analyze

Analyze

Read ../references/harness-folder-rules.md first. Read ../references/helper-surface.md when you need deterministic inspection, parsing, or local evidence collection.

Workflow

1. Inspect the current host state before making any recommendations.
2. Use deterministic helpers for facts:
   • host profile paths
   • discovered instruction files
   • settings parse results and MCP inventories
   • malformed settings or parse-error detection
   • discovered skills, agents, plugins, and local origin hints
   • the prior .harness snapshot when it exists
3. If .harness does not exist yet, treat that as normal. Use the live host files as the only starting point and describe the snapshot files that would be created.
4. Summarize what is clearly shared, clearly host-specific, malformed, stale, or ambiguous.
5. Ask only the follow-up questions that require semantic judgment:
   • whether similar guidance should become shared
   • whether a plugin or MCP definition should stay host-local
   • whether origin evidence is strong enough to record
6. If not --dry-run, refresh or initialize .harness so it captures:
   • the current host snapshot
   • the user's prior decisions
   • origin evidence and confidence notes
   • enough context to avoid asking the same question again later

Output

• inventory of current prompts, settings, skills, agents, plugins, and memory candidates
• current shared-vs-host-local opportunities
• malformed MCP/settings findings
• safe optimization ideas
• origin research hints for skills, agents, and plugins

Dry Run

• In --dry-run, do not write files.
• Show the current inventory, the proposed .harness snapshot updates, unresolved questions, and recommended next actions.
• If .harness is absent, show the snapshot files that would be initialized instead of assuming they already exist.

Origin Research

• Start with local evidence from manifests, cache metadata, package metadata, .git, README files, and known path conventions.
• Summarize that evidence before doing broader web research.
• When local evidence is incomplete, continue with GitHub or marketplace research and present confidence clearly to the user.