ix-debug

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 entry point.

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

Pro check (optional)

Run once at the start:

ix briefing --format json 2>&1

If it returns JSON with a revision field, Pro is available. Extract openBugs and recentDecisions for use in Pro steps below. If it errors, skip all [Pro] labeled steps.

[Pro] If openBugs is non-empty, scan for a known bug matching this symptom before proceeding. If found, surface it immediately — an existing bug record may already have candidates or a fix. [Pro] If recentDecisions is non-empty, scan for recent decisions or context that might explain the symptom or constrain the likely fix. Surface any relevant match before continuing.

Goal

Answer: where in the execution path is this likely failing, and why? Stop once you have 1–3 root cause candidates with supporting evidence.

Phase 1 — Locate the entry point (always)

ix locate $ARGUMENTS --format llm

If $ARGUMENTS is a symptom description rather than a symbol name, also run:

ix text "$ARGUMENTS" --limit 10 --format llm

Identify the most likely entry point (where the failure originates or first manifests).

Phase 2 — Explain (always)

ix explain <entry-point> --format llm

Extract: role, callers, callees, confidence. Identify whether this is:

• A boundary (external input, API, event) — failure likely from unexpected input
• An orchestrator — failure likely from wrong sequencing or state
• A utility/helper — failure likely from wrong assumptions by caller

Stop if: the explanation makes the failure source obvious → skip to Output.

Phase 3 — Decide: inline or delegate

Use the Phase 1–2 results to choose the path:

• Inline path (simple bug): the likely failure is still within a single subsystem, confidence is >= 0.7, and the entry point is not an orchestrator with more than 10 callees → continue to Phase 4.
• Delegate path (complex bug): confidence is < 0.7, OR the entry point is an orchestrator with more than 10 callees → use the Agent tool with subagent_type: "ix-memory:ix-bug-investigator" and pass the pre-computed context below.

You MUST pass pre-computed context so the agent skips redundant work. Launch the agent with:

Investigate: $ARGUMENTS

Pre-computed context (skip Steps 1–2): Entry point: [symbol, subsystem, file — from Phase 1] Entity type: [boundary / orchestrator / utility — from Phase 2] Explain output: [paste ix explain result]

Start from Step 3. The symptom is: [description]. The entry point classification suggests: [hint from Phase 2].

If the Agent tool is unavailable, continue inline through Phases 4–6, reduce breadth, preserve the 2-read cap, and surface uncertainty rather than over-reading.

Phase 4 — Trace the execution path (inline path)

ix trace <entry-point> --downstream --depth 2 --format llm

Walk the downstream path. At each step, look for:

• Functions that validate or transform state (potential incorrect assumptions)
• Cross-subsystem calls (where contracts might differ)
• Functions with high callee count (potential god functions, many failure points)

Narrow: Identify the 1–3 nodes most likely to contain the bug.

Delegate if: the trace crosses subsystem boundaries, reveals multiple plausible contract boundaries, or fans out widely enough that confidence drops below 0.7. Use the Phase 3 delegation prompt and tell the agent to start from Step 3.

Stop if: trace reveals an obvious candidate on a mostly single-subsystem path → proceed to Phase 6.

Phase 5 — Callers (inline path, if failure might come from upstream)

ix callers <entry-point> --limit 10 --format llm

Check whether the fault is in how this is called rather than in its own logic.

Phase 6 — Targeted code read (inline path, only at suspected failure points)

For each root cause candidate (max 2):

ix read <candidate-function> --format llm

Read the specific function only. Look for:

• Edge cases in input handling
• Assumptions about state that might be violated
• Missing null/error checks
• Incorrect sequencing

Hard limit: 2 ix read calls maximum. If still ambiguous, surface the candidates and uncertainty to the user.

Phase 7 — Synthesize

• If you delegated in Phase 3 or 4, present the agent's result directly. Do not re-run locate/explain/trace in the main thread.
• If you stayed inline, use the Output format below.
• If Pro is available and this is a new bug, append the bug-logging suggestion after the investigation output.

Output

## Debug: [entry point]

**Execution path:**
[entry-point] → [step] → [step] → [suspected failure point]

**Root cause candidates:**
1. [function/file] — [reason: what assumption might be wrong]
2. [function/file] — [reason]

**Evidence:**
- [what graph data supports each candidate]
- [what code read revealed, if any]

**Confidence:** [high / medium / low] — [why]

**Next steps:**
- Add logging at [specific point] to confirm
- Check [specific edge case] in [function]
- Run `/ix-investigate <X>` to understand [unclear component] more deeply

**[Pro]** If this is a new bug, log it:

    ix bug create "<symptom title>" --severity <low|medium|high|critical> --affects <entry-point>

(Omit if Pro unavailable or bug already tracked.)

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.