Expose the Stride review citation limit before writeback

docs/reference/protocols/goal-vision-replan-contract-v0.md

@@ -640,6 +640,11 @@ exact `replan-decision-*` id and separately records:
 - `goal_gain`: improved, flat, regressed or unknown;
 - delivered evidence references and the bounded lesson for the next experiment.
 
+The authoring contract exposes `review_reference_limit` (3). It applies separately
+to `previous_review.evidence_refs` and `explore_review.considered_refs`.
+
+编写契约的 `review_reference_limit`（3）分别限制上述两个引用列表。
+
 A claim needs evidence references; unavailable evidence remains unknown. These
 are explicit Agent assessments, not host certification of scientific truth.
 The next review retains the assessment with the original decision and subsequent

loopx/control_plane/stride_decision_schema.json

@@ -7,6 +7,7 @@
   "probe_field_limits": {"scope": 120, "prediction": 120, "adoption_criterion": 120, "cost_bound": 100, "minimum_evidence": 120},
   "axis_reason_limit": 96,
   "review_text_limit": 120,
+  "review_reference_limit": 3,
   "adoption_options": ["executed", "partially_executed", "not_executed", "unknown"],
   "information_options": ["new_evidence", "repeated_evidence", "unknown"],
   "gain_options": ["improved", "flat", "regressed", "unknown"],

loopx/control_plane/work_items/replan_context.ts

@@ -8,7 +8,7 @@ import {requireJsonObject, requireNonEmptyString} from "../runtime_decode.ts";
 import {FRESH_PATH_DISPOSITIONS, requiredSemanticOutcomes} from "./replan_semantics.ts";
 import {readReplanSnapshot} from "./replan_history_snapshot.ts";
 import {normalizeGoalPathDelta} from "../goals/vision_checkpoint.ts";
-import {normalizeRecordedStrideDecision, normalizeStrideStrategy, projectStrideExploreBasis} from "./stride_policy.ts";
+import {normalizeRecordedStrideDecision, normalizeStrideStrategy, projectStrideExploreBasis, STRIDE_REVIEW_REFERENCE_LIMIT} from "./stride_policy.ts";
 import {PROGRESS_RESULT_CLASSES} from "./replan_history.ts";
 import {MATERIAL_DELIVERY_OUTCOMES} from "./delivery_outcome.ts";
 
@@ -294,8 +294,8 @@ function reviewEvidence(rows: JsonObject[], request: JsonObject): JsonObject {
     throw new EffectRuntimeRequestError("previous decision changed; refresh replan context before reviewing its outcome");
   }
   const refs = request.evidence_refs;
-  if (!Array.isArray(refs) || refs.length > 3 || new Set(refs).size !== refs.length) {
-    throw new EffectRuntimeRequestError("previous review accepts at most three distinct evidence references");
+  if (!Array.isArray(refs) || refs.length > STRIDE_REVIEW_REFERENCE_LIMIT || new Set(refs).size !== refs.length) {
+    throw new EffectRuntimeRequestError(`previous review accepts at most ${STRIDE_REVIEW_REFERENCE_LIMIT} distinct evidence references`);
   }
   for (const ref of refs) identifier(ref, "review evidence reference");
   const source = objectOrNull(previous!._source_facts)!;

loopx/control_plane/work_items/stride_policy.ts

@@ -39,6 +39,7 @@ export const STRIDE_AXIS_NAMES = decisionSchema.axis_names;
 export const STRIDE_PLAN_LIMIT = decisionSchema.axis_plan_limit;
 export const STRIDE_REASON_LIMIT = decisionSchema.axis_reason_limit;
 export const STRIDE_REVIEW_LIMIT = decisionSchema.review_text_limit;
+export const STRIDE_REVIEW_REFERENCE_LIMIT = decisionSchema.review_reference_limit;
 // Persisted v1 records retain their original meaning; never offered for new reviews.
 export const RECORDED_STRIDE_SCHEMA = "stride_decision_v1";
 const recordedAdjustments = ["grow", "hold", "shrink"] as const;
@@ -52,9 +53,9 @@ function exactFields(value: JsonObject, fields: string[], label: string): void {
   if (Object.keys(value).some(key => !fields.includes(key))) throw new EffectRuntimeRequestError(`${label} contains unsupported fields`);
 }
 function refs(value: unknown, label: string): string[] {
-  if (!Array.isArray(value) || value.length > 3 || value.some(ref => typeof ref !== "string" ||
+  if (!Array.isArray(value) || value.length > STRIDE_REVIEW_REFERENCE_LIMIT || value.some(ref => typeof ref !== "string" ||
       !/^[A-Za-z0-9][A-Za-z0-9._:-]{0,127}$/.test(ref)) || new Set(value).size !== value.length) {
-    throw new EffectRuntimeRequestError(`${label} requires at most three distinct public-safe evidence references`);
+    throw new EffectRuntimeRequestError(`${label} requires at most ${STRIDE_REVIEW_REFERENCE_LIMIT} distinct public-safe evidence references`);
   }
   return value as string[];
 }
@@ -389,6 +390,7 @@ export function strideControlContract(context: JsonObject, obligationId: unknown
     axis_reason_limit: decisionSchema.axis_reason_limit, probe_field_limits: STRIDE_PROBE_LIMITS,
     decision_value_review: {...decisionValueReview},
     review_text_limit: decisionSchema.review_text_limit,
+    review_reference_limit: STRIDE_REVIEW_REFERENCE_LIMIT,
     previous_decision_ref: previous,
     ...(follow ? {decision_followthrough: follow} : {}),
     previous_review_options: {adoption: decisionSchema.adoption_options, information_gain: decisionSchema.information_options, goal_gain: decisionSchema.gain_options},
@@ -406,6 +408,6 @@ export function strideControlContract(context: JsonObject, obligationId: unknown
     required_on_periodic_review: true,
     rule: "Enabled periodic and observed-stall reviews require a current evidence-linked Vision path and stride decision before any settlement, including progress/successor/blocker/terminal exits. " +
       "Each axis needs plan/reason: grow useful completed scope on a supported route; switch names reference method and changed mechanism; verify resolves a grow-or-switch question. Apply these plans to the actual Todo scope and acceptance using the existing Todo authoring/update commands, then record the verified IDs in todo_delta. Retain unfinished useful work; do not create successor microtasks merely to count progress. On grow, consolidate fragmented setup, implementation, integration and decisive validation into a coherent accepted deliverable or complete interpretable experiment, including engineering/maintenance. The next admitted work must execute that scope, method and validation plan; changing review frequency or repeating the proposal is not adoption. Keep the configured review interval and preserve dependencies, acceptance, risk, authority and per-Turn settlement boundaries. " +
-      "Review the exact previous decision's adoption/information/goal gain and delivered Explore sources in their review fields. Missing/omitted/unread/unprojected evidence stays unknown, not exhausted. Guidance adds no fields/gates; the example is syntax, not an action recommendation. Claims certify no authority, execution or utility. Other legal outcomes cannot bypass a due review.",
+      "Review the exact previous decision's adoption/information/goal gain and delivered Explore sources in their review fields. Apply review_reference_limit separately to previous_review.evidence_refs and explore_review.considered_refs. Missing/omitted/unread/unprojected evidence stays unknown, not exhausted. Guidance adds no fields/gates; the example is syntax, not an action recommendation. Claims certify no authority, execution or utility. Other legal outcomes cannot bypass a due review.",
   };
 }

tests/control_plane/test_replan_progress_native_cli.py

@@ -347,6 +347,7 @@ def spend(turn: str, *, todo: str | None = TODO, obligation: str | None = None)
         assert fixture["index"].read_bytes() == original_index
 
         control = admitted["replan_action_packet"]["stride_control"]
+        assert control["review_reference_limit"] == 3
         decision = {**control["minimal_example"], "basis": "unexplained_gap", "action": "verify",
             "axes": {
                 "work_slice": {"plan": "Run and compare three complete assignment methods.",

tests/control_plane_ts/stride_policy.test.ts

@@ -94,6 +94,25 @@ test("Explore review keeps bare lowercase digest revisions and unavailable null
   }
 });
 
+test("each authored review list advertises and enforces its own three-reference bound", () => {
+  const control = strideControlContract({stride_strategy: "evidence_conditioned_v1"})!;
+  assert.equal(control.review_reference_limit, 3);
+  const evidence = ["evidence:first", "evidence:second", "evidence:third"];
+  const review = {reference_decision_ref: "replan-decision-" + "b".repeat(24), adoption: "unknown",
+    information_gain: "unknown", goal_gain: "unknown", lesson: "Interpret three observations together.",
+    evidence_refs: evidence};
+  const explore = {revision: "a".repeat(64), disposition: "used", considered_refs: evidence, reason: "Compared three findings."};
+  const decision = {...proposal, previous_review: review, explore_review: explore};
+  assert.deepEqual(normalizeStrideDecision(decision)[0]!.previous_review, review);
+  for (const extra of ["evidence:fourth", "evidence:first"]) {
+    const refs = extra === "evidence:first" ? [extra, extra] : [...evidence, extra];
+    assert.throws(() => normalizeStrideDecision({...decision, previous_review: {...review, evidence_refs: refs}}),
+      /previous_review\.evidence_refs requires at most 3 distinct/);
+    assert.throws(() => normalizeStrideDecision({...decision, explore_review: {...explore, considered_refs: refs}}),
+      /explore_review\.considered_refs requires at most 3 distinct/);
+  }
+});
+
 test("concrete independent plans retain action choice and fit the shared vision budget", () => {
   for (const basis of ["unexplained_gap", "uncertain_evidence", "supported_progress"]) {
     const decision = normalizeStrideDecision({...proposal, basis})[0]!;