@deksden-com/dd-flow-cli 0.9.0-beta.1 → 0.9.0-beta.7

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"); }

package/dist/services/vnext-plan.js

@@ -2,7 +2,7 @@ import crypto from "node:crypto";
 import fs from "node:fs";
 import path from "node:path";
 import { AppError } from "../shared/errors.js";
-import { validateCheckDeclaration, validateCheckPlacement, validateCodeCheckCommands } from "./code-checks.js";
+import { effectiveCheckDeclarations, readCodeCheckProfile, validateCheckDeclaration, validateCheckPlacement, validateCodeCheckCommands } from "./code-checks.js";
 import { requireProjectByRoot } from "./projects.js";
 import { resolveProjectRoot } from "../storage/paths.js";
 import { advanceFlowRun, appendFlowRunTimelineEvent, attachFlowRunStage, completeFlowRunStage, getFlowRunVariables, gitFacts } from "./runs.js";
@@ -29,13 +29,18 @@ export function startVnextPlan(context, input) {
     const home = requireHome(run);
     assertStageStartHookEvent(context, { projectId: project.id, eventKey: input.hookEventId, runId: run.id, stage: "plan", projectRoot, ...(input.contextSha256 ? { contextSha256: input.contextSha256 } : {}) });
     const workspaceRoute = requireVnextWorkspaceRoute({ projectRoot, runId: run.id, runHome: home, workspaceRoot: run.workspace_root, stage: "plan" });
-    const root = path.join(home, vnextStageDirectory("plan"));
-    fs.mkdirSync(root, { recursive: true });
     const protocols = protocolIds(home);
     if (!protocols.length)
         throw new AppError("not_found", "PLAN requires accepted PROTOCOLIZE protocols", 1);
+    // Static inputs are checked before PLAN creates a Work or materializes a
+    // draft. A rejected start is therefore side-effect free and safe to retry.
+    const template = read(path.join(projectRoot, ".memory-bank", "dd-flow", "vnext", "plan.md"));
+    assertProtocolWorkspace(run.workspace_root, protocols);
+    const { profile: codeCheckProfile } = readCodeCheckProfile(run.workspace_root);
     if (context.db.get("SELECT 1 FROM works WHERE project_id = ? AND run_id = ? AND task = ? AND status = 'running'", [project.id, run.id, planTask]))
         throw new AppError("invalid_work_state", "PLAN already has a running Work", 1, { run_id: run.id });
+    const root = path.join(home, vnextStageDirectory("plan"));
+    fs.mkdirSync(root, { recursive: true });
     const now = context.now();
     const rootWork = context.db.get("SELECT work_id FROM works WHERE project_id = ? AND run_id = ? AND parent_work_id IS NULL ORDER BY created_at LIMIT 1", [project.id, run.id]);
     if (!rootWork)
@@ -56,11 +61,9 @@ export function startVnextPlan(context, input) {
         context.db.exec("ROLLBACK");
         throw error;
     }
-    const template = read(path.join(projectRoot, ".memory-bank", "dd-flow", "vnext", "plan.md"));
     // A Desktop task may start above the materialized repository.  Lifecycle
     // prompts therefore hand agents write targets as absolute paths: relative
     // `.memory-bank/...` paths would otherwise silently land in the parent cwd.
-    assertProtocolWorkspace(run.workspace_root, protocols);
     const planPaths = protocols.map((id) => path.join(run.workspace_root, ".memory-bank", "protocol", id, "plan.json"));
     const mapPaths = protocols.map((id) => `${path.join(root, id, "aspect-map.json")}`);
     const owned = protocolOwnership(home, protocols);
@@ -76,12 +79,19 @@ export function startVnextPlan(context, input) {
     ]);
     const runVariables = getFlowRunVariables(context, { projectRoot, runId: run.id });
     const measuredCapacity = runVariables.variables[subagentCapacityKey];
+    const mergeRequired = runEndsAtMerge(context, projectRoot, run.id);
     const capacityContext = typeof measuredCapacity === "number" && Number.isInteger(measuredCapacity) && measuredCapacity >= 0
         ? `- The measured reviewer capacity is ${measuredCapacity}. This is a runtime fact for later PLAN-REVIEW dispatch; do not repeat the probe or invent a different value.`
         : "- Reviewer capacity is not measured yet. PLAN must not probe or launch reviewers; PLAN-REVIEW will measure it once if review is enabled.";
     const reviewGroupingRule = "Group only semantically compatible applicable aspects, preserving real trust, irreversible, high-risk and hard-dependency boundaries. Prefer the fewest groups that retain independent review value, normally one review wave. Put two or three compatible aspects in a group; do not create one group per aspect merely for convenience. A later PLAN-REVIEW dispatch measures current capacity once and schedules these semantic groups into waves; do not invent a capacity value here.";
     const checkProfile = path.join(run.workspace_root, ".memory-bank", "spec", "engineering", "code-check-profile.json");
-    const prompt = ["<stage_identity>", `- RUN: ${run.id}`, `- Work: ${planWorkId}`, "- stage: plan", "</stage_identity>", "", "<trusted_runtime_context>", "These facts were collected by dd-flow. Trust them; do not repeat CLI, Git, compatibility or permission discovery.", `- Project root: ${projectRoot}`, `- Workspace: ${run.workspace_root}`, `- Stage workspace: ${root}`, `- Git: ${JSON.stringify(gitFacts(run.workspace_root))}`, capacityContext, "</trusted_runtime_context>", "", "<workspace_contract>", `- route: ${workspaceRoute.route}`, `- feature branch: ${workspaceRoute.feature_branch ?? "not applicable"}`, `- base commit: ${workspaceRoute.base_ref ?? "not applicable"}`, `- write workspace: ${run.workspace_root}`, "The CLI has verified this frozen route. All project reads and writes for PLAN and later CODE happen in the write workspace; project root is only the stable runtime identity for lifecycle commands. Do not create, switch, merge or delete branches/worktrees.", "Keep the task runner's current cwd. Use the absolute paths in this packet instead of trying to set the provisioned workspace as a tool workdir.", "</workspace_contract>", "", "<accepted_inputs>", `- ${path.join(home, "01-specify", "specify.json")}`, `- ${path.join(home, "02-protocolize", "protocolize-result.json")}`, ...protocols.map((id) => `- ${path.join(run.workspace_root, ".memory-bank", "protocol", id, "summary.md")}`), "</accepted_inputs>", "", ...(fs.existsSync(checkProfile) ? ["<code_check_policy>", "You, not the CLI, select evidence for every accepted requirement and acceptance criterion. The profile only lists reusable aliases, mandatory project policy gates and guarded raw command prefixes. Inspect relevant package/test manifests before choosing a check. Do not classify checks by weight and do not omit a needed check because it looks expensive.", fs.readFileSync(checkProfile, "utf8").trim(), "</code_check_policy>", ""] : []), "<artifacts>", "The CLI has already materialized every artifact below as a partially filled draft. Edit these files in place; do not create replacements elsewhere.", "Prefilled and CLI-owned plan fields: schema_id, plan_id, protocol_id, initial revision and source_refs.", "Prefilled and CLI-owned aspect-map fields: schema_id, protocol_id, plan_id, plan revision, catalog_ref and every catalog aspect_id.", "You own the remaining semantic fields. Empty or missing semantic values are intentional draft markers and must be completed before validation.", ...planPaths.map((value) => `- partially filled plan: ${value}`), ...mapPaths.map((value) => `- partially filled aspect map: ${value}`), "</artifacts>", "", "<output_contract>", "Complete every named plan and aspect map in place. Do not create or edit code-work-batch.json: dd-flow derives it after validation.", "The CLI owns schema_id, plan_id, protocol_id, revision and source_refs. Preserve them exactly.", "Use protocol-plan@6. Its top-level checks[] is the single check catalog. Every check has id, command, purpose, run_at and availability. available means executable now. planned means one named P* Work first creates a NEW @check/... alias: planned therefore always needs provided_by and the exact alias definition. Every semantic @check alias, including an existing one, repeats its exact accepted profile command in definition so later stages can detect drift. Items and acceptance entries use check_refs only; never duplicate command declarations.", "For each R-* and AC-*, choose an actually relevant proof: an existing focused test, a new planned alias plus its provider Work, a project policy gate, or an honestly limited external/manual proof. Every plan item needs at least one check_ref. The CLI validates ids, provider ordering, materialization and guarded command policy; it never chooses a check for you. A provider Work may verify itself with the alias it has just created. A consumer must depend on that provider.", "Each plan item must name concrete existing source/test paths in required_read. planned_write_areas is optional: use stable component directories or files only when they help coordinate parallel Work; it is never a write allowlist. Reference every owned R-* and AC-* in one or more items; every AC-* needs an observable acceptance proof.", "For every selected check, inspect its command's launch path and the runtime entrypoints it starts. The fixture/reset process, service process and client process must observe one intended environment and data world. If a required runtime entrypoint needs a code change, make that change explicit in the Work task and its verification. Use planned_write_areas only to advertise likely concurrent overlap; do not treat it as ownership or assume another Work will repair an omitted change. If an independent infrastructure Work is clearer, plan that Work explicitly and order consumers after it.", reviewGroupingRule, "Complete compact contract and schema paths:", `- protocol plan schema: ${path.join(run.workspace_root, ".memory-bank", "dd-flow", "schemas", "vnext-protocol-plan.schema.json")}`, `- aspect map schema: ${path.join(run.workspace_root, ".memory-bank", "dd-flow", "schemas", "plan-aspect-map.schema.json")}`, "Minimal valid protocol-plan shape:", "```json", JSON.stringify(planExample(protocols[0]), null, 2), "```", "Minimal valid aspect-map shape:", "```json", JSON.stringify(aspectMapExample(protocols[0]), null, 2), "```", "</output_contract>", "", "<execution_commands>", "PLAN never launches independent reviewers or registers CODE Work.", "If PLAN needs a material user decision 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", pauseCommandTemplate, "```", "Ask the returned user_message, stop, and resume this same PLAN Work with the exact returned command.", "Validate both partially filled drafts after completing their semantic fields:", ...validationCommands.map((command) => `- ${command}`), "Finish PLAN only after all questions are resolved and both validation commands pass:", finishCommand, "The response returns the only PLAN-REVIEW start command. Follow it; do not start CODE directly.", "</execution_commands>", "", "<stage_instructions>", template, "</stage_instructions>", ""].join("\n");
+    const policyMergeAliases = codeCheckProfile?.mandatory_by_gate.merge ?? [];
+    const mergeContract = mergeRequired
+        ? ["<merge_gate_contract>", ...(policyMergeAliases.length
+                ? [`This RUN must reach MERGE. Project policy already supplies the mandatory merge gate${policyMergeAliases.length === 1 ? "" : "s"}: ${policyMergeAliases.join(", ")}. Do not duplicate them in semantic checks[]. Add another merge check only when the task genuinely needs additional evidence.`]
+                : ["This RUN must reach MERGE and project policy supplies no merge gate. Select at least one real top-level checks[] entry with run_at: merge. It may use an existing project alias or a planned alias materialised by a named P* provider Work. This is a planning obligation: do not defer it to CODE-REVIEW or MERGE."]), "The CLI validates the effective merge gate but never invents one or migrates an incompatible project policy.", "</merge_gate_contract>", ""]
+        : [];
+    const prompt = ["<stage_identity>", `- RUN: ${run.id}`, `- Work: ${planWorkId}`, "- stage: plan", "</stage_identity>", "", "<trusted_runtime_context>", "These facts were collected by dd-flow. Trust them; do not repeat CLI, Git, compatibility or permission discovery.", `- Project root: ${projectRoot}`, `- Workspace: ${run.workspace_root}`, `- Stage workspace: ${root}`, `- Git: ${JSON.stringify(gitFacts(run.workspace_root))}`, capacityContext, "</trusted_runtime_context>", "", "<workspace_contract>", `- route: ${workspaceRoute.route}`, `- feature branch: ${workspaceRoute.feature_branch ?? "not applicable"}`, `- base commit: ${workspaceRoute.base_ref ?? "not applicable"}`, `- write workspace: ${run.workspace_root}`, "The CLI has verified this frozen route. All project reads and writes for PLAN and later CODE happen in the write workspace; project root is only the stable runtime identity for lifecycle commands. Do not create, switch, merge or delete branches/worktrees.", "Keep the task runner's current cwd. Use the absolute paths in this packet instead of trying to set the provisioned workspace as a tool workdir.", "</workspace_contract>", "", "<accepted_inputs>", `- ${path.join(home, "01-specify", "specify.json")}`, `- ${path.join(home, "02-protocolize", "protocolize-result.json")}`, ...protocols.map((id) => `- ${path.join(run.workspace_root, ".memory-bank", "protocol", id, "summary.md")}`), "</accepted_inputs>", "", ...(fs.existsSync(checkProfile) ? ["<code_check_policy>", "You, not the CLI, select evidence for every accepted requirement and acceptance criterion. The profile only lists reusable aliases, mandatory project policy gates and guarded raw command prefixes. Inspect relevant package/test manifests before choosing a check. Do not classify checks by weight and do not omit a needed check because it looks expensive.", fs.readFileSync(checkProfile, "utf8").trim(), "</code_check_policy>", ""] : []), ...mergeContract, "<artifacts>", "The CLI has already materialized every artifact below as a partially filled draft. Edit these files in place; do not create replacements elsewhere.", "Prefilled and CLI-owned plan fields: schema_id, plan_id, protocol_id, initial revision and source_refs.", "Prefilled and CLI-owned aspect-map fields: schema_id, protocol_id, plan_id, plan revision, catalog_ref and every catalog aspect_id.", "You own the remaining semantic fields. Empty or missing semantic values are intentional draft markers and must be completed before validation.", ...planPaths.map((value) => `- partially filled plan: ${value}`), ...mapPaths.map((value) => `- partially filled aspect map: ${value}`), "</artifacts>", "", "<output_contract>", "Complete every named plan and aspect map in place. Do not create or edit code-work-batch.json: dd-flow derives it after validation.", "The CLI owns schema_id, plan_id, protocol_id, revision and source_refs. Preserve them exactly.", "Use protocol-plan@6. Its top-level checks[] is the single check catalog. Every check has id, command, purpose, run_at and availability. available means executable now. planned means one named P* Work first creates a NEW @check/... alias: planned therefore always needs provided_by and the exact alias definition. Every semantic @check alias, including an existing one, repeats its exact accepted profile command in definition so later stages can detect drift. Items and acceptance entries use check_refs only; never duplicate command declarations.", "For each R-* and AC-*, choose an actually relevant proof: an existing focused test, a new planned alias plus its provider Work, a project policy gate, or an honestly limited external/manual proof. Every plan item needs at least one check_ref. The CLI validates ids, provider ordering, materialization and guarded command policy; it never chooses a check for you. A provider Work may verify itself with the alias it has just created. A consumer must depend on that provider.", "Each plan item must name concrete existing source/test paths in required_read. planned_write_areas is optional: use stable component directories or files only when they help coordinate parallel Work; it is never a write allowlist. Reference every owned R-* and AC-* in one or more items; every AC-* needs an observable acceptance proof.", "For every selected check, inspect its command's launch path and the runtime entrypoints it starts. The fixture/reset process, service process and client process must observe one intended environment and data world. If a required runtime entrypoint needs a code change, make that change explicit in the Work task and its verification. Use planned_write_areas only to advertise likely concurrent overlap; do not treat it as ownership or assume another Work will repair an omitted change. If an independent infrastructure Work is clearer, plan that Work explicitly and order consumers after it.", reviewGroupingRule, "Complete compact contract and schema paths:", `- protocol plan schema: ${path.join(run.workspace_root, ".memory-bank", "dd-flow", "schemas", "vnext-protocol-plan.schema.json")}`, `- aspect map schema: ${path.join(run.workspace_root, ".memory-bank", "dd-flow", "schemas", "plan-aspect-map.schema.json")}`, "Minimal valid protocol-plan shape:", "```json", JSON.stringify(planExample(protocols[0]), null, 2), "```", "Minimal valid aspect-map shape:", "```json", JSON.stringify(aspectMapExample(protocols[0]), null, 2), "```", "</output_contract>", "", "<execution_commands>", "PLAN never launches independent reviewers or registers CODE Work.", "If PLAN needs a material user decision 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", pauseCommandTemplate, "```", "Ask the returned user_message, stop, and resume this same PLAN Work with the exact returned command.", "Validate both partially filled drafts after completing their semantic fields:", ...validationCommands.map((command) => `- ${command}`), "Finish PLAN only after all questions are resolved and both validation commands pass:", finishCommand, "The response returns the only PLAN-REVIEW start command. Follow it; do not start CODE directly.", "</execution_commands>", "", "<stage_instructions>", template, "</stage_instructions>", ""].join("\n");
     const artifactMaterialization = { status: "materialized", completeness: "partially_filled", plan_paths: planPaths, aspect_map_paths: mapPaths, cli_owned_plan_fields: ["schema_id", "plan_id", "protocol_id", "revision", "source_refs"], cli_owned_aspect_map_fields: ["schema_id", "protocol_id", "plan_id", "plan_revision", "catalog_ref", "aspects[].aspect_id"], validation_commands: validationCommands };
     const promptPath = path.join(root, "stage-prompt.md");
     fs.writeFileSync(promptPath, prompt);
@@ -177,6 +187,15 @@ export function validateVnextPlanArtifacts(context, input) {
             failures.push(validationFailure(file, error));
         }
     }
+    if (!failures.length) {
+        try {
+            validatePsetCheckIdentity(plans);
+            validateRequiredMergeGate(context, input, plans);
+        }
+        catch (error) {
+            failures.push(validationFailure(batch, error));
+        }
+    }
     if (!failures.length) {
         const temporaryBatch = `${batch}.tmp-${crypto.randomUUID()}`;
         try {
@@ -285,7 +304,7 @@ function stageStartedAt(home, fallback) { try {
 catch {
     return fallback;
 } }
-function validationFailure(file, error) { return { file, message: error instanceof Error ? error.message : String(error), ...(error instanceof AppError ? { details: error.details } : {}) }; }
+function validationFailure(file, error) { return { file, message: error instanceof Error ? error.message : String(error), ...(error instanceof AppError ? { code: error.code, details: error.details } : {}) }; }
 function acceptedObligations(home) {
     const file = path.join(home, "01-specify", "specify.json");
     if (!fs.existsSync(file))
@@ -395,6 +414,7 @@ function projectCodeWorkBatch(input) {
             // stale as soon as its provider performs its declared work.
             required_read: [...new Set([
                     ...orientation,
+                    path.relative(input.workspaceRoot, file).split(path.sep).join("/"),
                     ...item.execution_context.required_read,
                     ...existingDocumentPaths,
                     ...(value.checks.some((check) => check.availability === "planned" && check.provided_by === item.id)
@@ -547,9 +567,14 @@ function validatePlanSemantics(file, ownedRefs, acceptedRefs) {
         for (const id of acceptance.plan_item_ids ?? [])
             if (!ids.has(id))
                 throw new AppError("validation", "PLAN acceptance references an unknown item", 2, { file, criterion_id: acceptance.criterion_id, plan_item_id: id });
-        for (const id of acceptance.check_refs ?? [])
-            if (!checks.has(id))
+        for (const id of acceptance.check_refs ?? []) {
+            const check = checks.get(id);
+            if (!check)
                 throw new AppError("check_reference_unknown", "PLAN acceptance references an unknown check", 2, { file, criterion_id: acceptance.criterion_id, check_id: id });
+            for (const itemId of acceptance.plan_item_ids ?? [])
+                if (check.availability === "planned" && check.provided_by !== itemId && !ancestors(itemId).has(check.provided_by))
+                    throw new AppError("check_consumer_not_ordered_after_provider", "Acceptance may consume a planned check only after its provider", 2, { file, criterion_id: acceptance.criterion_id, item: itemId, check_id: id, provider: check.provided_by });
+        }
     }
     for (const obligation of ownedRefs)
         if (!realized.has(obligation))
@@ -558,3 +583,26 @@ function validatePlanSemantics(file, ownedRefs, acceptedRefs) {
         if (!plan.acceptance.some((acceptance) => acceptance.criterion_id === obligation))
             throw new AppError("validation", "Every owned AC-* needs an observable PLAN acceptance entry", 2, { file, criterion_id: obligation });
 }
+function validatePsetCheckIdentity(plans) {
+    const owners = new Map();
+    for (const plan of plans)
+        for (const check of plan.value.checks) {
+            const prior = owners.get(check.id);
+            if (prior)
+                throw new AppError("duplicate_pset_check_id", "PLAN check ids must be unique across the whole PSET", 2, { check_id: check.id, protocols: [prior, plan.protocolId] });
+            owners.set(check.id, plan.protocolId);
+        }
+}
+function runEndsAtMerge(context, projectRoot, runId) {
+    const project = requireProjectByRoot(context, resolveProjectRoot(projectRoot));
+    const row = context.db.get("SELECT index_json FROM runs WHERE project_id = ? AND id = ?", [project.id, runId]);
+    return JSON.parse(row?.index_json ?? "{}").execution_profile?.settings?.stop_target === "merge_completed";
+}
+function validateRequiredMergeGate(context, input, plans) {
+    if (!runEndsAtMerge(context, input.projectRoot, input.runId))
+        return;
+    const declared = plans.flatMap(({ value }) => value.checks);
+    if (effectiveCheckDeclarations(input.workspaceRoot ?? input.projectRoot, declared, ["merge"]).length === 0) {
+        throw new AppError("merge_gate_plan_missing", "PLAN for a RUN ending in MERGE must declare at least one semantic or project-policy merge check", 2, { run_id: input.runId });
+    }
+}