refactor(code-review): excise inert Premise category layer; SCHEMA_VERSION 2 (v3.0.0)

CHANGELOG.md

@@ -4,6 +4,20 @@ All notable changes to the claude-plugins project will be documented in this fil
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/). Entries are listed newest-first; each plugin section is treated as released when merged to `main`.
 
+### code-review v3.0.0
+
+#### Removed
+- Retired the `Premise` finding category and its entire processing layer, now inert since the premise reviewer was removed (no producer emits `category: "Premise"`). Dropped `Premise` from the `CATEGORIES` schema enum; removed the two verdict-precedence rules that gated on Premise findings (the Premise priority-0 → `CHANGES_REQUESTED` rule and the cumulative-Premise → `NEEDS_ATTENTION` rule) along with the `_count_gateable_premise_medium` helper, the `premise_cumulative_medium` operator threshold, and its `justification_rate_alert` companion; removed the Premise-scoped `stats` telemetry sub-blocks (`by_subcategory`, `justification`, `premise_cumulative_medium_count`); removed the verifier "always verify Premise" eligibility branch and the Premise extra-strictness blocks from `verifier_prompt.txt`; and deleted the four `premise_*.md` fix templates with their `/fix` dispatch rows. The shared/BHA reviewer prompts keep the generic "PREMISE:" reasoning step (what the code is supposed to do), and the general author-justification machinery — the `justified[]` bucket, `JUSTIFIED-VALID`/`JUSTIFIED-INVALID` verdicts, the verifier's J1/J2 audit, and the justification-audit learning stream — is retained; only the Premise-specific pieces were removed.
+- Removed the four now-unreachable Premise reasoning-certificate kinds (`necessity`, `cohesion`, `workaround`, `complexity`) from `REASONING_CERTIFICATE_KINDS`; only the active reviewer/reasoning-step kinds (`impact`, `test_quality`, `sibling_pattern`, `bha`, `bhb`, `auditor`) remain.
+
+#### Changed
+- Bumped `SCHEMA_VERSION` 1 → 2 for the `Finding` + `ResultEnvelope` contract to reflect the removed `Premise` category and telemetry keys. Because the schema version is folded into the prompt/cache hash, this invalidates the Bug Hunter A and verification caches once on rollout. `verdict-thresholds.json` now exposes only `impact_cumulative` (the FEA-1401 cumulative Impact gate, PLN-726 OQ#6), and `stats` retains `impact_cumulative_count` in place of the removed premise count.
+- Pinned `effort: high` in the `code-review-worker` and `code-review-worker-graph` subagent definitions so the reviewer fleet's reasoning depth no longer drops when `/code-review` is run from a lower session effort level. Reasoning effort, unlike `model`, has no per-invocation override — a spawned reviewer otherwise inherits the session effort — so pinning it in frontmatter holds every reviewer at `high` regardless of the session level (`high` is valid on both Opus and Sonnet, the route-assigned reviewer models). `start.md` and `README.md` now document the effort axis alongside the existing session-model cost guidance.
+
+#### Fixed
+- Corrected documentation and a test left stale by the `Premise` removal and the renumbered verdict-precedence list. The cumulative Impact verdict gate is now referenced by name (FEA-1401 / PLN-726 OQ#6) in `start.md` and `README.md` rather than a bare ordinal that collided with `SCHEMA.md`'s sequential rules, and a numbering note in `SCHEMA.md` records that `_compute_canonical_verdict`'s plan-derived rule labels do not map 1:1 to that list. The `SCHEMA.md` deferred-fixture count now matches `_DEFERRED_FIXTURES` (3 deferred; `golden_injection_quarantine` is listed as a shipped fixture). The schema test that exercised the removed `stats.justification` / `stats.by_subcategory` telemetry sub-blocks now targets the live `stats.verification` sub-block (`justified_valid_count` / `justified_invalid_count` / `by_reviewer`).
+- Restored the operator-facing Impact gate count in GitHub-mode output: the `github-review.md` Verifier Stats block now shows `Impact gateable count` (`stats.impact_cumulative_count`, gate threshold `impact_cumulative`) to match the local-mode presenter — the removed Premise cumulative-gate display had left GitHub mode with no line for the envelope's sole operator-tunable verdict-gate count. Also clarified the `golden_schema_v1_round_trip` fixture description to reflect that it round-trips a `schema_version: 1` finding through the current `schema_version: 2` envelope (a v1-finding backward-compat probe), instead of implying a stale v1-only test.
+
 ### code-review v2.37.1
 
 #### Fixed

plugins/code-review/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "code-review",
   "description": "Code review plugin",
-  "version": "2.37.1",
+  "version": "3.0.0",
   "author": {
     "name": "ClosedLoop",
     "email": "support@closedloop.ai"

plugins/code-review/README.md

@@ -7,7 +7,7 @@ A multi-agent code review plugin for Claude Code that performs deep, partitioned
 - **Multi-agent parallel review**: Splits changed files into partitions and spawns concurrent reviewer agents (Bug Hunter A, plus domain specialists) to review each partition independently
 - **Deterministic hygiene checks**: Pattern-based checks for CI artifacts, sensitive file exposure, and path leakage — zero LLM tokens required
 - **Risk-based model routing**: Scores each file partition by risk (size, file type, LOC) and routes high-risk partitions to more capable models
-- **Cost-optimized orchestration**: The orchestrator walk is mechanical (run helper, read JSON, honor gates) and carries no diff or large artifacts in its own context, so it is cheap to run on a lower-cost session model while spawned reviewer and verifier subagents keep their own route-assigned models — run `/code-review` from a standard-context Sonnet session for the cheapest orchestrator without changing review quality
+- **Cost-optimized orchestration**: The orchestrator walk is mechanical (run helper, read JSON, honor gates) and carries no diff or large artifacts in its own context, so it is cheap to run on a lower-cost session model while spawned reviewer and verifier subagents keep their own route-assigned models — run `/code-review` from a standard-context Sonnet session for the cheapest orchestrator without changing review quality (the worker subagents pin `effort: high`, so lowering the session effort too won't reduce reviewer reasoning depth)
 - **Finding validation and deduplication**: Normalizes severity, filters low-confidence findings, deduplicates near-duplicate issues via Jaccard similarity, and validates line numbers against the actual diff
 - **Incremental reviews**: Tracks prior review state to diff only new commits since the last successful review (auto-incremental mode)
 - **Caching**: Content-addressed cache keyed on prompt hash and diff tip to skip re-reviewing unchanged partitions
@@ -260,21 +260,19 @@ Each finding includes: file path, line number, severity, category, issue title,
 
 Operator-tunable knobs live under `.closedloop-ai/settings/`. All files are optional; absent or malformed entries fall back to built-in defaults.
 
-### `verdict-thresholds.json` (PLN-721, PLN-773)
+### `verdict-thresholds.json` (FEA-1401)
 
-Tunes the verdict-precedence gates and the operator-facing telemetry alerts:
+Tunes the verdict-precedence gates:
 
 ```json
 {
-  "premise_cumulative_medium": 3,
-  "justification_rate_alert": 0.30
+  "impact_cumulative": 2
 }
 ```
 
 | Key | Default | Effect |
 |---|---|---|
-| `premise_cumulative_medium` | `3` | Trigger `NEEDS_ATTENTION` when at least N MEDIUM Premise findings survive verification on the same PR, even if no individual finding is HIGH. Set to a very large number (e.g. `999`) to disable. Values below 1 are ignored. |
-| `justification_rate_alert` | `0.30` | Threshold above which `stats.justification.threshold_alert` flips to `true` and the Verifier Stats footer flags the run. PLN-721 §Telemetry: "if > ~30%, authors likely gaming the hatch." Footer-only alert (does not modify the verdict). Set to `1.0` to disable. Values outside `[0.0, 1.0]` are ignored. |
+| `impact_cumulative` | `2` | Trigger `NEEDS_ATTENTION` when at least N BLOCKING/HIGH `ImpactAnalysis` findings survive verification on the same PR (the cumulative Impact gate — SCHEMA.md §5 verdict-precedence Rule 4 / FEA-1401 OQ#6), even if no single finding would gate on its own. Set to a very large number (e.g. `999`) to disable. Values below 1 are ignored. |
 
 ### `verification-gates.json` (PLN-722)
 
@@ -307,8 +305,7 @@ The presenter (local mode `start.md`, GitHub mode `code-review-verifier-stats.md
 
 - Per-reviewer FP rate (`stats.verification.by_reviewer[*].fp_rate`)
 - Override count per reviewer (`stats.verification.by_reviewer[*].re_asserted`)
-- Premise justification rate + rejection rate (`stats.justification.*`)
-- Premise findings partitioned by subcategory (`stats.by_subcategory`)
+- Justified-finding counts (`stats.verification.justified_valid_count` / `justified_invalid_count`)
 
 `pending-learnings/premise-justifications.jsonl` and `pending-learnings/verifier-overrides.jsonl` feed `self-learning:process-learnings` so the verifier's J2 (responsiveness) threshold and the per-reviewer FP-rate gate can tune over time. Both jsonl writers serialize via `fcntl.flock` so concurrent runs each get exactly one well-formed line per event.
 

plugins/code-review/SCHEMA.md

@@ -37,7 +37,7 @@ shape. Producers may emit dicts directly; the Python convenience type lives in
   "system_marker": "<from the canonical enum (Section 3); null when finding_scope == 'diff'>",
 
   // ── Classification ────────────────────────────────────────
-  "category": "Correctness | Code Quality | Documentation | Hygiene | Repo Hygiene | Premise | ImpactAnalysis | TestQuality | Coverage | InjectionAttempt | CompanionChange | Security",
+  "category": "Correctness | Code Quality | Documentation | Hygiene | Repo Hygiene | ImpactAnalysis | TestQuality | Coverage | InjectionAttempt | CompanionChange | Security",
   "subcategory": "<category-specific; nullable>",
 
   // ── Severity ──────────────────────────────────────────────
@@ -181,7 +181,7 @@ The terminal artifact of every review run.
   // ── Stats ─────────────────────────────────────────────────
   "stats": {
     "by_severity": {"BLOCKING": <int>, "HIGH": <int>, "MEDIUM": <int>},
-    "by_category": {"Correctness": <int>, "Premise": <int>, ...},
+    "by_category": {"Correctness": <int>, ...},
     "by_reviewer": {"<reviewer>": {"verified": <int>, "rejected": <int>, "tentative": <int>, "justified": <int>}},
     "by_finding_scope": {"diff": <int>, "system": <int>, "pr_metadata": <int>},
     "verification": {
@@ -194,7 +194,7 @@ The terminal artifact of every review run.
       "skipped_count": <int>,
       "false_positive_rate": <float>
     },
-    "premise_cumulative_medium_count": <int>,
+    "impact_cumulative_count": <int>,
     "agent_failures": [{"agent_id": "<id>", "reason": "<string>"}]
   },
 
@@ -270,13 +270,14 @@ The `verdict` subcommand applies these rules in order; the first match wins.
 1. Any coverage gap with `required: true` (foundation) → **CHANGES_REQUESTED**
 2. Any BLOCKING finding (verified or system-scoped) → **CHANGES_REQUESTED**
 3. Any HIGH finding (verified or system-scoped) → **NEEDS_ATTENTION**
-4. ≥ N MEDIUM Premise findings (plan 02; default N=3) → **NEEDS_ATTENTION**
-5. ≥ M BLOCKING/HIGH Impact Analysis findings (plan 06; default M=2) → **NEEDS_ATTENTION**
-6. Any TENTATIVE finding (plan 03) → **NEEDS_ATTENTION**
-7. Otherwise → **APPROVED**
+4. ≥ M BLOCKING/HIGH Impact Analysis findings (plan 06; default M=2) → **NEEDS_ATTENTION**
+5. Any TENTATIVE finding (plan 03) → **NEEDS_ATTENTION**
+6. Otherwise → **APPROVED**
 
 `verdict_reason` cites the specific finding(s) that produced the verdict.
 
+**Numbering note.** This list is a clean sequential summary. The implementation (`_compute_canonical_verdict`) carries plan-derived rule labels — including the `2.5` (mandatory-human-review short-circuit) and `3.5` (TENTATIVE fall-through) sub-rules, and the cumulative Impact gate labeled **Rule 6** (PLN-726 OQ#6) — so the code's labels do not map 1:1 to the numbers above (e.g. the Impact gate is item 4 here but "Rule 6" in code).
+
 The verdict subcommand writes `<CR_DIR>/verdict.json` with both the canonical verdict and a `verdict` string compatible with `run-loop.sh` (which keys on the legacy form):
 - APPROVED → approve
 - NEEDS_ATTENTION → needs_attention
@@ -648,12 +649,12 @@ the same normalization path the assertion uses, so a subsequent
 no-flag run sees byte-identical output. Updates are reviewed in the
 commit diff, not auto-merged.
 
-**Phase 8 ships 3 fixtures end-to-end**
+**Phase 8 shipped 3 fixtures end-to-end**
 (`golden_minimal_correctness`, `golden_all_categories`,
 `golden_schema_v1_round_trip`) plus a byte-identical determinism test
-for `prepare-run`. The 6 fixtures requiring plans 01/02/03/05/06
-(`golden_premise_*`, `golden_impact_with_callsites`,
-`golden_coverage_gap`, `golden_injection_quarantine`,
+for `prepare-run`; PLN-720 promoted a 4th, `golden_injection_quarantine`.
+The remaining 3 fixtures requiring plans 03/05/06
+(`golden_impact_with_callsites`, `golden_coverage_gap`,
 `golden_budget_exceeded`) have reserved directories with READMEs and
 are skipped via a `_DEFERRED_FIXTURES` map in the test module until
 their dependent plans land. Phase 4b will extend the harness to walk

plugins/code-review/agents/code-review-worker-graph.md

@@ -2,6 +2,7 @@
 name: code-review-worker-graph
 description: Graph-aware code review worker for the cross-file reviewers (Impact Analyzer, Bug Hunter B, fast-path). Identical to code-review-worker but adds read-only codebase-memory-mcp tools for precise cross-file usage discovery. Use only for reviewers whose role prompt loads the codebase knowledge graph protocol.
 tools: Read, Write, Grep, Glob, mcp__codebase-memory-mcp__search_graph, mcp__codebase-memory-mcp__trace_path, mcp__codebase-memory-mcp__get_code_snippet, mcp__codebase-memory-mcp__search_code
+effort: high  # pinned so a lowered session effort can't cut reviewer reasoning depth (no per-Task override; frontmatter is the only lever). Not redundant with the default — do not remove. Rationale: start.md "Orchestrator model (cost)".
 ---
 
 # Code Review Worker (graph-aware)

plugins/code-review/agents/code-review-worker.md

@@ -2,6 +2,7 @@
 name: code-review-worker
 description: Worker agent for partitioned code review tasks. Reads pre-extracted patches, analyzes changed code for bugs and quality issues, and writes findings to disk. Use when spawning background review agents that need file access.
 tools: Read, Write, Grep, Glob
+effort: high  # pinned so a lowered session effort can't cut reviewer reasoning depth (no per-Task override; frontmatter is the only lever). Not redundant with the default — do not remove. Rationale: start.md "Orchestrator model (cost)".
 ---
 
 # Code Review Worker