ix-infrastructure/ix-claude-plugin/ix-investigate

ix-claude-plugin shared model

Check command -v ix first. If unavailable, use Grep + Read as fallback.

Argument parsing

Strip --save and any following path token from $ARGUMENTS before resolving the target.

• If --save <path> is present, set SAVE_PATH to that path.
• If --save is present without a path, auto-generate ix-investigate-<target-slug>.md in cwd (target slug = the target with spaces and slashes replaced by -).
• If --save is absent, SAVE_PATH is empty — do not write a file.

Pro check (optional)

Do not run ix briefing — you have already been told. The plugin's UserPromptSubmit hook injects an [ix] Session briefing once per TTL whenever Pro is reachable. Seen one this session? Pro is available; read recentDecisions from it. Not seen one? Skip every [Pro] step. Asking costs a whole turn to learn something already in your context.

Goal

Answer: what is this, how does it connect, and what's the execution path? Stop as soon as those three questions can be answered accurately.

Phase 1 — Locate (always)

ix locate $ARGUMENTS --format llm

If multiple matches: use --kind, --path, or --pick N to resolve. Do not proceed until the entity is unambiguous.

If ix locate returns nothing: try ix text $ARGUMENTS --limit 10 --format llm.

Phase 2 — Explain (always)

ix explain <resolved-symbol> --format llm

Extract: role, importance, caller count, callee count, confidence score.

If the resolved entity is a class or module, also run:

ix overview <resolved-symbol> --format llm

This reveals internal structure (members, sub-components) without reading source.

Orphan check: If fan_in = 0 AND fan_out = 0 in the ix explain output:

• Report: "Symbol is a graph orphan — no detected dependencies. Either the graph needs a refresh (ix map) or the file has no parseable import/call relationships."
• Suggest ix map --silent from the project root as the first step (ix map takes a directory; it rejects a single file).
• Stop here — skip Phases 3–5 unless the user specifically asks for source-level inspection.

Evaluate: Is the explanation sufficient to answer the question?

Stop if: explain gave clear role, purpose, and connection summary → skip to Output.

Phase 3 — Connections (run only if caller/callee detail needed)

Run only the directions you need — not both by default:

# If "who uses this" matters:
ix callers <symbol> --limit 15 --format llm
 
# If "what does this do internally" matters:
ix callees <symbol> --limit 15 --format llm

Stop if: you now know who uses it and what it depends on.

Phase 4 — Trace (run only if execution flow is unclear)

ix trace <symbol> --depth 2 --format llm

One trace only. Pick the most representative direction (--upstream or --downstream) based on the question.

Stop if: execution path is now clear.

Phase 5 — Code read (last resort only)

Only if the above steps leave a specific implementation question unanswered:

ix read <symbol> --format llm

Read the symbol only — never the full file. If the symbol is a class, read the specific method suspected.

Hard limit: One ix read call maximum. If still unclear after reading, surface the ambiguity to the user rather than reading more.

Phase 6 — Design context [Pro]

If Pro is available and recentDecisions from the briefing is non-empty, check for decisions affecting this symbol:

ix decisions --topic <resolved-symbol> --format text

Include any relevant decisions in the output under Design context.

Output

## [Symbol] — Investigation
 
**What it is:** [kind, file, subsystem — from graph]
**Role:** [orchestrator / boundary / helper / utility / etc.]
 
**Execution flow:**
[downstream: what it calls → what those call, 2 levels max]
[upstream: who calls it, top 5]
 
**Key connections:**
- Depends on: [top 3 callees]
- Used by: [top 3 callers with their subsystem]
 
**Design context:** [Pro only — relevant recorded decisions, or omit section if none]
 
**Evidence quality:** [strong / partial / uncertain] — [one-line reason]
 
**Next step:**
- [most useful follow-up based on findings]

If confidence < 0.7 in ix output, label those claims as [uncertain] and recommend ix map to refresh.

Save step (only if `SAVE_PATH` is non-empty):

• Write the full output above to SAVE_PATH using the Write tool.
• Confirm to the user: Saved to <SAVE_PATH>.
• Do not write the file if --save was not passed.