TraceDecay Dev: Introspecting TraceDecay Usage

Use this skill for TraceDecay self-improvement driven by real agent behavior, not guesswork. Keep the pass evidence-first: logs and summaries identify the failure mode, then code, tests, or bundled skills encode the fix.

Required Audit Lanes

1. Adoption and telemetry: Use tracedecay:diagnosing-analytics. Run tracedecay analytics diagnostics (add --all when the question is cross-project). Compare hook/tool volume with TraceDecay tool usage.
2. Past-session behavior: Use tracedecay:managing-session-context. Start with tracedecay_message_search; use tracedecay_lcm_status before deeper LCM work, then tracedecay_lcm_grep or replay only as needed.
3. Managed skills and automation: Use tracedecay:inspecting-managed-skills. Start with tracedecay_skill_list; view skills or run artifacts only when the list points to a specific stale, patched, failed, or unused item.
4. Durable memory: Use tracedecay:project-memory. Search with tracedecay_fact_store before re-deriving project decisions. For curation, run read-only inventory and dry-runs first; never remove facts without explicit approval.
5. Code-health and implementation target: Use tracedecay:code-health. Let weak dimensions and hotspots decide whether the fix is code, tests, docs, evals, or skill guidance.

Interpretation Rules

• High hook volume with low TraceDecay tool usage is an adoption gap. Improve trigger text, tool hints, or eval coverage before blaming users.
• Managed skills with many patches and zero successful uses are validation failures. Inspect their evidence and either tighten the managed-skill loop or add a bundled operator skill that prevents the repeated mistake.
• LCM depth and compression ratios are health signals, not direct quality scores. Use raw/session recall to confirm what agents actually missed.
• Memory curation candidates are proposals, not permission. Prefer merge or update when provenance matters; deletion needs fresh user approval.
• Analytics fields can be global. project_id: null means the run was global, not broken.

Improvement Loop

1. Capture counts and examples from every required audit lane.
2. Group evidence into one failure class: discovery, fallback, diagnostics interpretation, automation validation, memory curation, or code structure.
3. Choose the smallest durable fix:
   • bundled skill update when agents know the feature but skip the workflow;
   • eval/test when the workflow should be enforced;
   • code change when the data is unavailable, misleading, or too hard to interpret;
   • managed-skill action when the issue lives only in the profile store.
4. Verify with the narrowest relevant command, then run the matching skill contract or code test.

Helper script

Run scripts/project-analytics.sh as the fast first pass over Lane 1 (adoption) and Lane 4 (durable memory). It prints what tracedecay analytics diagnostics does not: a per-tool mcp_tool_call breakdown with error counts, and fact-store adoption — how many times facts were retrieved/accessed ("seen") versus rated helpful/unhelpful, the seen:feedback ratio, and the transport-agnostic feedback ledger. It prefers the CLI/tools and drops to SQL only for those gaps, resolving store paths from tracedecay tool storage_status. Add --all for a cross-project breakdown.

Deliverable

Report the exact commands/tools used, headline counts, cited sessions or run ids when available, the selected failure class, the improvement made, and the verification result. Include any tracedecay_metrics: savings line.