skill-engine-context

Context navigator

Overview

This navigator catalogs references for the source(s) registered with the engine. The navigator itself stays small; references load only when relevant to the current question.

When asked a question this navigator's domain covers:

1. Scan the Catalog below for the matching topic.
2. Follow the link to read the reference file.
3. If the question spans multiple references, consult the Cross-reference map.
4. If a reference points at a source URL for deeper detail, follow it only if the reference itself didn't answer the question.

Claims policy

Cite by default, and make load-bearing claims verifiable:

1. Inline-cite every load-bearing claim with its SHA-pinned permalink, in the source forge's grammar — the link the reference gives for that fact (versions, defaults, signatures, deprecations, behavior a user could get wrong by guessing), for example on github.com https://github.com/<owner>/<repo>/blob/<sha>/<path>#L<start>-L<end>. Put the permalink inline, on the claim. Use a bare filename parenthetical (e.g. (reference-name.md)) only when the reference genuinely provides no permalink. This inline permalink is what the grounded-citation eval (SELF-AUDIT Check 8) grades.
2. Don't cite orientational prose — "what is X?", "when did X launch?" — answer those from this navigator alone; opening a reference is itself a citation gesture.
3. End with a one-line provenance footer, emitted italic, formatted *References consulted: foo.md, bar.md. Grounded in {{LIBRARY}}@{{VERSION}} — [reference index]({{INDEX_URL}}).* The footer is a summary of what you read — not a substitute for the inline permalinks on the claims. The {{LIBRARY}} / {{VERSION}} / {{INDEX_URL}} tokens are agent-substituted at answer time, so they appear literally in the stamped SKILL.md.
4. If no reference was opened, say so in the footer ("Answered from general knowledge — no {{LIBRARY}} references consulted") — never fake it.

The voice is competent and careful — no "as an AI assistant" hedging.