grimoire-know

[Required Reading]

1. Read ${CLAUDE_PLUGIN_ROOT}/GRIMOIRE-CONVENTIONS.md. Only § Skill composition, § Project context and § Historic (read-only consumption) apply to this skill. § TDD, § Sub-agent spawning rules for execute/quick, § Commits, § IDE-aware review, § User-facing language, and § Pause-point pattern do not apply — this skill is pure Q&A and never mutates the repo, makes no commits, and creates no pages.
2. If .grimoire/PROJECT.md exists in the project root, read it for project context. If it does not exist, proceed without it and, at the end of the run, suggest the user run grimoire-init once to enrich future answers (non-blocking — never refuse to answer because PROJECT.md is missing).
3. If .grimoire/HISTORIC.md exists, read it for recent-page context. If older context seems relevant to the question, browse .grimoire/bag/historic/ (newest suffix first). Missing → proceed without it.

[Objective] Answer the following question about the repository or the application it builds, faithfully and read-only:

$ARGUMENTS

[Hard constraints — read-only]

• Never use Write, Edit, NotebookEdit, or any file-mutating tool against the consumer project.
• Never run shell commands that mutate state (git commit, git push, git checkout -b, file creation/deletion, package installs, etc.). Read-only Bash (ls, cat, git log, git diff, git status, find, grep) is allowed.
• Never create a page folder under .grimoire/pages/ and never write to .grimoire/HISTORIC.md or .grimoire/PROJECT.md.
• Never make commits.
• When uncertain, say so plainly instead of guessing.

[Phase 1: Question analysis]

• Classify whether the question is answerable from the repo alone, requires web research, or both.
• Identify which files or directories are most relevant to the question without reading more than needed.
• Decide upfront whether the sub-agent will need WebSearch/WebFetch, so it can be briefed accordingly.

[Phase 2: Investigation]

• Spawn one general-purpose sub-agent with Read, Grep, Glob, read-only Bash, WebSearch, and WebFetch.
• Brief the sub-agent with a self-contained prompt that includes the user's question, the relevant files/dirs identified in Phase 1, and explicit instructions to:
  • Inspect only the files needed to answer the question.
  • Use WebSearch/WebFetch when the answer depends on facts outside the repo (third-party libs, standards, recent docs, API behavior).
  • Collect every URL it actually consulted.
  • Flag any claim it is not confident about, instead of guessing.
  • Return a structured response with: (a) the answer, (b) an "uncertain about" list (empty if none), and (c) the source links it used (empty if none).

[Phase 3: Response] Render the sub-agent's findings to the user in chat:

• Direct answer first. Lead with the answer to the question — no preamble.
• Uncertain about. Include this section only if the sub-agent's "uncertain about" list is non-empty; mark it clearly so the user can verify those claims.
• References. Include this section at the end only when the sub-agent actually consulted the web, listing each URL on its own line. If no web sources were used, omit the section entirely.
• No file changes, no commits, no HISTORIC.md updates.

[Phase 4: Cleanup] Nothing to clean up — this skill is fully ephemeral. If .grimoire/PROJECT.md was missing during Phase 1, remind the user that running grimoire-init once would enrich future answers. End.