fix(review): resolve dead aiReviewCloseConfidence floor with a configurable disposition

.gittensory.yml.example

@@ -344,12 +344,22 @@ gate:
     # String or null. Default: null (the key record's model, else a conservative
     # per-provider default).
     model: null
-    # Minimum calibrated AI-reviewer confidence (0-1) recorded for cache and
-    # calibration context. Under `mode: block`, consensus and split AI-review
-    # defects still block regardless of this floor. Number 0–1, or null.
-    # Default: null (engine uses 0.93). Config-as-code only — no DB column or
-    # dashboard toggle; this can only be set here.
+    # Minimum calibrated AI-reviewer confidence (0-1). Under `mode: block`,
+    # consensus and split AI-review defects still BLOCK the gate regardless of
+    # this floor — what varies below it is `lowConfidenceDisposition` below.
+    # Number 0-1, or null. Default: null (engine uses 0.93). Config-as-code
+    # only — no DB column or dashboard toggle; this can only be set here.
     closeConfidence: null
+    # Disposition for a sub-closeConfidence-floor consensus/split defect (#4603).
+    #   one_shot         — ignore the floor; always one-shot-close (today's
+    #                      pre-#4603 behavior). Opt-in only.
+    #   hold_for_review  — DEFAULT. Still blocks the merge, but holds for a
+    #                      maintainer instead of one-shot-closing.
+    #   advisory_only    — drop to a fully non-blocking advisory below the floor.
+    # one_shot | hold_for_review | advisory_only, or null. Default: null (engine
+    # uses hold_for_review). DB-backed (dashboard-settable too, via the
+    # maintainer AI-review settings); this overrides the stored value.
+    lowConfidenceDisposition: null
     # Per-repo override of the self-host operator's dual-AI combine strategy (#2567).
     # single | consensus | synthesis, or null. Default: null (the operator's own
     # AI_REVIEW_PLAN.combine env default, itself "consensus" if unset). A refinement

apps/gittensory-ui/public/openapi.json

@@ -9554,6 +9554,16 @@
           },
           "agentGlobalFreezeOverride": {
             "type": "boolean"
+          },
+          "aiReviewLowConfidenceDisposition": {
+            "type": "string",
+            "nullable": true,
+            "enum": [
+              "one_shot",
+              "hold_for_review",
+              "advisory_only",
+              null
+            ]
           }
         },
         "required": [

config/examples/gittensory.full.yml

@@ -357,12 +357,22 @@ gate:
     # String or null. Default: null (the key record's model, else a conservative
     # per-provider default).
     model: null
-    # Minimum calibrated AI-reviewer confidence (0-1) recorded for cache and
-    # calibration context. Under `mode: block`, consensus and split AI-review
-    # defects still block regardless of this floor. Number 0–1, or null.
-    # Default: null (engine uses 0.93). Config-as-code only — no DB column or
-    # dashboard toggle; this can only be set here.
+    # Minimum calibrated AI-reviewer confidence (0-1). Under `mode: block`,
+    # consensus and split AI-review defects still BLOCK the gate regardless of
+    # this floor — what varies below it is `lowConfidenceDisposition` below.
+    # Number 0-1, or null. Default: null (engine uses 0.93). Config-as-code
+    # only — no DB column or dashboard toggle; this can only be set here.
     closeConfidence: null
+    # Disposition for a sub-closeConfidence-floor consensus/split defect (#4603).
+    #   one_shot         — ignore the floor; always one-shot-close (today's
+    #                      pre-#4603 behavior). Opt-in only.
+    #   hold_for_review  — DEFAULT. Still blocks the merge, but holds for a
+    #                      maintainer instead of one-shot-closing.
+    #   advisory_only    — drop to a fully non-blocking advisory below the floor.
+    # one_shot | hold_for_review | advisory_only, or null. Default: null (engine
+    # uses hold_for_review). DB-backed (dashboard-settable too, via the
+    # maintainer AI-review settings); this overrides the stored value.
+    lowConfidenceDisposition: null
     # Per-repo override of the self-host operator's dual-AI combine strategy (#2567).
     # single | consensus | synthesis, or null. Default: null (the operator's own
     # AI_REVIEW_PLAN.combine env default, itself "consensus" if unset). A refinement

migrations/0140_ai_review_low_confidence_disposition.sql

@@ -0,0 +1,8 @@
+-- AI-review low-confidence disposition (#4603, resolving the dead aiReviewCloseConfidence floor audit
+-- finding). Governs what happens when an ai_consensus_defect/ai_review_split finding's confidence is BELOW
+-- the configured aiReviewCloseConfidence floor: 'hold_for_review' (default -- flips the undocumented
+-- unconditional-close drift from commit 311b7613d/#1781 back to a safe default) routes the would-be close
+-- through the existing held-for-manual-review mechanism instead of one-shot-closing; 'one_shot' keeps
+-- today's unconditional-close behavior (opt-in); 'advisory_only' drops a sub-floor finding to fully
+-- non-blocking. See src/rules/advisory.ts's isConfiguredGateBlocker and gittensory-gate-setting-wiring.
+ALTER TABLE repository_settings ADD COLUMN ai_review_low_confidence_disposition TEXT NOT NULL DEFAULT 'hold_for_review';

packages/gittensory-engine/src/advisory/gate-advisory.ts

@@ -40,10 +40,17 @@ export type GateCheckPolicy = {
   /** When `block`, a dual-model AI consensus defect (`ai_consensus_defect` finding) becomes a hard
    *  blocker. Defaults to advisory — AI never blocks unless the maintainer opts in. */
   aiReviewGateMode?: GateRuleMode | undefined;
-  /** Minimum calibrated confidence (0-1) configured for AI close calibration. AI defect findings still block under
-   *  `aiReviewGateMode: block` even when below this floor; the floor remains configurable context, never a guardrail
-   *  that downgrades a blocker to manual review. `null`/undefined ⇒ the 0.93 default. */
+  /** Minimum calibrated confidence (0-1) configured for AI close calibration. AI defect findings still BLOCK the
+   *  gate under `aiReviewGateMode: block` even when below this floor — the floor never turns a real defect into a
+   *  non-blocker on its own. What varies below the floor is {@link aiReviewLowConfidenceDisposition}. `null`/
+   *  undefined ⇒ the 0.93 default. */
   aiReviewCloseConfidence?: number | null | undefined;
+  /** Disposition for a sub-floor `ai_consensus_defect`/`ai_review_split` finding (#4603) — see the host copy's
+   *  doc comment (`src/rules/advisory.ts` / `src/types.ts`) for the full semantics. `null`/undefined ⇒
+   *  `hold_for_review` (the shipped default). Only `advisory_only` changes what `isConfiguredGateBlocker` returns
+   *  for these codes here; `one_shot`/`hold_for_review` are indistinguishable to this predictor (the
+   *  `hold_for_review` vs `one_shot` difference is a disposition-planner concern this predictor doesn't model). */
+  aiReviewLowConfidenceDisposition?: "one_shot" | "hold_for_review" | "advisory_only" | undefined;
   readinessScore?: number | null | undefined;
   /** When `block`, the deterministic slop score becomes a hard blocker once `slopRisk >= slopGateMinScore`
    *  (default threshold 60, the `high` band). Defaults to off/advisory — slop never blocks unless opted in. */
@@ -542,11 +549,20 @@ function isConfiguredGateBlocker(finding: AdvisoryFinding, policy: GateCheckPoli
   // most conservative AI signal (two independent models) but still confirmed-contributor gated by
   // evaluateGateCheck, and advisory by default.
   // A consensus defect (both reviewers) OR a SPLIT (one reviewer flagged a blocker the other did not) both block
-  // when aiReviewGateMode is `block`. The configured close-confidence floor remains calibration context; it does
-  // not turn a blocker into a manual hold for normal contributors. (#ai-review-split)
+  // when aiReviewGateMode is `block`. (#ai-review-split) The close-confidence floor + disposition (#4603, mirrors
+  // the host copy in src/rules/advisory.ts -- this predictor package doesn't thread aiReviewLowConfidenceDisposition
+  // through predicted-gate.ts's own policy-building call yet, same deliberate partial-wiring precedent as
+  // linkedIssueSatisfactionGateMode, so this branch only ever sees the default `hold_for_review` here today) decide
+  // what happens to a SUB-floor finding: `one_shot`/`hold_for_review` both still block here; only `advisory_only`
+  // demotes a sub-floor finding to a non-blocker.
   if (code === "ai_consensus_defect" || code === "ai_review_split") {
-    void (policy.aiReviewCloseConfidence ?? DEFAULT_AI_REVIEW_CLOSE_CONFIDENCE);
-    return gatePolicyBlocks(policy.aiReviewGateMode, "advisory");
+    if (!gatePolicyBlocks(policy.aiReviewGateMode, "advisory")) return false;
+    if ((policy.aiReviewLowConfidenceDisposition ?? "hold_for_review") === "advisory_only") {
+      const floor = policy.aiReviewCloseConfidence ?? DEFAULT_AI_REVIEW_CLOSE_CONFIDENCE;
+      const confidence = finding.confidence ?? 1;
+      if (confidence < floor) return false;
+    }
+    return true;
   }
   if (code === REVIEW_THREAD_BLOCKER_CODE) return true;
   // A leaked-secret finding (`secret_leak`) ALWAYS hard-blocks: a committed credential must be removed and

packages/gittensory-engine/src/focus-manifest.ts

@@ -10,6 +10,7 @@
 import { parse as parseYaml } from "yaml";
 import type {
   AdvisoryAiRoutingConfig,
+  AiReviewLowConfidenceDisposition,
   CombineStrategy,
   GatePolicyPack,
   GateRuleMode,
@@ -101,6 +102,11 @@ export type FocusManifestGateConfig = {
   /** `gate.aiReview.closeConfidence` (#7): minimum calibrated AI-reviewer confidence (0-1) for an AI defect to BLOCK
    *  under `aiReview.mode: block`. null (unset) ⇒ the gate's 0.93 default. Clamped to [0,1] at parse time. */
   aiReviewCloseConfidence: number | null;
+  /** `gate.aiReview.lowConfidenceDisposition` (#4603): disposition for a sub-`closeConfidence`-floor
+   *  `ai_consensus_defect`/`ai_review_split` finding. null (unset) ⇒ `hold_for_review` (the shipped default).
+   *  DB-backed (dashboard-settable too, via the `/ai-review` route); this overrides the stored value -- mirrors
+   *  `aiReviewMode` above, not the config-as-code-only `closeConfidence` sibling field just above. */
+  aiReviewLowConfidenceDisposition: AiReviewLowConfidenceDisposition | null;
   /** `gate.aiReview.combine` (#2567): per-repo override of the self-host operator's `AI_REVIEW_PLAN.combine`
    *  boot default (single/consensus/synthesis). null (unset) ⇒ the operator's plan (or `consensus`). A
    *  REFINEMENT only — see {@link aiReviewOnMerge} for the operator-floor clamp `runGittensoryAiReview` applies
@@ -911,6 +917,7 @@ const EMPTY_GATE_CONFIG: FocusManifestGateConfig = {
   aiReviewModel: null,
   aiReviewAllAuthors: null,
   aiReviewCloseConfidence: null,
+  aiReviewLowConfidenceDisposition: null,
   aiReviewCombine: null,
   aiReviewOnMerge: null,
   aiReviewReviewers: null,
@@ -1263,6 +1270,12 @@ function parseGateConfig(value: JsonValue | undefined, warnings: string[]): Focu
     aiReviewModel: normalizeOptionalString(aiReviewRecord?.model, "gate.aiReview.model", warnings),
     aiReviewAllAuthors: normalizeOptionalBoolean(aiReviewRecord?.allAuthors, "gate.aiReview.allAuthors", warnings),
     aiReviewCloseConfidence: normalizeOptionalConfidence(aiReviewRecord?.closeConfidence, "gate.aiReview.closeConfidence", warnings),
+    aiReviewLowConfidenceDisposition: normalizeOptionalEnum(
+      aiReviewRecord?.lowConfidenceDisposition,
+      "gate.aiReview.lowConfidenceDisposition",
+      ["one_shot", "hold_for_review", "advisory_only"] as const,
+      warnings,
+    ),
     aiReviewCombine: normalizeOptionalEnum(aiReviewRecord?.combine, "gate.aiReview.combine", ["single", "consensus", "synthesis"] as const, warnings),
     aiReviewOnMerge: normalizeOptionalEnum(aiReviewRecord?.onMerge, "gate.aiReview.onMerge", ["either", "both"] as const, warnings),
     aiReviewReviewers: normalizeOptionalReviewers(aiReviewRecord?.reviewers, "gate.aiReview.reviewers", warnings),
@@ -1309,6 +1322,7 @@ function parseGateConfig(value: JsonValue | undefined, warnings: string[]): Focu
     gate.aiReviewModel !== null ||
     gate.aiReviewAllAuthors !== null ||
     gate.aiReviewCloseConfidence !== null ||
+    gate.aiReviewLowConfidenceDisposition !== null ||
     gate.aiReviewCombine !== null ||
     gate.aiReviewOnMerge !== null ||
     gate.aiReviewReviewers !== null ||
@@ -1365,6 +1379,7 @@ export function gateConfigToJson(gate: FocusManifestGateConfig): JsonValue {
     gate.aiReviewModel !== null ||
     gate.aiReviewAllAuthors !== null ||
     gate.aiReviewCloseConfidence !== null ||
+    gate.aiReviewLowConfidenceDisposition !== null ||
     gate.aiReviewCombine !== null ||
     gate.aiReviewOnMerge !== null ||
     gate.aiReviewReviewers !== null
@@ -1376,6 +1391,7 @@ export function gateConfigToJson(gate: FocusManifestGateConfig): JsonValue {
     if (gate.aiReviewModel !== null) aiReview.model = gate.aiReviewModel;
     if (gate.aiReviewAllAuthors !== null) aiReview.allAuthors = gate.aiReviewAllAuthors;
     if (gate.aiReviewCloseConfidence !== null) aiReview.closeConfidence = gate.aiReviewCloseConfidence;
+    if (gate.aiReviewLowConfidenceDisposition !== null) aiReview.lowConfidenceDisposition = gate.aiReviewLowConfidenceDisposition;
     if (gate.aiReviewCombine !== null) aiReview.combine = gate.aiReviewCombine;
     if (gate.aiReviewOnMerge !== null) aiReview.onMerge = gate.aiReviewOnMerge;
     if (gate.aiReviewReviewers !== null) {

packages/gittensory-engine/src/types/manifest-deps-types.ts

@@ -18,6 +18,10 @@ export type CombineStrategy = "single" | "consensus" | "synthesis";
 
 export type OnMerge = "either" | "both";
 
+// Disposition for a sub-aiReviewCloseConfidence-floor ai_consensus_defect/ai_review_split finding (#4603) --
+// see src/types.ts's mirror of this type (AiReviewLowConfidenceDisposition) for the full semantics of each value.
+export type AiReviewLowConfidenceDisposition = "one_shot" | "hold_for_review" | "advisory_only";
+
 // #4110: `request_changes`/`comment` were REMOVED (see src/types.ts's mirror of this type for why).
 // `"advisory"` (#4535) is a NEW, actually-wired value -- see src/types.ts's mirror for the full rationale.
 export type ScreenshotTableGateAction = "close" | "advisory";

packages/gittensory-engine/test/gate-advisory-ai-review-low-confidence-disposition.test.ts

@@ -0,0 +1,75 @@
+import { test } from "node:test";
+import assert from "node:assert/strict";
+
+import { evaluateGateCheck } from "../dist/advisory/gate-advisory.js";
+import type { Advisory, AdvisoryFinding } from "../dist/types/predicted-gate-types.js";
+
+const DEFAULT_AI_REVIEW_CLOSE_CONFIDENCE = 0.93;
+
+function consensusAdvisory(confidence: number): Advisory {
+  const finding: AdvisoryFinding = {
+    code: "ai_consensus_defect",
+    title: "AI reviewers agree on a likely critical defect",
+    severity: "critical",
+    detail: "Both reviewers flagged the same blocker.",
+    confidence,
+  };
+  return {
+    id: "advisory-1",
+    targetType: "pull_request",
+    targetKey: "JSONbored/gittensory#1",
+    repoFullName: "JSONbored/gittensory",
+    conclusion: "action_required",
+    severity: "critical",
+    title: "Gittensory review",
+    summary: "",
+    findings: [finding],
+    generatedAt: "2026-07-10T00:00:00.000Z",
+  };
+}
+
+// #4603 (gate-decision twin of src/rules/advisory.ts, kept in sync per checkGateDecisionVersionBump): mirrors
+// the host copy's own regression tests for isConfiguredGateBlocker's aiReviewLowConfidenceDisposition branch.
+
+test("at-or-above-floor confidence blocks identically across all three dispositions", () => {
+  const advisory = consensusAdvisory(DEFAULT_AI_REVIEW_CLOSE_CONFIDENCE);
+  for (const disposition of ["one_shot", "hold_for_review", "advisory_only"] as const) {
+    const evaluation = evaluateGateCheck(advisory, { aiReviewGateMode: "block", aiReviewLowConfidenceDisposition: disposition });
+    assert.equal(evaluation.conclusion, "failure");
+  }
+});
+
+test("sub-floor + one_shot or hold_for_review still blocks (today's behavior, unchanged at the gate level)", () => {
+  const advisory = consensusAdvisory(DEFAULT_AI_REVIEW_CLOSE_CONFIDENCE - 0.1);
+  assert.equal(evaluateGateCheck(advisory, { aiReviewGateMode: "block", aiReviewLowConfidenceDisposition: "one_shot" }).conclusion, "failure");
+  assert.equal(evaluateGateCheck(advisory, { aiReviewGateMode: "block", aiReviewLowConfidenceDisposition: "hold_for_review" }).conclusion, "failure");
+  // Unset ⇒ hold_for_review is the default.
+  assert.equal(evaluateGateCheck(advisory, { aiReviewGateMode: "block" }).conclusion, "failure");
+});
+
+test("sub-floor + advisory_only drops the finding to fully non-blocking", () => {
+  const advisory = consensusAdvisory(DEFAULT_AI_REVIEW_CLOSE_CONFIDENCE - 0.1);
+  const evaluation = evaluateGateCheck(advisory, { aiReviewGateMode: "block", aiReviewLowConfidenceDisposition: "advisory_only" });
+  assert.equal(evaluation.conclusion, "success");
+  assert.equal(evaluation.blockers.length, 0);
+});
+
+test("respects a custom aiReviewCloseConfidence floor under advisory_only", () => {
+  const advisory = consensusAdvisory(0.5);
+  assert.equal(
+    evaluateGateCheck(advisory, { aiReviewGateMode: "block", aiReviewLowConfidenceDisposition: "advisory_only", aiReviewCloseConfidence: 0.6 }).conclusion,
+    "success",
+  );
+  assert.equal(
+    evaluateGateCheck(advisory, { aiReviewGateMode: "block", aiReviewLowConfidenceDisposition: "advisory_only", aiReviewCloseConfidence: 0.4 }).conclusion,
+    "failure",
+  );
+});
+
+test("aiReviewGateMode !== block stays non-blocking regardless of disposition", () => {
+  const advisory = consensusAdvisory(DEFAULT_AI_REVIEW_CLOSE_CONFIDENCE - 0.1);
+  for (const disposition of ["one_shot", "hold_for_review", "advisory_only"] as const) {
+    const evaluation = evaluateGateCheck(advisory, { aiReviewGateMode: "advisory", aiReviewLowConfidenceDisposition: disposition });
+    assert.equal(evaluation.conclusion, "success");
+  }
+});