From skeft-qa
Interactively review the System-2 dev-process harvest — promote agent-reviewed findings to human-reviewed (or close them), and triage GAP-A gate proposals. Use when you want to sign off on what the autonomous loop learned about HOW it ran.
How this skill is triggered — by the user, by Claude, or both
Slash command
/skeft-qa:debriefThis skill is limited to the following tools:
The summary Claude sees in its skill listing — used to decide when to auto-load this skill
<!-- INVOCABILITY (spec 11): `disable-model-invocation: true` is the deliberate USER-ONLY
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").
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.
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).
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:
human-reviewed (the lesson is confirmed),status: closed, with a one-line evidence of why it's resolved —
fixed, superseded, or promoted to a System-1 gate),agent-reviewed (not yet confirmed).Apply the user's choice via the register CLI (NOT by hand-editing the markdown — let the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 (its full JSON record
from the register) with the user's decision applied ("tier": "human-reviewed" to promote, or
"status": "closed" + "evidence": "..." to close):
cd "<repo>" && echo '<F-json>' | uv run python scripts/system2_register.py --upsert
(upsert raises tier monotonically — agent-reviewed → human-reviewed is allowed; it never
downgrades.) A closed finding drops out of the register-wide active-issues view on the
next harvest refresh.
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.
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.
npx claudepluginhub netrxn/sk_eft_hawking --plugin skeft-qaProvides a checklist for code reviews covering functionality, security, performance, maintainability, tests, and quality. Use for pull requests, audits, team standards, and developer training.