Explanation Traces: Structured Decision Query

Overview

This skill reads the current session's session-trace.json and presents routing decisions, agent selections, skill phase transitions, and gate verdicts as a human-readable timeline. Every answer comes from what the trace actually recorded at decision time -- never from post-hoc reconstruction or rationalization.

Key constraints baked into the workflow:

• Read-only: this skill never modifies trace files or any other file
• Answers must come from recorded trace data, not from memory or inference about what "probably happened"
• If no trace file exists, explain how to enable tracing rather than guessing at decisions
• When the user asks about a specific decision, filter to that decision -- do not dump the full log
• Timestamps and evidence fields are authoritative; do not paraphrase away precision

Instructions

Phase 1: LOCATE

Goal: Find the session trace file.

Step 1: Search for trace data

Check for session-trace.json in these locations, in order:

1. Current working directory: ./session-trace.json
2. .claude/ directory: .claude/session-trace.json
3. Glob fallback: **/session-trace.json (in case a custom path was used)

Use the Read tool on the first match found.

Step 2: Handle missing trace

If no session-trace.json exists anywhere, stop and inform the user:

No session-trace.json found.

To enable decision tracing, add a hook that writes routing decisions,
agent selections, skill phase transitions, and gate verdicts to
session-trace.json. See: skills/explanation-traces/references/trace-schema.md
for the expected JSON format.

Do not fabricate trace data. Do not attempt to reconstruct decisions from memory or conversation history. The entire value of this skill is that it reads what was actually recorded -- without a trace file, there is nothing to read.

GATE: Trace file found and readable. Proceed only when gate passes.

Phase 2: PARSE

Goal: Extract decision points from the trace and filter to the user's query.

Step 1: Read the trace file

Parse the JSON from session-trace.json. Extract the decisions array. Each entry contains:

Step 2: Filter if user asked about a specific decision

If the user's query targets a specific decision (e.g., "why did you pick the code reviewer?", "why that agent?"), filter the decisions array:

Step 3: Validate data integrity

Check that each decision entry has all required fields. Flag any entries missing evidence or alternatives -- these are lower-confidence records where the trace was incomplete.

GATE: Decisions parsed and filtered. At least one decision entry available. Proceed only when gate passes.

Phase 3: PRESENT

Goal: Format trace data as a human-readable decision timeline.

Step 1: Build the timeline

For each decision entry, format as:

[TIMESTAMP] TYPE
  Decision: CHOSEN
  Alternatives: ALTERNATIVES (or "none recorded")
  Evidence: EVIDENCE
  Context: CONTEXT

Order chronologically. Group consecutive entries of the same type under a shared heading if there are more than 5 entries total, to prevent wall-of-text for long sessions.

Step 2: Highlight the answer to the user's question

If the user asked "why did you choose X?", lead with the specific decision entry that answers their question, then show surrounding context:

You asked: "Why did you choose the code reviewer?"

Decision found at [TIMESTAMP]:
  Chosen: reviewer-code agent
  Alternatives: reviewer-domain, reviewer-perspectives
  Evidence: Request matched "review this function" trigger; code-specific
            keywords ("function", "bug") scored highest for reviewer-code
  Context: User said "review this function for bugs"

--- Full session timeline (3 decisions) ---
[... remaining entries ...]

Step 3: Flag gaps honestly

If the trace has gaps (missing timestamps, empty evidence fields, decisions without alternatives), say so explicitly:

Note: [N] decision(s) have incomplete evidence fields. These entries
show WHAT was decided but not WHY -- the recording hook may not have
captured full context for these decisions.

Do not fill gaps with speculation. Incomplete data presented honestly is more useful than complete-looking data that includes fabrication.

GATE: Timeline presented. User's question answered from trace data. Done.

Examples

Example 1: General session review

User says: "Show me the decision log"

skill: explanation-traces

Actions:

1. Locate session-trace.json (Phase 1)
2. Parse all decision entries (Phase 2)
3. Present full chronological timeline (Phase 3) Result: Complete session decision history with evidence for each choice

Example 2: Specific agent question

User says: "Why did you pick that agent?"

skill: explanation-traces "why that agent?"

Actions:

1. Locate session-trace.json (Phase 1)
2. Filter to agent-selection entries (Phase 2)
3. Present the most recent agent selection with evidence, then full timeline for context (Phase 3) Result: Evidence-backed explanation of agent choice, not a post-hoc rationalization

Example 3: Gate failure investigation

User says: "Why did the gate fail?"

skill: explanation-traces "gate failure"

Actions:

1. Locate session-trace.json (Phase 1)
2. Filter to gate-verdict entries, especially failures (Phase 2)
3. Present gate verdicts with the evidence that caused pass/fail (Phase 3) Result: Specific gate failure reason from the trace, with what was expected vs. what was found

Anti-Patterns

Anti-Pattern 1: Post-Hoc Rationalization

Wrong: Reconstructing "why" from memory when the trace file is missing. Right: If no trace file exists, say so and explain how to enable tracing. Never fabricate an explanation.

Anti-Pattern 2: Paraphrasing Away Evidence

Wrong: "The router probably picked that agent because it seemed relevant." Right: Quote the exact evidence field: "Trigger 'review this function' matched reviewer-code with score 0.92."

Anti-Pattern 3: Dumping Raw JSON

Wrong: Printing the entire session-trace.json as-is. Right: Parse, filter to the user's question, and format as a readable timeline with clear labels.

Anti-Pattern 4: Filling Gaps with Inference

Wrong: "The alternatives field is empty, but it likely considered agents X and Y." Right: "No alternatives were recorded for this decision -- the trace is incomplete at this point."

Anti-Pattern 5: Ignoring the User's Specific Question

Wrong: Always showing the full timeline regardless of what the user asked. Right: Lead with the specific answer, then offer full context as supplementary detail.

Error Handling

Error: No Trace File Found

Cause: Tracing hook not enabled or session-trace.json was cleaned up. Solution: Inform user that no trace exists. Point to skills/explanation-traces/references/trace-schema.md for the schema a tracing hook should produce. Do not reconstruct decisions from conversation history.

Error: Trace File Is Malformed JSON

Cause: Partial write, race condition, or manual corruption. Solution: Report the parse error with the specific line/character offset. Attempt to extract any valid decision entries before the corruption point. Flag that the trace is incomplete.

Error: Trace File Has No Decision Entries

Cause: Hook is writing the file but not recording decisions (e.g., only session metadata). Solution: Report that the trace exists but contains zero decision entries. Check whether the decisions array is present but empty vs. missing entirely, and report which case applies.

Error: User Asks About a Decision Not in the Trace

Cause: The specific decision the user is asking about was not recorded. Solution: Show what IS in the trace and explain that the requested decision type was not captured. Suggest which hook event type would need to emit that decision.

References

Reference Files

• ${CLAUDE_SKILL_DIR}/references/trace-schema.md: JSON schema for session-trace.json with field descriptions and example entries