# Agent Learning Report — YYYY-MM-DD

**Repo:** `<absolute path or repo name>`
**Personal repo:** `<absolute path>`
**Modes run:** baseline-repo, distill-sessions, map-capability
**Session window:** last 30 days (or specify)
**Sessions sampled:** N of M (note sampling if not all)

---

## Summary

One paragraph. What did this run produce that is new vs. the prior report (if any)?
What is the most important gate added or removed?

---

## confirmed_current

Verified this run. Each line includes a source path or command.

- [purpose] <one factual sentence>. Source: <file>:<line>
- [entrypoint] <command and what it does>. Source: <file>:<line>
- [validation] <validation gate command>. Source: <file>:<line>
- [gotcha] <one quoted heading, ≤25 words>. Source: <file>:<line>

---

## memory_derived

From prior memory, distilled sessions, or earlier reports. Not re-verified this run.

- [<domain>] <factual sentence>. Origin: <session-ref or prior-report-date>
- [<domain>] <observation grounded in count>. Origin: distill-sessions, evidence: N of M sessions over <window>

---

## needs_verification

Plausible but not confirmed. Each item includes the exact verification action.

- [<area>] <claim that needs checking>. Verify: <command or file path>
- [<area>] <claim>. Verify: <command>

---

## agent_compensation

The actionable output. Per-domain gates for what future agents MUST do before recommending action in that domain. Read by any agent (Claude, Codex, Hermes, etc.).

### domain: <name>

- **level:** <0–4, or range>
- **evidence_summary:** <2–3 sentences>
- **gates:**
  - category: <docs-check | repo-gate | live-check | evidence-quote | handoff-check>
    gate: <specific, verifiable, cheap action>
  - category: <…>
    gate: <…>

### domain: <name>

- **level:** <…>
- **evidence_summary:** <…>
- **gates:**
  - …

---

## self_healing_loop

- failure_signal -> candidate_gate -> validation_status -> next_session_load. Source: <corpus or baseline>
- strongest_current_gate: <domain> requires `<gate-category>` before action. Source: <corpus or baseline>

---

## skill_inventory

- [skill_inventory] available_skills: <N>. source: <skill-map>
- [skill_inventory] invalid_skills: <N>. source: <skill-map>

---

## skill_usage

- [skill_usage] expected: <skill names or none>. count: <N>. source: <skill-usage>
- [skill_usage] loaded: <skill names or none>. count: <N>. source: <skill-usage>
- [skill_usage] applied: <skill names or none>. count: <N>. source: <skill-usage>

---

## skill_health

- [skill_health] missed_expected_skill: <skill>. source: <skill-usage>; count: <N>; verify: rerun evaluate_skill_routing.py for the matching scope.

---

## skill_compensation

- [skill_compensation] category: candidate_skill_adjustment; skill: <skill>; signal: <correlation-only signal>; gate: <specific future-agent action>. source: <skill-impact>; count: <N>; verify: rerun evaluate_skill_impact.py after the next session batch.

---

## proposed_evergreen_diffs

*(Optional — omit section entirely if nothing to propose.)*

The skill does NOT auto-apply changes to soul.md / system.md / preferences.md / etc. These are suggestions only. The user decides whether to apply.

### file: <path>

- **before:** <current text>
- **after:** <proposed text>
- **rationale:** <one sentence; references at least one agent_compensation gate above>

---

## deferred_phases

*(Optional — omit if no change since last report.)*

Single paragraph noting which deferred phases (write-personal, brief-agent) the user might want next, and why.

---

## Report metadata

- Generated by: agent-learning-compounder skill, MVP
- Generated at: <ISO timestamp>
- Prior report referenced: <YYYY-MM-DD.md filename or "none">
- Length: <line count> lines
