ad-level-up

Mechanical shape:

THIS SESSION:
  1. State the candidate + its citation (or stop — not grounded).
  2. Trace it to a root cause.
  3. Run the four anti-overfitting gates (all must pass) + the effectiveness pass.
  4. Place it deterministically; draft the minimal edit (do NOT write yet).
  5. Adversarial multi-lens review of the candidate (already-covered / coherence / placement).
  6. PRESENT the proposal + rationale. Apply ONLY on explicit approval, one item at a time.

The whole point is that a full or drifting context never edits the rules unsupervised — so the write is always downstream of an explicit human OK.

<background_information> The companion that evolves the rule-set ad-audit audits against. Curation is a WRITE operation behind a human gate — distinct from the read-only ad-audit. The mechanism (four anti-overfitting gates, effectiveness pass, deterministic placement, adversarial multi-lens review, hard human gate) is ADR-0037; the rule-set locations are ADR-0035 (machine store) and ADR-0043 (project layer at .agentic/rules/, committed or machine-local via .git/info/exclude — a project rule shadows a conflicting machine-store rule, and the audit reports the shadowing). On Claude Code the multi-lens review fans out Task subagents; on Codex it runs inline with an optional user-initiated rule-candidate-reviewer escalation. The skill owns the terse rule-set only — bigger decisions route to ad-adr / ad-guidelines. </background_information>

Running ad-level-up (Codex single-pass, human-gated). I will state the candidate + evidence, trace its root cause, run the four anti-overfitting gates + effectiveness pass, place it, draft the minimal edit, run an adversarial multi-lens review, then PRESENT a proposal. I will NOT write anything to the rule-set until you explicitly approve — one item at a time.

Step 1 — state candidate + evidence. One sentence + the citation (finding / PR / transcript / file:line, or the ad-audit handoff). If it cannot be cited, stop — not grounded.

Step 2 — trace to root cause. Attach the candidate to the upstream cause (investigation / grounding / verification gap), not the surface symptom.

Step 3 — four anti-overfitting gates. All must pass; reject the rest out loud: (a) recurrence or deliberate decision; (b) generalisation (a class of future work); (c) load-bearing root cause; (d) proportionate cost (earns its keep against adherence decay).

Step 4 — effectiveness pass. Classify (improvement / correction / increment / merge / extend / reject — prefer sharpening/merging/extending). Check redundancy (does an existing rule / binding doc already watch this?), real-behaviour (a cited observed defect, not a hypothetical), and dead-rule sweep (flag a subsumed rule for retirement).

Step 5 — deterministic placement. Assign to the group whose grounding-target it matches; keep the group set minimal (new group only when a defect class is uncovered by every existing group; split only when a group grows too large); next stable id.

Step 6 — draft the minimal edit. Pick the layer from the rule's own content (generalizes beyond this repo → machine store at $AGENTIC_RULES_DIR else ~/.agentic/rules/; this-project convention → .agentic/rules/ at the repo root), stating the recommendation for the user to confirm; read the target file(s) first. On first project-rule creation ask committed vs machine-local; machine-local wires .agentic/rules/ into .git/info/exclude (on approval, never .gitignore). When you find an .agentic/rules/ that is neither committed nor excluded (e.g. a fresh clone of a machine-local repo), re-offer the committed-vs-machine-local choice. Draft the terse imperative rule (and its rationale when worth keeping). Do NOT write to disk.

Step 7 — adversarial multi-lens review. Refute the candidate against the actual current rule files along three lenses: (a) already-covered (cite the covering rule → reject/merge), (b) coherence + necessity (minimal change? clears the gates?), (c) placement. Run inline; for a substantive rule, recommend the user spawn the bundled rule-candidate-reviewer subagent once per lens for true isolation. Filter and reclassify; do not write.

Step 8 — report + gate. Present the exact delta (old → new per file, at the resolved path), the gate table, the effectiveness line, the adversarial-review verdicts (citing the covering rule on any reject/merge), a plain-language rationale, and a "considered but rejected as overfitting" list. Apply ONLY on the user's explicit approval, one item at a time; then stop and await the next item.

<output_contract>

• A proposal (delta + gate table + effectiveness line + adversarial verdicts + rationale + rejected-as-overfitting list) — presented, not applied.
• On explicit approval: the minimal edit to the rule-set at the ADR-0035 location, one item at a time. Nothing written without approval. </output_contract>

<optional_escalation> For a substantive candidate, spawn the bundled rule-candidate-reviewer subagent (.codex/agents/rule-candidate-reviewer.toml) once per lens (already-covered / coherence / placement) against the actual rule files, for true fresh-context isolation. Pass the handoff inline in the spawn prompt — the candidate, its evidence, and the resolved rule-set path (a curation candidate is small; no persisted trail file is needed). Union the verdicts into the proposal; the human gate is unchanged. </optional_escalation>