@deksden-com/dd-flow-cli 0.9.0-beta.0 → 0.9.0-beta.7

package/dist/services/vnext-code.js

@@ -4,7 +4,7 @@ import fs from "node:fs";
 import path from "node:path";
 import { AppError } from "../shared/errors.js";
 import { resolveProjectRoot } from "../storage/paths.js";
-import { aggregateCheckDeclarations, checkReceipts, codeExecutionEnvironment, finalCodeCheckDeclarations, runCodeChecks, unchangedFinalGateFailures, workspaceFingerprint } from "./code-checks.js";
+import { aggregateCheckDeclarations, checkReceipts, codeExecutionEnvironment, finalCodeCheckDeclarations, runCodeChecks, unchangedFinalGateFailures, workspaceChangedPaths, workspaceFingerprint } from "./code-checks.js";
 import { requireProjectByRoot } from "./projects.js";
 import { appendFlowRunTimelineEvent, advanceFlowRun, attachFlowRunStage, completeFlowRunStage, completeFlowRun, getFlowRunVariables, gitFacts } from "./runs.js";
 import { validateSchema } from "./schema-validation.js";
@@ -135,8 +135,22 @@ export async function finishVnextCode(context, input) {
     }
     // Reject a malformed semantic receipt before running the expensive gate.
     const verification = verificationForFinish(context, { file: input.verificationFile, projectRoot: run.workspace_root, runId: run.id });
-    if (verification.verdict !== "passed" || verification.unresolved.length > 0) {
-        throw new AppError("code_verification_incomplete", "CODE verification is not accepted; create or complete the necessary repair Work first", 2, { verification });
+    if (verification.verdict === "blocked") {
+        return { ok: true, run_id: run.id, stage, outcome: "blocked", verification, instruction: "CODE remains running. Resolve the stated external or user-input blocker in this same coordinator session, update code-verification.json, then invoke this same stage finish command again." };
+    }
+    if (verification.verdict === "passed" && verification.unresolved.length > 0) {
+        throw new AppError("verification_contradictory", "A passed CODE verification cannot contain unresolved obligations", 2, { unresolved: verification.unresolved });
+    }
+    if (verification.verdict === "needs_repair") {
+        const repair = addVnextCodeRepair(context, {
+            projectRoot,
+            runId: run.id,
+            originWorkIds: works.filter((work) => work.status === "completed").map((work) => work.work_id),
+            semanticUnresolved: verification.unresolved.length ? verification.unresolved : [verification.summary],
+            verificationPath: input.verificationFile,
+            objective: verification.summary
+        });
+        return { ok: true, run_id: run.id, stage, outcome: "repair_required", verification, repair, instruction: "The semantic verification is not accepted. Run the returned repair Work, then update code-verification.json and invoke this same stage finish command again." };
     }
     const checks = finalCodeCheckDeclarations(run.workspace_root, works.flatMap((work) => packet(work)?.checks ?? []));
     const unchangedFailures = unchangedFinalGateFailures(context, { projectId: project.id, runId: run.id, workspaceRoot: run.workspace_root, declarations: checks });
@@ -174,18 +188,26 @@ export async function finishVnextCode(context, input) {
     }
     const allReceipts = checkReceipts(context, { projectId: project.id, runId: run.id });
     const finalReceipts = latestReceiptsByCommand(allReceipts);
+    const projectedVerification = verificationProjection(works, finalReceipts);
+    const unresolvedAcceptance = (projectedVerification.acceptance ?? []).filter((item) => item.status === "unresolved");
+    if (unresolvedAcceptance.length)
+        throw new AppError("code_acceptance_unresolved", "CODE cannot finish until every due acceptance criterion has current checks and evidence", 2, { unresolved: unresolvedAcceptance });
     const now = context.now();
     const timing = stageTiming(home, stage, now);
-    const changedPaths = [...new Set(works.flatMap((work) => resultPaths(work.result)))];
+    const reportedPaths = [...new Set(works.flatMap((work) => resultPaths(work.result)))];
+    const observedChangedPaths = workspaceChangedPaths(run.workspace_root);
+    const changedPaths = observedChangedPaths ?? reportedPaths;
+    const missingReportedPaths = observedChangedPaths === null ? [] : reportedPaths.filter((item) => !observedChangedPaths.includes(item));
+    if (missingReportedPaths.length)
+        throw new AppError("changed_path_not_materialized", "CODE Work reported paths that are not changed in the accepted workspace", 2, { paths: missingReportedPaths });
     const next = nextAction(run, changedPaths);
-    writeProtocolFlowStatus(run.workspace_root, home, run.id, next === "start_code_review" ? "CODE complete; CODE-REVIEW is next." : next === "start_merge" ? "CODE complete; MERGE is queued." : "CODE complete; this RUN reached its configured terminal boundary.");
     const report = {
         schema_id: "dd-flow/stage-report@1",
         run_id: run.id,
         stage,
         generated_at: now,
         verdict: "done",
-        verification: verificationProjection(works, finalReceipts),
+        verification: projectedVerification,
         semantic: {
             result: `Completed ${works.length} CODE Work item${works.length === 1 ? "" : "s"}; ${finalReceipts.filter((receipt) => receipt.status === "passed").length} current final check receipt${finalReceipts.length === 1 ? "" : "s"} passed.`,
             acceptance: [...expected],
@@ -213,6 +235,10 @@ export async function finishVnextCode(context, input) {
         artifacts: { json: "stage-report.json", markdown: "stage-report.md", html: "stage-report.html", summary: "stage-report.md" },
         validation: { permission_scope: "known_targets_only", memory_bank_scope: "changed_files_and_links_only", status: "passed" }
     };
+    // Creating the MERGE request is part of accepting this terminal handoff.
+    // Do it before materialising a terminal stage report or changing RUN state,
+    // so an invalid merge policy leaves CODE safely running and retryable.
+    const mergeRequest = next === "start_merge" ? ensureVnextMergeRequest(context, { projectRoot, runId: run.id, acceptedPaths: changedPaths }) : null;
     writeReport(root, report);
     completeFlowRunStage(context, {
         projectRoot,
@@ -237,7 +263,6 @@ export async function finishVnextCode(context, input) {
     else {
         advanceFlowRun(context, { projectRoot, runId: run.id, status: "running", verdict: "code_completed", nextAction: next });
     }
-    const mergeRequest = next === "start_merge" ? ensureVnextMergeRequest(context, { projectRoot, runId: run.id }) : null;
     return {
         ok: true,
         run_id: run.id,
@@ -255,8 +280,8 @@ export function addVnextCodeRepair(context, input) {
     const home = requireHome(run);
     if (!input.objective.trim())
         throw new AppError("validation", "Repair objective must not be empty", 2);
-    if (!input.checkReceiptId && !input.reviewFindingIds?.length) {
-        throw new AppError("validation", "Repair needs a failed check receipt or CODE-REVIEW finding", 2);
+    if (!input.checkReceiptId && !input.reviewFindingIds?.length && !input.semanticUnresolved?.length) {
+        throw new AppError("validation", "Repair needs failed check evidence, a CODE-REVIEW finding, or an unresolved CODE verification", 2);
     }
     const receipt = input.checkReceiptId ? context.db.get("SELECT id, declaration_id, run_id, command, status, receipt_path, stdout_path, stderr_path FROM check_receipts WHERE project_id = ? AND id = ?", [
         project.id,
@@ -269,11 +294,19 @@ export function addVnextCodeRepair(context, input) {
     }
     const origins = [...new Set(input.originWorkIds)].map((id) => requireCodeWork(context, project.id, run.id, id));
     const packets = origins.map((work) => packet(work));
-    const invariantPackets = receipt
+    const invariantPackets = receipt || input.semanticUnresolved?.length
         ? codeWorks(context, project.id, run.id).map((work) => packet(work)).filter((value) => !value.repair)
         : packets;
     const first = packets[0];
-    const declaredChecks = finalCodeCheckDeclarations(run.workspace_root, packets.flatMap((value) => value.checks));
+    // A later repair may be caused by any accepted aggregate declaration, not
+    // only the checks copied into its immediate repair parent.  Otherwise a
+    // repair that fixes one failed aggregate check can make a second, already
+    // declared aggregate failure impossible to repair.  The immutable source
+    // of repair eligibility is the original CODE graph.
+    const declaredChecks = finalCodeCheckDeclarations(run.workspace_root, codeWorks(context, project.id, run.id)
+        .map((work) => packet(work))
+        .filter((value) => !value.repair)
+        .flatMap((value) => value.checks));
     const failedCheck = receipt ? declaredChecks.find((check) => check.id === receipt.declaration_id) : undefined;
     if (receipt && !failedCheck) {
         throw new AppError("repair_check_declaration_missing", "Repair receipt is not backed by an accepted CODE check declaration", 2, {
@@ -281,7 +314,8 @@ export function addVnextCodeRepair(context, input) {
             declaration_id: receipt.declaration_id
         });
     }
-    const key = input.reviewFindingIds?.length ? "code-review-repair" : "code-gate-repair";
+    const semanticRepair = Boolean(input.semanticUnresolved?.length);
+    const key = input.reviewFindingIds?.length ? "code-review-repair" : semanticRepair ? "code-verification-repair" : "code-gate-repair";
     const receiptWriteScope = receipt ? receiptRepairPaths(run.workspace_root, receipt) : [];
     const repair = {
         schema_id: "dd-flow/code-work-packet@5",
@@ -291,22 +325,23 @@ export function addVnextCodeRepair(context, input) {
         repair: {
             origin_work_ids: origins.map((work) => work.work_id),
             ...(receipt ? { check_receipt_id: receipt.id, failure_receipt_path: receipt.receipt_path } : {}),
-            ...(input.reviewFindingIds?.length ? { review_finding_ids: unique(input.reviewFindingIds), review_evidence_refs: unique(input.reviewEvidenceRefs ?? []), review_check_refs: unique(input.reviewCheckRefs ?? (input.reviewChecks ?? []).map((check) => check.id)) } : {})
+            ...(input.reviewFindingIds?.length ? { review_finding_ids: unique(input.reviewFindingIds), review_evidence_refs: unique(input.reviewEvidenceRefs ?? []), review_check_refs: unique(input.reviewCheckRefs ?? (input.reviewChecks ?? []).map((check) => check.id)) } : {}),
+            ...(semanticRepair ? { semantic_unresolved: unique(input.semanticUnresolved ?? []), verification_path: input.verificationPath } : {})
         },
         task: input.objective,
         semantic_spine: {
-            user_outcome: receipt ? `Restore the accepted behavior after aggregate failure: ${input.objective}` : `Resolve accepted CODE-REVIEW finding: ${input.objective}`,
+            user_outcome: receipt ? `Restore the accepted behavior after aggregate failure: ${input.objective}` : semanticRepair ? `Close the unresolved CODE verification: ${input.objective}` : `Resolve accepted CODE-REVIEW finding: ${input.objective}`,
             component_responsibility: "Diagnose and repair the failed accepted CODE result without changing unrelated behavior.",
             must_preserve: unique(invariantPackets.flatMap((value) => value.semantic_spine.must_preserve)),
             non_goals: unique(invariantPackets.flatMap((value) => value.semantic_spine.non_goals)),
-            acceptance_contribution: receipt ? `Make failed check pass: ${receipt.command}` : `Resolve CODE-REVIEW finding(s): ${input.reviewFindingIds.join(", ")}`
+            acceptance_contribution: receipt ? `Make failed check pass: ${receipt.command}` : semanticRepair ? `Resolve CODE verification gap(s): ${input.semanticUnresolved.join("; ")}` : `Resolve CODE-REVIEW finding(s): ${input.reviewFindingIds.join(", ")}`
         },
         requirements: uniqueBy(invariantPackets.flatMap((value) => value.requirements), (value) => value.id),
         acceptance: uniqueBy(invariantPackets.flatMap((value) => value.acceptance), (value) => JSON.stringify(value)),
         // A CODE-REVIEW repair changes delivered code/evidence, never the
         // already accepted PLAN or its ownership declaration.
         document_updates: [],
-        required_read: unique([...(receipt ? [receipt.receipt_path] : input.reviewEvidenceRefs ?? []), ...packets.flatMap((value) => value.required_read)]),
+        required_read: unique([...(receipt ? [receipt.receipt_path] : input.reviewEvidenceRefs ?? []), ...(semanticRepair && input.verificationPath ? [input.verificationPath] : []), ...packets.flatMap((value) => value.required_read)]),
         discovery_boundary: unique(packets.flatMap((value) => value.discovery_boundary)),
         // These are collision-avoidance hints. Receipt paths enrich the coordinator
         // picture but never restrict the repair's project-local edits.
@@ -314,7 +349,7 @@ export function addVnextCodeRepair(context, input) {
         // Receipts record the resolved shell command. A repair must retain the
         // accepted declaration (including its immutable @check alias) and only
         // change when it runs, so work finish can validate it again.
-        checks: selectRepairChecks({ ...(failedCheck ? { failedCheck } : {}), ...(input.reviewChecks ? { reviewChecks: input.reviewChecks } : {}) }),
+        checks: selectRepairChecks({ ...(failedCheck ? { failedCheck } : {}), ...(input.reviewChecks ? { reviewChecks: input.reviewChecks } : {}), ...(semanticRepair ? { semanticChecks: declaredChecks } : {}) }),
         provides_checks: [],
         stop_conditions: unique([
             ...invariantPackets.flatMap((value) => value.stop_conditions),
@@ -339,16 +374,18 @@ export function addVnextCodeRepair(context, input) {
         type: "code_repair_created",
         work_id: id,
         origin_work_ids: input.originWorkIds,
-        ...(receipt ? { check_receipt_id: receipt.id } : { review_finding_ids: input.reviewFindingIds })
+        ...(receipt ? { check_receipt_id: receipt.id } : input.reviewFindingIds?.length ? { review_finding_ids: input.reviewFindingIds } : { semantic_unresolved: input.semanticUnresolved })
     });
     refreshRunWorkProjection(context, project.id, run.id);
     return { ok: true, run_id: run.id, repair_work_id: id, start_command: workStartCommand(context, work) };
 }
-/** The repair Work runs only the check that motivated it; the stage retains the aggregate gate. */
+/** Retain the causal declaration for worker context; run_at keeps aggregate gates at stage scope. */
 export function selectRepairChecks(input) {
-    const checks = input.failedCheck
-        ? [{ ...input.failedCheck, purpose: "Re-run the failed accepted check.", run_at: "work" }]
-        : (input.reviewChecks ?? []).map((check) => ({ ...check, purpose: `Prove the CODE-REVIEW repair: ${check.purpose}`, run_at: "work" }));
+    // Gate placement belongs to the accepted PLAN. `runCodeChecks` executes only
+    // work-scoped entries, while retaining the aggregate declaration tells the
+    // worker exactly what will be rerun by the coordinator.
+    const candidates = input.failedCheck ? [input.failedCheck] : input.reviewChecks ?? input.semanticChecks ?? [];
+    const checks = candidates.map((check) => ({ ...check, purpose: `${input.semanticChecks ? "Prove the CODE verification repair" : "Prove the CODE-REVIEW repair"}: ${check.purpose}` }));
     return uniqueBy(checks, (check) => check.id);
 }
 function receiptRepairPaths(workspaceRoot, receipt) {

package/dist/services/vnext-merge.js

@@ -14,6 +14,7 @@ import { writeStageReport } from "./stage-report-renderer.js";
 import { bindStageCoordinatorWork, createChildWork, finishFanInWork, finishWork, refreshRunWorkProjection, startStageCoordinatorWork } from "./work-registry.js";
 import { checkReceipts, effectiveCheckDeclarations, readCodeCheckProfile, runCodeChecks, unchangedFinalGateFailures, workspaceFingerprint } from "./code-checks.js";
 import { applyExternalStageContext } from "./stage-context.js";
+import { ensureLaneWorkspace, heartbeatLaneLock, releaseLaneLock, waitAcquireLaneLock } from "./lanes.js";
 const stage = "merge";
 const stageDir = vnextStageDirectory(stage);
 export function isVnextMergeRun(context, input) {
@@ -35,18 +36,49 @@ export function ensureVnextMergeRequest(context, input) {
     const sourceBranch = gitValue(run.workspace_root, ["branch", "--show-current"]);
     if (!sourceBranch)
         throw new AppError("merge_source_invalid", "MERGE requires a named source branch", 1, { workspace_root: run.workspace_root });
-    commitPendingSource(run.workspace_root, run.id, ignoredGitPaths(context, project.id));
-    const sourceCommit = gitValue(run.workspace_root, ["rev-parse", "HEAD"]);
     const targetBranch = workspaceRoute(home).integration_branch;
-    if (!sourceCommit || !targetBranch)
-        throw new AppError("merge_source_invalid", "MERGE could not freeze source commit or target branch", 1, { source_commit: sourceCommit, target_branch: targetBranch });
     const targetWorkspace = projectRoot;
-    const enqueueTarget = gitValue(targetWorkspace, ["rev-parse", targetBranch], true);
+    // Validate every immutable prerequisite before touching the source tree. A
+    // rejected MERGE handoff must not leave a new commit behind or let a caller
+    // subsequently mark CODE/CODE-REVIEW complete without a queue request.
+    if (!targetBranch)
+        throw new AppError("merge_source_invalid", "MERGE could not determine the target branch", 1, { target_branch: targetBranch });
     const semantic = planChecks(run.workspace_root, protocols);
     validateMergeAcceptance(run.workspace_root, protocols, semantic);
     const effective = effectiveCheckDeclarations(run.workspace_root, semantic, ["merge"]);
     if (run.workspace_root !== targetWorkspace && effective.length === 0)
         throw new AppError("merge_gate_missing", "A real source/target integration requires at least one semantic or policy merge check", 2, { run_id: run.id, protocol_ids: protocols });
+    const ignored = ignoredGitPaths(context, project.id);
+    const accepted = input.acceptedPaths ? [...new Set(input.acceptedPaths.filter((item) => !ignored.includes(item)))].sort() : null;
+    const freezeRoot = path.join(home, stageDir);
+    const freezeFile = path.join(freezeRoot, "source-freeze.json");
+    fs.mkdirSync(freezeRoot, { recursive: true });
+    let sourceCommit;
+    let sourcePaths;
+    if (fs.existsSync(freezeFile)) {
+        const frozen = JSON.parse(fs.readFileSync(freezeFile, "utf8"));
+        sourceCommit = frozen.run_id === run.id && typeof frozen.source_commit === "string" ? frozen.source_commit : "";
+        sourcePaths = Array.isArray(frozen.paths) ? frozen.paths : [];
+        if (!sourceCommit || gitValue(run.workspace_root, ["rev-parse", "HEAD"]) !== sourceCommit || (accepted && JSON.stringify(accepted) !== JSON.stringify(sourcePaths)))
+            throw new AppError("merge_source_freeze_conflict", "Existing MERGE source freeze does not match the accepted source", 2, { freeze_file: freezeFile });
+    }
+    else {
+        sourcePaths = [...new Set(meaningfulStatus(run.workspace_root, ignored).map(statusPath))].sort();
+        if (accepted) {
+            const unexpected = sourcePaths.filter((item) => !accepted.includes(item));
+            const absent = accepted.filter((item) => !sourcePaths.includes(item));
+            if (unexpected.length || absent.length)
+                throw new AppError("merge_source_drift", "MERGE source differs from the accepted terminal-stage file set", 2, { unexpected, absent });
+        }
+        const sourceFingerprint = workspaceFingerprint(run.workspace_root);
+        commitPendingSource(run.workspace_root, run.id, ignored);
+        sourceCommit = gitValue(run.workspace_root, ["rev-parse", "HEAD"]);
+        if (sourceCommit)
+            fs.writeFileSync(freezeFile, `${JSON.stringify({ schema_id: "dd-flow/merge-source-freeze@1", run_id: run.id, source_workspace: run.workspace_root, source_branch: sourceBranch, source_commit: sourceCommit, precommit_workspace_fingerprint: sourceFingerprint, paths: sourcePaths, at: context.now() }, null, 2)}\n`);
+    }
+    if (!sourceCommit)
+        throw new AppError("merge_source_invalid", "MERGE could not freeze source commit", 1, { source_commit: sourceCommit, target_branch: targetBranch });
+    const enqueueTarget = gitValue(targetWorkspace, ["rev-parse", targetBranch], true);
     const root = requireRootWork(context, project.id, run.id);
     let child;
     let requestId;
@@ -83,38 +115,50 @@ export async function startVnextMerge(context, input) {
     fs.mkdirSync(workRoot, { recursive: true });
     const resultPath = path.join(workRoot, "result.json");
     const promptPath = path.join(workRoot, "prompt.md");
-    if (["active", "action_required"].includes(request.status)) {
-        const pendingWork = context.db.get("SELECT status FROM works WHERE work_id = ?", [request.executor_work_id]);
-        const binding = pendingWork?.status === "created"
-            ? startStageCoordinatorWork(context, { workId: request.executor_work_id, hookEventId: input.hookEventId, stage, projectRoot, ...(input.contextSha256 ? { contextSha256: input.contextSha256 } : {}) })
-            : bindStageCoordinatorWork(context, { workId: request.executor_work_id, hookEventId: input.hookEventId, stage, promptPath, resultPath, ...(input.contextSha256 ? { contextSha256: input.contextSha256 } : {}) });
-        const prompt = fs.existsSync(promptPath) ? fs.readFileSync(promptPath, "utf8") : mergePrompt(context, { run, request, resultPath, projectRoot });
-        if (!fs.existsSync(promptPath))
-            fs.writeFileSync(promptPath, prompt);
+    ensureLaneWorkspace(context, { projectRoot, lane: "merge", workspacePath: request.target_workspace, branch: request.target_branch });
+    const lane = await waitAcquireLaneLock(context, { projectRoot, lane: "merge", workerId: request.executor_work_id, workspacePath: request.target_workspace, timeoutSeconds: 3600, pollIntervalSeconds: 15, ttlSeconds: 7200, reason: `MERGE ${request.merge_request_id}`, ...(input.progress ? { progress: input.progress } : {}) });
+    if (!lane.acquired)
+        throw new AppError("merge_lane_wait_failed", "MERGE could not acquire the project integration lane", 1, { merge_request_id: request.merge_request_id, status: lane.status ?? "unknown" });
+    try {
+        if (["active", "action_required"].includes(request.status)) {
+            const pendingWork = context.db.get("SELECT status FROM works WHERE work_id = ?", [request.executor_work_id]);
+            const binding = pendingWork?.status === "created"
+                ? startStageCoordinatorWork(context, { workId: request.executor_work_id, hookEventId: input.hookEventId, stage, projectRoot, ...(input.contextSha256 ? { contextSha256: input.contextSha256 } : {}) })
+                : bindStageCoordinatorWork(context, { workId: request.executor_work_id, hookEventId: input.hookEventId, stage, promptPath, resultPath, ...(input.contextSha256 ? { contextSha256: input.contextSha256 } : {}) });
+            const prompt = fs.existsSync(promptPath) ? fs.readFileSync(promptPath, "utf8") : mergePrompt(context, { run, request, resultPath, projectRoot });
+            if (!fs.existsSync(promptPath))
+                fs.writeFileSync(promptPath, prompt);
+            const externalContext = applyExternalStageContext({ stageRoot: root, promptPath, ...(input.externalContext ? { loaded: input.externalContext } : {}) });
+            ensureMergeStageAttached(context, run, projectRoot);
+            return startPacket(context, run, request, binding, promptPath, resultPath, prompt, projectRoot, externalContext);
+        }
+        const current = requireRequest(context, request.merge_request_id);
+        const targetHead = gitValue(current.target_workspace, ["rev-parse", current.target_branch], true);
+        const checkedOutTarget = gitValue(current.target_workspace, ["branch", "--show-current"], true);
+        if (checkedOutTarget !== current.target_branch)
+            throw new AppError("merge_target_branch_mismatch", "Integration workspace is not on the configured target branch", 1, { expected: current.target_branch, actual: checkedOutTarget, target_workspace: current.target_workspace });
+        const targetStatus = meaningfulStatus(current.target_workspace, ignoredGitPaths(context, project.id));
+        if (targetStatus.length)
+            throw new AppError("merge_target_dirty", "Integration workspace must be clean before MERGE starts", 1, { target_workspace: current.target_workspace, status: targetStatus });
+        const claimed = context.db.run("UPDATE merge_requests SET status = 'active', execution_target_head = ?, lock_acquired_at = ?, checkpoint = 'baseline_locked', updated_at = ? WHERE merge_request_id = ? AND status IN ('queued','dispatching','action_required')", [targetHead, context.now(), context.now(), current.merge_request_id]);
+        if (claimed.changes !== 1 && requireRequest(context, current.merge_request_id).status !== "active")
+            throw new AppError("merge_claim_conflict", "MERGE request could not acquire the integration lane", 1, { merge_request_id: current.merge_request_id });
+        const started = startStageCoordinatorWork(context, { workId: current.executor_work_id, hookEventId: input.hookEventId, stage, projectRoot, ...(input.contextSha256 ? { contextSha256: input.contextSha256 } : {}) });
+        const prompt = mergePrompt(context, { run, request: requireRequest(context, current.merge_request_id), resultPath, projectRoot });
+        fs.writeFileSync(promptPath, prompt);
         const externalContext = applyExternalStageContext({ stageRoot: root, promptPath, ...(input.externalContext ? { loaded: input.externalContext } : {}) });
-        ensureMergeStageAttached(context, run, projectRoot);
-        return startPacket(context, run, request, binding, promptPath, resultPath, prompt, projectRoot, externalContext);
+        context.db.run("UPDATE work_sessions SET prompt_path = ?, result_path = ?, updated_at = ? WHERE work_id = ? AND status = 'running'", [promptPath, resultPath, context.now(), current.executor_work_id]);
+        attachFlowRunStage(context, { projectRoot, runId: run.id, stage, dir: stageDir, status: "running", dataSchemaId: "dd-flow/stage-report@2" });
+        appendFlowRunTimelineEvent(context, project.id, run.id, { type: "merge_started", merge_request_id: current.merge_request_id, work_id: current.executor_work_id, target_head: targetHead });
+        return startPacket(context, run, current, started, promptPath, resultPath, prompt, projectRoot, externalContext);
+    }
+    catch (error) {
+        const latest = requireRequest(context, request.merge_request_id);
+        if (["active", "dispatching"].includes(latest.status))
+            context.db.run("UPDATE merge_requests SET status = 'action_required', last_error_json = ?, updated_at = ? WHERE merge_request_id = ?", [JSON.stringify({ code: "merge_start_failed", error: error instanceof Error ? error.message : String(error) }), context.now(), latest.merge_request_id]);
+        releaseMergeLane(context, projectRoot, latest, `MERGE ${latest.merge_request_id} start failed`);
+        throw error;
     }
-    await waitForTurn(context, request, input.progress);
-    const current = requireRequest(context, request.merge_request_id);
-    const targetHead = gitValue(current.target_workspace, ["rev-parse", current.target_branch], true);
-    const checkedOutTarget = gitValue(current.target_workspace, ["branch", "--show-current"], true);
-    if (checkedOutTarget !== current.target_branch)
-        throw new AppError("merge_target_branch_mismatch", "Integration workspace is not on the configured target branch", 1, { expected: current.target_branch, actual: checkedOutTarget, target_workspace: current.target_workspace });
-    const targetStatus = meaningfulStatus(current.target_workspace, ignoredGitPaths(context, project.id));
-    if (targetStatus.length)
-        throw new AppError("merge_target_dirty", "Integration workspace must be clean before MERGE starts", 1, { target_workspace: current.target_workspace, status: targetStatus });
-    const claimed = context.db.run("UPDATE merge_requests SET status = 'active', execution_target_head = ?, lock_acquired_at = ?, checkpoint = 'baseline_locked', updated_at = ? WHERE merge_request_id = ? AND status IN ('queued','dispatching','action_required')", [targetHead, context.now(), context.now(), current.merge_request_id]);
-    if (claimed.changes !== 1 && requireRequest(context, current.merge_request_id).status !== "active")
-        throw new AppError("merge_claim_conflict", "MERGE request could not acquire the integration lane", 1, { merge_request_id: current.merge_request_id });
-    const started = startStageCoordinatorWork(context, { workId: current.executor_work_id, hookEventId: input.hookEventId, stage, projectRoot, ...(input.contextSha256 ? { contextSha256: input.contextSha256 } : {}) });
-    const prompt = mergePrompt(context, { run, request: requireRequest(context, current.merge_request_id), resultPath, projectRoot });
-    fs.writeFileSync(promptPath, prompt);
-    const externalContext = applyExternalStageContext({ stageRoot: root, promptPath, ...(input.externalContext ? { loaded: input.externalContext } : {}) });
-    context.db.run("UPDATE work_sessions SET prompt_path = ?, result_path = ?, updated_at = ? WHERE work_id = ? AND status = 'running'", [promptPath, resultPath, context.now(), current.executor_work_id]);
-    attachFlowRunStage(context, { projectRoot, runId: run.id, stage, dir: stageDir, status: "running", dataSchemaId: "dd-flow/stage-report@2" });
-    appendFlowRunTimelineEvent(context, project.id, run.id, { type: "merge_started", merge_request_id: current.merge_request_id, work_id: current.executor_work_id, target_head: targetHead });
-    return startPacket(context, run, current, started, promptPath, resultPath, prompt, projectRoot, externalContext);
 }
 function startPacket(context, run, request, binding, promptPath, resultPath, prompt, projectRoot, externalContext) {
     return { ok: true, run_id: run.id, stage, merge_request_id: request.merge_request_id, work_id: request.executor_work_id, session_binding: binding.session_binding ?? binding, prompt_path: promptPath, result_path: resultPath, worker_prompt_markdown: prompt, ...(externalContext ? { external_context: externalContext } : {}), queue: queueStatus(context, request), effective_checks: effectiveMergeChecks(run, request), next: { apply_command: applyCommand(context, request, projectRoot), finish_command: finishCommand(context, run.id, request, projectRoot) } };
@@ -125,6 +169,7 @@ export function applyVnextMerge(context, input) {
     if (existing.executor_work_id === input.workId && ["apply_recorded", "integration_committed", "bootstrap_ready", "checks_passed", "delivered", "finalized"].includes(existing.checkpoint) && fs.existsSync(receiptFile))
         return JSON.parse(fs.readFileSync(receiptFile, "utf8"));
     const request = requireOwnedActiveRequest(context, input);
+    heartbeatLaneLock(context, { projectRoot: input.projectRoot, lane: "merge", workerId: request.executor_work_id, workspacePath: request.target_workspace, ttlSeconds: 7200 });
     assertExecutionBaseline(request);
     input.progress?.(`applying ${request.source_commit} to ${request.target_branch}`);
     let outcome = "applied_clean";
@@ -157,6 +202,7 @@ export async function finishVnextMerge(context, input) {
         return { ok: true, resumed: true, run_id: run.id, stage, outcome: "completed", merge_request_id: existing.merge_request_id, work_id: existing.executor_work_id, integration_commit: existing.integration_commit, report_path: path.join(requireHome(run), stageDir, "stage-report.json"), next_action: "merge_completed" };
     }
     let request = requireOwnedActiveRequest(context, input);
+    heartbeatLaneLock(context, { projectRoot, lane: "merge", workerId: request.executor_work_id, workspacePath: request.target_workspace, ttlSeconds: 7200 });
     if (request.run_id !== run.id)
         throw new AppError("merge_request_mismatch", "MRG does not belong to RUN", 2);
     if (request.checkpoint === "baseline_locked" || !fs.existsSync(applyReceiptPath(context, request)))
@@ -218,6 +264,7 @@ export async function finishVnextMerge(context, input) {
     completeFlowRun(context, { projectRoot, runId: run.id, status: "done", verdict: "merge_completed", nextAction: undefined });
     refreshRunWorkProjection(context, project.id, run.id);
     appendFlowRunTimelineEvent(context, project.id, run.id, { type: "merge_completed", merge_request_id: request.merge_request_id, integration_commit: requireRequest(context, request.merge_request_id).integration_commit });
+    releaseMergeLane(context, projectRoot, request, `MERGE ${request.merge_request_id} completed`);
     return { ok: true, run_id: run.id, stage, outcome: "completed", merge_request_id: request.merge_request_id, work_id: request.executor_work_id, integration_commit: requireRequest(context, request.merge_request_id).integration_commit, report_path: path.join(root, "stage-report.json"), next_action: "merge_completed" };
 }
 export function getVnextMergeRequest(context, input) { const project = requireProjectByRoot(context, resolveProjectRoot(input.projectRoot)); const request = input.requestId ? requireRequest(context, input.requestId) : input.runId ? requestForRun(context, project.id, input.runId) : null; if (!request || request.project_id !== project.id)
@@ -277,15 +324,28 @@ function performConfiguredCleanup(run, request) {
     const policy = executionSettings(run).merge_cleanup?.source ?? "retain";
     const receipt = cleanupReceiptPath(run);
     let action = "retained";
-    if (policy === "delete_after_success" && path.resolve(request.source_workspace) !== path.resolve(request.target_workspace)) {
-        git(request.target_workspace, ["worktree", "remove", request.source_workspace]);
-        git(request.target_workspace, ["branch", "-d", request.source_branch]);
-        action = "deleted";
+    let error = null;
+    try {
+        if (policy === "delete_after_success" && path.resolve(request.source_workspace) !== path.resolve(request.target_workspace)) {
+            git(request.target_workspace, ["worktree", "remove", request.source_workspace]);
+            git(request.target_workspace, ["branch", "-d", request.source_branch]);
+            action = "deleted";
+        }
+    }
+    catch (cause) {
+        // Cleanup is post-delivery hygiene. It must never invalidate an integration
+        // commit which already passed the merge gate and was delivered.
+        action = "action_required";
+        error = cause instanceof Error ? cause.message : String(cause);
+    }
+    fs.writeFileSync(receipt, `${JSON.stringify({ schema_id: "dd-flow/merge-cleanup-receipt@1", merge_request_id: request.merge_request_id, policy, action, source_workspace: request.source_workspace, source_branch: request.source_branch, ...(error ? { error } : {}), at: new Date().toISOString() }, null, 2)}\n`);
+}
+function releaseMergeLane(context, projectRoot, request, reason) {
+    try {
+        releaseLaneLock(context, { projectRoot, lane: "merge", workerId: request.executor_work_id, workspacePath: request.target_workspace, reason });
     }
-    fs.writeFileSync(receipt, `${JSON.stringify({ schema_id: "dd-flow/merge-cleanup-receipt@1", merge_request_id: request.merge_request_id, policy, action, source_workspace: request.source_workspace, source_branch: request.source_branch, at: new Date().toISOString() }, null, 2)}\n`);
+    catch { /* An expired lease cannot invalidate an already completed MERGE. */ }
 }
-function waitForTurn(context, request, progress) { return new Promise((resolve) => { const poll = () => { const current = requireRequest(context, request.merge_request_id); const ahead = queueAhead(context, current); const active = context.db.get("SELECT merge_request_id FROM merge_requests WHERE project_id = ? AND status IN ('active','action_required','recovery_required') AND merge_request_id <> ? LIMIT 1", [current.project_id, current.merge_request_id]); if (ahead === 0 && !active)
-    return resolve(); progress?.(`MERGE ${current.merge_request_id} is waiting: ${ahead} request(s) ahead; next update in 15 seconds`); setTimeout(poll, 15_000); }; poll(); }); }
 function queueAhead(context, request) { return context.db.get("SELECT COUNT(*) AS count FROM merge_requests WHERE project_id = ? AND status NOT IN ('completed','failed','cancelled') AND (created_at < ? OR (created_at = ? AND merge_request_id < ?))", [request.project_id, request.created_at, request.created_at, request.merge_request_id])?.count ?? 0; }
 function queueStatus(context, request) { return { position: queueAhead(context, request) + 1, requests_ahead: queueAhead(context, request), status: request.status, route: request.execution_route }; }
 function requireOwnedActiveRequest(context, input) { const project = requireProjectByRoot(context, resolveProjectRoot(input.projectRoot)); const request = requireRequest(context, input.requestId); if (request.project_id !== project.id || request.executor_work_id !== input.workId || !["active", "action_required"].includes(request.status))

package/dist/services/vnext-plan-review.js

@@ -14,6 +14,7 @@ import { requireVnextWorkspaceRoute } from "./vnext-workspace-policy.js";
 import { nextWorkId } from "./ids.js";
 import { vnextStageDirectory } from "../domain/stage-catalog.js";
 import { writeStageReport } from "./stage-report-renderer.js";
+import { assertPortableArtifactRef } from "./portable-refs.js";
 import { applyExternalStageContext } from "./stage-context.js";
 import { capacityProbe, readFanoutDescriptor, subagentCapacityKey, writeFanoutDescriptor } from "./vnext-fanout.js";
 const stage = "plan-review";
@@ -48,7 +49,7 @@ export function startVnextPlanReview(context, input) {
     if (existing?.status === "running") {
         const promptPath = path.join(root, "stage-prompt.md");
         const externalContext = applyExternalStageContext({ stageRoot: root, promptPath, ...(input.externalContext ? { loaded: input.externalContext } : {}) });
-        return { ...preparedExisting(context, { projectRoot, projectId: project.id, run, root, workId: existing.work_id, hookEventId: input.hookEventId, ...(input.contextSha256 ? { contextSha256: input.contextSha256 } : {}), requested, effective, groups, batchChecksum: checksum(batch), planChecksum: checksum(path.join(planRoot, "stage-report.json")) }), ...(externalContext ? { external_context: externalContext } : {}) };
+        return { ...preparedExisting(context, { projectRoot, projectId: project.id, run, root, workId: existing.work_id, hookEventId: input.hookEventId, ...(input.contextSha256 ? { contextSha256: input.contextSha256 } : {}), requested, effective, groups, batchChecksum: checksum(batch), planChecksum: planSetChecksum(home, run.workspace_root) }), ...(externalContext ? { external_context: externalContext } : {}) };
     }
     const priorStage = readIndex(run).stage_runs?.find((entry) => entry.stage === stage);
     if (priorStage?.status === "done") {
@@ -61,7 +62,7 @@ export function startVnextPlanReview(context, input) {
     fs.mkdirSync(root, { recursive: true });
     const now = context.now();
     const batchChecksum = checksum(batch);
-    const planChecksum = checksum(path.join(planRoot, "stage-report.json"));
+    const planChecksum = planSetChecksum(home, run.workspace_root);
     if (effective === "off") {
         const failures = validateVnextPlanArtifacts(context, { projectRoot, workspaceRoot: run.workspace_root, runId: run.id, home, protocols: protocolIds });
         if (failures.length)
@@ -131,7 +132,7 @@ export function dispatchVnextPlanReview(context, input) {
     const pending = groups.filter((group) => !latest(group));
     if (pending.length) {
         const file = path.join(root, ".dispatch.json");
-        writeJson(file, { works: pending.map((group) => ({ key: group.key, task: group.task, depends_on: group.depends_on, launch_policy: "fresh_agent_required", result_schema: "dd-flow/plan-review-result@1", payload: { kind: "plan-review", group: { key: group.key, aspect_ids: group.aspect_ids } } })) });
+        writeJson(file, { works: pending.map((group) => ({ key: group.key, task: group.task, depends_on: group.depends_on, launch_policy: "fresh_agent_required", result_schema: "dd-flow/plan-review-result@1", payload: { kind: "plan-review", read_only: true, group: { key: group.key, aspect_ids: group.aspect_ids } } })) });
         try {
             addWorkBatch(context, { parentWorkId: parent.work_id, file });
         }
@@ -190,14 +191,18 @@ export function finishVnextPlanReview(context, input) {
     if (!Number.isInteger(reviewedPlanRevision) || reviewedPlanRevision < 1)
         throw new AppError("runtime_missing", "PLAN-REVIEW starting revision is unavailable", 1);
     const planRevision = currentPlanRevision(home, run.workspace_root);
-    if (["blocked", "failed", "cancelled"].includes(decision.outcome)) {
+    const currentPlanChecksum = planSetChecksum(home, run.workspace_root);
+    if (decision.outcome === "blocked") {
+        throw new AppError("stage_pause_required", "A user or decision blocker must pause the current PLAN-REVIEW Work with `dd-flow stage pause`; do not finish the stage as blocked", 2, { run_id: run.id, stage, work_id: parent.work_id });
+    }
+    if (["failed", "cancelled"].includes(decision.outcome)) {
         preserveDecisionReceipt(root, decisionPath);
         return finishTerminalReview(context, { projectRoot, projectId: project.id, run, root, planRoot, batch, parentWorkId: parent.work_id, children, contextFile, decision });
     }
     if (children.some((child) => child.status === "created" || child.status === "running"))
         throw new AppError("worker_jobs_incomplete", "PLAN-REVIEW finish requires all reviewer Works to settle", 1, { works: children.map((child) => ({ work_id: child.work_id, status: child.status })) });
     const latestChildren = latestReviewChildren(children);
-    const evidenceFailures = reviewerEvidenceFailures(context, { projectId: project.id, parentWorkId: parent.work_id, children: latestChildren, groups, planRevision: reviewedPlanRevision });
+    const evidenceFailures = reviewerEvidenceFailures(context, { projectId: project.id, parentWorkId: parent.work_id, children: latestChildren, groups, planRevision: reviewedPlanRevision, workspaceRoot: run.workspace_root, runHome: home, runId: run.id });
     if (evidenceFailures.length)
         throw new AppError("review_evidence_invalid", "PLAN-REVIEW reviewer evidence is incomplete, stale, or not isolated", 2, { errors: evidenceFailures });
     if (decision.outcome !== "accepted")
@@ -209,17 +214,21 @@ export function finishVnextPlanReview(context, input) {
     const needsCorrection = latestChildren.some((child) => ["needs_changes", "blocked"].includes(reviewerVerdict(child.result)));
     if (latestChildren.some((child) => child.status !== "completed"))
         throw new AppError("blocked", "PLAN-REVIEW cannot accept incomplete reviewer evidence", 2);
-    const materialFindingIds = canonicalReviewerFindings(latestChildren).filter(({ finding }) => ["blocker", "high", "medium"].includes(finding.severity)).map(({ finding_ref }) => finding_ref);
-    const decidedFindingIds = new Set(decision.finding_decisions.map((finding) => finding.finding_ref).filter((id) => Boolean(id)));
-    if (new Set(materialFindingIds).size !== materialFindingIds.length || materialFindingIds.some((id) => !decidedFindingIds.has(id)))
-        throw new AppError("review_evidence_invalid", "Every material reviewer finding needs one stable decision", 2, { material_finding_ids: materialFindingIds, decided_finding_ids: [...decidedFindingIds] });
+    const canonicalFindings = canonicalReviewerFindings(latestChildren);
+    const allFindingIds = canonicalFindings.map(({ finding_ref }) => finding_ref);
+    const materialFindingIds = canonicalFindings.filter(({ finding }) => ["blocker", "high", "medium"].includes(finding.severity)).map(({ finding_ref }) => finding_ref);
+    const decisionFindingIds = decision.finding_decisions.map((finding) => finding.finding_ref).filter((id) => Boolean(id));
+    const decidedFindingIds = new Set(decisionFindingIds);
+    const unknownDecisions = decisionFindingIds.filter((id) => !allFindingIds.includes(id));
+    if (new Set(allFindingIds).size !== allFindingIds.length || decidedFindingIds.size !== decisionFindingIds.length || unknownDecisions.length || materialFindingIds.some((id) => !decidedFindingIds.has(id)))
+        throw new AppError("review_evidence_invalid", "Reviewer findings and coordinator decisions must use unique known canonical references, with every material finding decided", 2, { material_finding_ids: materialFindingIds, decided_finding_ids: decisionFindingIds, unknown_decisions: unknownDecisions });
     if (needsCorrection) {
         if (decision.correction.status !== "applied")
             throw new AppError("validation", "Reviewer findings require an in-place correction receipt", 2);
         if (decision.correction.previous_plan_revision !== reviewedPlanRevision || planRevision <= reviewedPlanRevision || decision.correction.changed_paths.length === 0)
             throw new AppError("validation", "Applied PLAN correction must advance revision and list changed paths", 2, { reviewed_plan_revision: reviewedPlanRevision, final_plan_revision: planRevision });
     }
-    else if (decision.correction.status !== "not_required" || planRevision !== reviewedPlanRevision) {
+    else if (decision.correction.status !== "not_required" || planRevision !== reviewedPlanRevision || currentPlanChecksum !== contextFile.system?.plan_checksum) {
         throw new AppError("validation", "A clean PLAN review must not claim an unverified correction", 2);
     }
     const planFailures = validateVnextPlanArtifacts(context, { projectRoot, workspaceRoot: run.workspace_root, runId: run.id, home, protocols: protocolIdsForReview(home) });
@@ -233,7 +242,7 @@ export function finishVnextPlanReview(context, input) {
             throw new AppError("validation", "Generated CODE batch must not be listed as an agent-authored correction", 2, { changed_paths: decision.correction.changed_paths });
     }
     preserveDecisionReceipt(root, decisionPath);
-    const finalPlanChecksum = checksum(path.join(run.workspace_root, ".memory-bank", "protocol", protocolIdsForReview(home)[0], "plan.json"));
+    const finalPlanChecksum = planSetChecksum(home, run.workspace_root);
     const finalBatchChecksum = checksum(batch);
     const registered = registerCode(context, project.id, run.id, batch);
     const now = context.now();
@@ -258,7 +267,7 @@ function orchestratorPrompt(context, input) {
     const decision = path.join(input.root, "decision.json");
     const revision = currentPlanRevision(input.home, input.run.workspace_root);
     const workspaceContract = ["<workspace_contract>", `- route: ${input.workspaceRoute.route}`, `- feature branch: ${input.workspaceRoute.feature_branch ?? "not applicable"}`, `- base commit: ${input.workspaceRoute.base_ref ?? "not applicable"}`, `- read/write workspace: ${input.run.workspace_root}`, "The CLI verified this frozen route. All plan and correction writes belong in the named workspace; project root remains only the stable lifecycle identity. Do not create, switch, merge or delete branches/worktrees.", "</workspace_contract>"].join("\n");
-    return ["<stage_identity>", `- RUN: ${input.run.id}`, `- Work: ${input.workId}`, "- Stage: plan-review", `- Mode: ${input.effective}`, "</stage_identity>", "", "<trusted_runtime_context>", "These facts were collected by dd-flow. Trust them; do not repeat CLI, Git, compatibility, permission or schema discovery.", `- Project root: ${input.projectRoot}`, `- Stage workspace: ${input.root}`, `- PLAN revision: ${revision}`, `- PLAN report checksum: ${input.planChecksum}`, `- Generated CODE batch checksum: ${input.batchChecksum}`, "</trusted_runtime_context>", "", workspaceContract, "", "<review_groups>", ...input.groups.map((group) => `- ${group.key}: ${group.aspect_ids.join(", ")}`), "</review_groups>", "", "<execution_commands>", `Dispatch fresh reviewers: ${dispatchCommand(context, input.run.id, input.projectRoot)}`, `If dispatch requests capacity, run exactly one concurrent fan-out of ${capacityProbeFanoutSize} probes. This measures the harness limit; it is not a task to obtain ${capacityProbeFanoutSize} successful probes. Start #01…#${capacityProbeFanoutSize} once, all together, using all-settled handling so one rejection does not hide the other outcomes. A rejected launch is expected evidence. Never retry, replace, or add a probe. Each started probe calls no tools, reads no files, creates no children, waits ${capacityProbeHoldSeconds} seconds, then returns exactly AGENT-NN. For cleanup, wait at most ${capacityProbeDeadlineSeconds} seconds from the first launch, terminate every unfinished probe, then release every finished probe session that the harness permits. Only after that cleanup record the number of launches that started successfully, not the number of replacement attempts or late completions: ${capacityRecordCommand(context, input.run.id, input.projectRoot, "<successful-initial-launches>")}. Capacity probes are not Works and are never registered.`, "After dispatch, launch at most the measured capacity at once. If more independent reviewer Works remain, wait for the current wave to settle, then start the unchanged queued Works in the next wave. A reviewer launch rejected before it starts is not review evidence: do not create a replacement; wait for a running wave to settle and start that same queued Work. Each reviewer must be a genuinely fresh harness child Session. The lifecycle adapter binds that observed Session; do not bind or supply a Session ID manually. Reviewers are read-only and must not create children. As soon as a reviewer result is accepted, release that reviewer Session when the harness permits; do not let finished disposable workers occupy slots before the next wave.", "Review the execution environment of every selected check as part of its proof: a reset/fixture process, service process and client process must share the intended data and configuration world. A runtime entrypoint that can break that invariant must be explicit in one Work's task and verification and ordered before its consumer. planned_write_areas may advertise likely overlap, but do not treat them as ownership; required_read alone is not a delivery plan.", "If the final decision needs user input with no reasonable default, run this exact one-command heredoc, replacing only its placeholder body. The heredoc is the permitted stdin form; do not use cat, a pipe, a temporary file or a second shell command:", "```sh", input.pauseCommandTemplate, "```", "Ask the returned user_message, stop, then resume this same PLAN-REVIEW Work. Do not write decision.json or finish first.", `When all reviewer results are complete and every user question is resolved, classify every material finding, fix accepted findings in this same PLAN-REVIEW Work, then write ${decision} and finish: ${finishCommand(context, input.run.id, input.projectRoot, decision)}`, "Reviewer findings use local FIND-NNN ids. dd-flow exposes each finding to this coordinator as WRK-.../FIND-NNN; use that canonical finding_ref in the decision.", "A completed reviewer result with needs_changes or blocked is evidence, not the stage outcome. Classify its material findings and apply accepted fixes in this one review pass; do not start a second review automatically. Only a missing, malformed or unfinished reviewer result blocks the stage. For an accepted correction, increment PLAN revision and update only plan.json and the relevant aspect map. Do not edit or list code-work-batch.json: the CLI validates final PLAN and regenerates it. If no material correction is needed, set correction.status=not_required. The CLI checks mechanical handoff coherence; it does not prove semantic correctness.", "```json", JSON.stringify({ schema_id: "dd-flow/plan-review-decision@3", outcome: "accepted | blocked | failed | cancelled", summary: "Concise evidence-backed final decision.", finding_decisions: [{ finding_ref: "WRK-001-review/FIND-001", decision: "accepted_fix | rejected | deferred_as_DEF | requires_user | duplicate", reason: "Why." }], correction: { status: "not_required | applied", previous_plan_revision: revision, changed_paths: [], summary: "No material correction was needed, or summarize the applied correction." } }, null, 2), "```", "</execution_commands>", "", "<stage_instructions>", template, "</stage_instructions>", ""].join("\n");
+    return ["<stage_identity>", `- RUN: ${input.run.id}`, `- Work: ${input.workId}`, "- Stage: plan-review", `- Mode: ${input.effective}`, "</stage_identity>", "", "<trusted_runtime_context>", "These facts were collected by dd-flow. Trust them; do not repeat CLI, Git, compatibility, permission or schema discovery.", `- Project root: ${input.projectRoot}`, `- Stage workspace: ${input.root}`, `- PLAN revision: ${revision}`, `- PLAN report checksum: ${input.planChecksum}`, `- Generated CODE batch checksum: ${input.batchChecksum}`, "</trusted_runtime_context>", "", workspaceContract, "", "<review_groups>", ...input.groups.map((group) => `- ${group.key}: ${group.aspect_ids.join(", ")}`), "</review_groups>", "", "<execution_commands>", `Dispatch fresh reviewers: ${dispatchCommand(context, input.run.id, input.projectRoot)}`, `If dispatch requests capacity, run exactly one concurrent fan-out of ${capacityProbeFanoutSize} probes. This measures the harness limit; it is not a task to obtain ${capacityProbeFanoutSize} successful probes. Start #01…#${capacityProbeFanoutSize} once, all together, using all-settled handling so one rejection does not hide the other outcomes. A rejected launch is expected evidence. Never retry, replace, or add a probe. Each started probe calls no tools, reads no files, creates no children, waits ${capacityProbeHoldSeconds} seconds, then returns exactly AGENT-NN. For cleanup, wait at most ${capacityProbeDeadlineSeconds} seconds from the first launch, terminate every unfinished probe, then release every finished probe session that the harness permits. Only after that cleanup record the number of launches that started successfully, not the number of replacement attempts or late completions: ${capacityRecordCommand(context, input.run.id, input.projectRoot, "<successful-initial-launches>")}. Capacity probes are not Works and are never registered.`, "After dispatch, launch at most the measured capacity at once. If more independent reviewer Works remain, wait for the current wave to settle, then start the unchanged queued Works in the next wave. A reviewer launch rejected before it starts is not review evidence: do not create a replacement; wait for a running wave to settle and start that same queued Work. Each reviewer must be a genuinely fresh harness child Session. The lifecycle adapter binds that observed Session; do not bind or supply a Session ID manually. Reviewers are read-only and must not create children. As soon as a reviewer result is accepted, release that reviewer Session when the harness permits; do not let finished disposable workers occupy slots before the next wave.", "Review the execution environment of every selected check as part of its proof: a reset/fixture process, service process and client process must share the intended data and configuration world. A runtime entrypoint that can break that invariant must be explicit in one Work's task and verification and ordered before its consumer. planned_write_areas may advertise likely overlap, but do not treat them as ownership; required_read alone is not a delivery plan.", "If the final decision needs user input with no reasonable default, run this exact one-command heredoc, replacing only its placeholder body. The heredoc is the permitted stdin form; do not use cat, a pipe, a temporary file or a second shell command:", "```sh", input.pauseCommandTemplate, "```", "Ask the returned user_message, stop, then resume this same PLAN-REVIEW Work. Do not write decision.json or finish first.", `When all reviewer results are complete and every user question is resolved, classify every material finding, fix accepted findings in this same PLAN-REVIEW Work, then write ${decision} and finish: ${finishCommand(context, input.run.id, input.projectRoot, decision)}`, "Reviewer findings use local FIND-NNN ids. dd-flow exposes each finding to this coordinator as WRK-.../FIND-NNN; use that canonical finding_ref in the decision.", "A completed reviewer result with needs_changes or blocked is evidence, not the stage outcome. Classify its material findings and apply accepted fixes in this one review pass; do not start a second review automatically. Only a missing, malformed or unfinished reviewer result blocks the stage. For an accepted correction, increment PLAN revision and update only plan.json and the relevant aspect map. Do not edit or list code-work-batch.json: the CLI validates final PLAN and regenerates it. If no material correction is needed, set correction.status=not_required. The CLI checks mechanical handoff coherence; it does not prove semantic correctness.", "```json", JSON.stringify({ schema_id: "dd-flow/plan-review-decision@3", outcome: "accepted | failed | cancelled", summary: "Concise evidence-backed final decision.", finding_decisions: [{ finding_ref: "WRK-001-review/FIND-001", decision: "accepted_fix | rejected | deferred_as_DEF | requires_user | duplicate", reason: "Why." }], correction: { status: "not_required | applied", previous_plan_revision: revision, changed_paths: [], summary: "No material correction was needed, or summarize the applied correction." } }, null, 2), "```", "</execution_commands>", "", "<stage_instructions>", template, "</stage_instructions>", ""].join("\n");
 }
 function reviewGroups(home, workspaceRoot) {
     const root = path.join(home, "03-plan");
@@ -419,6 +428,14 @@ function reviewerEvidenceFailures(context, input) {
         const actual = result.aspects.map((item) => item.aspect_id);
         if (actual.length !== group.aspect_ids.length || new Set(actual).size !== actual.length || group.aspect_ids.some((id) => !actual.includes(id)) || result.aspects.some((item) => item.evidence_refs.length === 0))
             failures.push({ group: group.key, work_id: child.work_id, error: "reviewer_aspect_coverage_invalid", expected_aspects: group.aspect_ids, actual_aspects: actual });
+        for (const ref of result.aspects.flatMap((item) => [...item.evidence_refs, ...item.findings.flatMap((finding) => finding.evidence_refs)])) {
+            try {
+                assertPortableArtifactRef(ref, { workspaceRoot: input.workspaceRoot, runHome: input.runHome, runId: input.runId });
+            }
+            catch (error) {
+                failures.push({ group: group.key, work_id: child.work_id, error: "reviewer_evidence_ref_invalid", ref, detail: error instanceof Error ? error.message : String(error) });
+            }
+        }
         const workSession = context.db.get("SELECT session_id, status FROM work_sessions WHERE work_id = ? ORDER BY created_at DESC LIMIT 1", [child.work_id]);
         if (!workSession?.session_id || workSession.status !== "completed" || workSession.session_id === parent?.session_id)
             failures.push({ group: group.key, work_id: child.work_id, error: "reviewer_session_not_fresh", parent_session_id: parent?.session_id ?? null, reviewer_session_id: workSession?.session_id ?? null });
@@ -445,7 +462,7 @@ function finishTerminalReview(context, input) {
     context.db.run("UPDATE works SET status = ?, result = ?, completed_at = ?, updated_at = ? WHERE work_id = ? AND status = 'running'", [childStatus, input.decision.summary, now, now, rootWork]);
     closeRunningWorkSession(context, rootWork, childStatus, now);
     const settledChildren = context.db.all("SELECT work_id, task, status, result, created_at FROM works WHERE project_id = ? AND run_id = ? AND parent_work_id = ? AND task <> 'Capacity probe: return ready and finish this Work.'", [input.projectId, input.run.id, input.parentWorkId]);
-    const report = reportFor({ run: input.run, mode: input.contextFile.system?.effective_mode ?? "standard", requested: input.contextFile.system?.requested_mode ?? "auto", outcome: input.decision.outcome, groups: input.contextFile.groups ?? [], batchChecksum: input.contextFile.system?.batch_checksum ?? checksum(input.batch), planChecksum: input.contextFile.system?.plan_checksum ?? checksum(path.join(input.planRoot, "stage-report.json")), code: {}, now, projectRoot: input.projectRoot, decision: input.decision, children: latestReviewChildren(settledChildren), flow: flowCommand(context) });
+    const report = reportFor({ run: input.run, mode: input.contextFile.system?.effective_mode ?? "standard", requested: input.contextFile.system?.requested_mode ?? "auto", outcome: input.decision.outcome, groups: input.contextFile.groups ?? [], batchChecksum: input.contextFile.system?.batch_checksum ?? checksum(input.batch), planChecksum: input.contextFile.system?.plan_checksum ?? planSetChecksum(requireHome(input.run), input.run.workspace_root), code: {}, now, projectRoot: input.projectRoot, decision: input.decision, children: latestReviewChildren(settledChildren), flow: flowCommand(context) });
     writeReport(input.root, report);
     refreshRunWorkProjection(context, input.projectId, input.run.id);
     const stageStatus = input.decision.outcome === "cancelled" ? "skipped" : input.decision.outcome === "failed" ? "failed" : "blocked";
@@ -497,3 +514,4 @@ catch {
 } }
 function writeJson(file, value) { const temporary = `${file}.${crypto.randomUUID()}.tmp`; fs.writeFileSync(temporary, `${JSON.stringify(value, null, 2)}\n`); fs.renameSync(temporary, file); }
 function checksum(file) { return crypto.createHash("sha256").update(fs.readFileSync(file)).digest("hex"); }
+function planSetChecksum(home, workspaceRoot) { const entries = protocolIdsForReview(home).sort().map((protocolId) => { const file = path.join(workspaceRoot, ".memory-bank", "protocol", protocolId, "plan.json"); return { protocol_id: protocolId, path: path.relative(workspaceRoot, file).split(path.sep).join("/"), sha256: checksum(file) }; }); return crypto.createHash("sha256").update(JSON.stringify(entries)).digest("hex"); }