debrief

name	debrief
description	Interactively review the System-2 dev-process harvest — promote agent-reviewed findings to human-reviewed (or close them), and triage GAP-A gate proposals. Runs over the register-aware harvest's already-synthesized register (it files/combines/re-opens continuously); debrief is the human governor — only it promotes to human-reviewed. Use when you want to sign off on what the autonomous loop learned about HOW it ran.
disable-model-invocation	true
allowed-tools	Bash(cd *), Bash(uv run python *), Read

Interactively promote System-2 findings to human-reviewed and triage GAP-A proposals. This is the periodic human governor step — never auto-run; never auto-edit CLAUDE.md / hooks / roadmaps (those need their own explicit sign-off — spec 1 principle 6, "self-improving, never self-mutating").

1. Resolve the repo root (cwd-robust — works from the workspace root OR inside the repo): !`R=$(uv run --no-sync python "${CLAUDE_PLUGIN_ROOT}/scripts/harness_common_cli.py" repo-root 2>/dev/null); test -n "$R" || R=$(git rev-parse --show-toplevel 2>/dev/null); echo "${R:-UNRESOLVED}"` (harness repo_root(), the same resolver the hooks use, with a git rev-parse fallback). UNRESOLVED only if launched entirely outside the workspace — then ask the user to cd into SK_EFT_Hawking/ and re-run. Use <repo> below.

2. Read the register + proposals. Read <repo>/docs/dev-loops/SYSTEM2_REGISTER.md (the ## Open section) and any drafts under <repo>/docs/dev-loops/proposals/. Cluster the agent-reviewed findings (use the per-occurrence goal_id goal-pointer to group by origin — goal-authoring vs other-harness-component vs tactical friction).

3. Present one cluster at a time to the user with: the title, why, how-to-apply, tally (distinct compact-events), and goal-origin. For each, ask the user to choose:

   • promote → human-reviewed (the lesson is confirmed). For an OPEN ISSUE this records the confirmation. For a process win, promotion triggers the win LIFECYCLE: promote → integrate the best practice into the harness (a CLAUDE.md line / the relevant skill / context bootstrap — wins are NOT injected via the active-issues view, so the integration IS the win's in-loop visibility) → then close it (it now lives in the harness). Propose the exact edit + target and apply it on the user's OK (the per-win sign-off principle 6 requires); record the integration target in evidence.
   • close → resolved (status: closed, one-line evidence of why — fixed, superseded, integrated into the harness, or promoted to a System-1 gate),
   • misfile → status: misfiled (it was never a real finding — harvest noise that slipped through). The harvest no longer writes misfiled, so this human sweep is the ONLY path into ## Misfiled.
   • leave → keep agent-reviewed (not yet confirmed).

4. Apply the user's choice via the register CLI (NOT by hand-editing the markdown — let the register own dedup / tier-monotonicity / round-trip). For a finding F with the decision applied:

   • promote (set "tier":"human-reviewed"): cd "<repo>" && echo '<F-json>' | uv run python scripts/system2_register.py --upsert --promote — the --promote flag is REQUIRED. Without it, --upsert deterministically clamps the tier to agent-reviewed (the structural block on harvest self-promotion to human-reviewed — the tier-as-visibility gaming vector). Only /debrief passes --promote.
   • close / misfile / leave: the same --upsert without --promote ("status":"closed" or "misfiled" + "evidence"; leave = no write). A closed/misfiled finding drops out of the register-wide active-issues view on the next refresh.

5. GAP-A proposals. For each proposals/<id>.md, present the proposed structural prevention. On the user's sign-off it becomes a tracked build task (a new validate.py check / goal-prompt reference tweak / automation) — never auto-applied here. Note the disposition; do not implement the gate in this skill.

6. After the pass, refresh the active-issues view so closures take effect immediately: cd "<repo>" && uv run python scripts/system2_register.py --write-active-issues. Summarize what was promoted / closed / left and which proposals were signed off.