feat: require concrete evidence for progress claims

hooks/force-recall/handler.ts

@@ -124,6 +124,9 @@ function buildGateLockInput(wrapperResult: any): Record<string, unknown> {
         concreteNextAction: requiredNextAction,
       }
     : null;
+  const progressEvidenceReason = progressEvidence
+    ? ""
+    : "progression claim requires concrete evidence such as sessionKey, runId, modified_files, or verification result";
   const hasExternalizedCheckpointEvidence = wrapperResult.silentLaunchOk === true
     && typeof wrapperResult.taskRecord?.task_name === "string"
     && wrapperResult.taskRecord.task_name.trim().length > 0;
@@ -145,6 +148,7 @@ function buildGateLockInput(wrapperResult: any): Record<string, unknown> {
     statusSummary: claimedProgression,
     executionEvidence,
     progressEvidence,
+    progressEvidenceReason,
     sessionKey: typeof progressEvidence?.sessionKey === "string" ? progressEvidence.sessionKey : "",
     runId: typeof progressEvidence?.runId === "string" ? progressEvidence.runId : "",
     modified_files: Array.isArray(progressEvidence?.modified_files) ? progressEvidence.modified_files : [],
@@ -221,6 +225,7 @@ function buildGateLockBlock(gateLockResult: GateLockResult | null): string {
       "gateRequired=unknown",
       "- ENFORCEMENT: Gate-lock evaluator unavailable; keep existing long-task safeguards in force.",
       "- ENFORCEMENT: Do not claim you have progressed into the next task or are already pushing the next step unless you have concrete progress evidence such as a sessionKey, runId, modified_files record, verification result, actual dispatch, tool calls, file changes, or a persisted checkpoint artifact.",
+      "- ENFORCEMENT: Hook inputs for any progression claim should carry progressEvidence (or equivalent concrete fields) so the gate can verify the claim.",
       "- HARD_GATE: Evaluator unavailable is not permission to claim silent continuation or next-task progression without verifiable progress evidence.",
       "- HARD_GATE: Fall back to a non-silent, evidence-preserving follow-up if you cannot prove checkpoint state or concrete execution.",
       "[/LONG_TASK_GATE_LOCK]",
@@ -239,6 +244,7 @@ function buildGateLockBlock(gateLockResult: GateLockResult | null): string {
     })),
     ...((gateLockResult.allowedResponseModes ?? []).map((mode) => `allowedResponseMode=${mode}`)),
     "- ENFORCEMENT: Do not claim you have progressed into the next task or are already pushing the next step unless you have concrete progress evidence such as a sessionKey, runId, modified_files record, verification result, actual dispatch, tool calls, file changes, or a persisted checkpoint artifact.",
+    "- ENFORCEMENT: Hook input should include progressEvidence (or equivalent concrete fields) whenever a progression claim is present.",
     "- ENFORCEMENT: Forbidden path: plain-text handoff that pretends the long task is already continuing without an externalized checkpoint.",
     "- ENFORCEMENT: Forbidden path: stating you have already entered the next task/step when the record only contains planning language and no concrete execution evidence.",
   ];
@@ -246,6 +252,7 @@ function buildGateLockBlock(gateLockResult: GateLockResult | null): string {
   if (gateLockResult.gateStatus === "fail") {
     lines.push("- HARD_GATE: Block any plain-text handoff or silent-continuation claim when externalized checkpoint evidence is missing.");
     lines.push("- HARD_GATE: Block any reply path that says you already moved into the next task or are advancing the next step without concrete progress evidence.");
+    lines.push("- HARD_GATE: If a progression claim exists, the hook input must supply progressEvidence (or equivalent concrete fields) before the claim can pass gate.");
     lines.push("- HARD_GATE: Do not say you are already on the next task, already dispatched follow-up work, or already progressing in background unless you can point to a sessionKey, runId, modified_files record, verification result, actual tool execution, file changes, emitted messages, or checkpoint records.");
     lines.push("- HARD_GATE: If required evidence is missing, ask for/produce the checkpoint or downgrade to a non-silent, evidence-preserving follow-up.");
     lines.push("- HARD_GATE: If owner decision is involved, do not replace button-path closure with plain-text handoff.");

scripts/long_task_gate_lock.mjs

@@ -197,7 +197,7 @@ function claimsProgressionWithoutEvidence(input) {
 
   if (!hasNonEmptyString(progressionClaim)) return false;
 
-  return !(hasProgressEvidence(input) || hasExecutionEvidence(input));
+  return !hasProgressEvidence(input);
 }
 
 function evaluateGate(input) {

scripts/test_force_recall_long_task_preflight.mjs

@@ -69,8 +69,10 @@ async function main() {
     'reason=claimed progression without concrete progress evidence is forbidden',
     'requiredEvidence=progressEvidence',
     'requiredValue=sessionKey, runId, modified_files, verification result, or equivalent concrete progress evidence',
+    'ENFORCEMENT: Hook input should include progressEvidence (or equivalent concrete fields) whenever a progression claim is present.',
     'HARD_GATE: Block any plain-text handoff or silent-continuation claim when externalized checkpoint evidence is missing.',
     'HARD_GATE: Block any reply path that says you already moved into the next task or are advancing the next step without concrete progress evidence.',
+    'HARD_GATE: If a progression claim exists, the hook input must supply progressEvidence (or equivalent concrete fields) before the claim can pass gate.',
     'HARD_GATE: Do not say you are already on the next task, already dispatched follow-up work, or already progressing in background unless you can point to a sessionKey, runId, modified_files record, verification result, actual tool execution, file changes, emitted messages, or checkpoint records.',
     'ENFORCEMENT: Forbidden path: plain-text handoff that pretends the long task is already continuing without an externalized checkpoint.',
     'ENFORCEMENT: Forbidden path: stating you have already entered the next task/step when the record only contains planning language and no concrete execution evidence.',