install

pi install npm:@firstpick/pi-extension-feature-system-prompt

Pi discovers feature-system-prompt.ts through the package manifest. The feature-development-workflow skill must also be enabled.

Behavior

Before the main agent starts, the extension first applies a conservative local fast path for explicit non-feature work and known short continuations. Ambiguous or potentially additive requests use an isolated, tool-free Pi session with the active conversation model. The classifier must return an exact JSON object with only kind and reason; kind is one exact request kind, while reason is normalized to one plain-text line and bounded to 500 characters. The feature kinds are feature_lightweight and feature_complex; either appends a short complexity context and a mandatory bridge telling the parent to load the enabled feature-development-workflow skill with read. Other successful classifications leave the system prompt unchanged.

The bridge preserves progressive disclosure rather than duplicating the skill body. Before selecting the bridge, the extension resolves the advertised <location> for feature-development-workflow from <available_skills>, verifies that its SKILL.md is readable and non-empty, and—when classified complex—verifies that references/COMPLEX-FEATURE-CONTRACT.md beside it is also readable and non-empty. A failed availability check injects a configuration-error policy instead of the bridge.

The fast path recognizes only narrow bug, review, research, explanatory-question, and troubleshooting forms, and refuses prompts with additive capability signals. A narrow local continuation check resolves short, explicit follow-ups from the previous effective kind; when no prior kind exists, the active-model classifier remains authoritative. Session starts and tree navigation reset that in-memory continuation state.

In RPC mode only, the extension publishes two replayable statuses. The existing "feature-decision-output" key carries versionless JSON such as {"kind":"feature_complex","reason":"The request crosses the extension and UI contract."}. The separate "feature-category" key remains exactly "lightweight-feature" or "complex-feature". A consumer must require the category to match the structured decision kind; for rolling compatibility, it may also accept the legacy exact-label decision payloads "feature_lightweight" and "feature_complex" with a deterministic fallback reason. Non-feature, unavailable, invalid, and reset states clear both statuses with undefined. Known local continuations replay the prior structured feature decision, including its reason.

Configuration and requirements

Enable feature-development-workflow through Pi settings or a package entry. The extension does not read an external feature prompt file and does not bundle a second copy of feature policy.

Reload Pi after changing the extension, skill, or settings so resource discovery and extension state are rebuilt.

Security and privacy boundaries

• The local fast path receives only the current request and does not persist it.
• The model classifier receives only bounded current request text and, for a conservative continuation, bounded previous request text plus its effective classification.
• The nested classifier uses an in-memory Pi session, no tools, and a minimal resource loader to avoid recursive extension loading.
• The extension does not read the enabled skill during classification or send skill contents to the classifier.
• The untrusted classifier reason is used only in the serialized RPC status. It is never interpolated into the privileged system prompt; routing and injected complexity context use only the validated exact kind.
• Classification telemetry and prompt text are not persisted by this extension.

Failure behavior and limitations

If no active model is available, the classifier fails, times out, or returns malformed structured output, the extension appends a short fallback telling the parent to classify from request and repository evidence, then load the feature skill only when applicable. JSON with an unknown kind, missing or extra fields, a non-string reason, or an empty normalized reason is invalid.

A successfully classified feature receives the bridge only after the availability check passes. If the enabled skill is missing from the current system prompt or a required file is unavailable, unreadable, or empty, the extension injects a configuration-error policy directing the parent not to implement the feature until configuration is restored.

The availability check is machine-enforced; following the loaded policy remains an agent instruction rather than an authorization system.

Development

npm test
npm run check
npm run smoke
npm pack --dry-run --json

Tests cover taxonomy and strict decision parsing, bounded classifier input and reasons, continuation decision reuse, feature-only bridge injection without reason interpolation, fail-closed language, lifecycle resets, structured RPC statuses, and inert registration.

License

MIT