@deksden-com/dd-flow-cli 0.9.0-beta.1 → 0.9.0-beta.10

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

@@ -14,14 +14,12 @@ 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";
+import { readFanoutDescriptor, subagentCapacityKey, writeFanoutDescriptor } from "./vnext-fanout.js";
 const stage = "plan-review";
 const stageDir = vnextStageDirectory(stage);
 const reviewTask = "Independently review the accepted PLAN and decide whether CODE may open.";
-const capacityProbeFanoutSize = capacityProbe.fanout_size;
-const capacityProbeHoldSeconds = capacityProbe.probe_hold_seconds;
-const capacityProbeDeadlineSeconds = capacityProbe.cleanup_deadline_seconds;
 export function isVnextPlanReviewRun(context, input) {
     const project = requireProjectByRoot(context, resolveProjectRoot(input.projectRoot));
     return Boolean(context.db.get("SELECT id FROM runs WHERE project_id = ? AND id = ? AND flow_kind = 'vnext_protocolize'", [project.id, input.runId]));
@@ -48,7 +46,7 @@ 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 +59,7 @@ 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)
@@ -125,13 +123,13 @@ export function dispatchVnextPlanReview(context, input) {
     if (capacity.pending)
         return capacity.response;
     if (capacity.availableSlots < 1)
-        throw new AppError("no_subagent_capacity", "PLAN-REVIEW requires one fresh reviewer Session but no probe launch succeeded", 1, { run_id: run.id });
+        throw new AppError("subagent_capacity_unqualified", "PLAN-REVIEW requires one qualified native subagent slot", 1, { run_id: run.id });
     const existing = context.db.all("SELECT work_id, task, status, result, created_at, launch_policy FROM works WHERE project_id = ? AND run_id = ? AND parent_work_id = ?", [project.id, run.id, parent.work_id]);
     const latest = (group) => existing.find((item) => item.task === group.task);
     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 });
         }
@@ -161,14 +159,14 @@ export function recordVnextPlanReviewCapacity(context, input) {
     const prior = variables.variables[reviewCapacityKey];
     if (typeof prior === "number") {
         if (prior === input.availableSlots)
-            return { ok: true, run_id: run.id, capacity: { available_slots: prior, source: "bounded_probe", reused: true } };
+            return { ok: true, run_id: run.id, capacity: { available_slots: prior, source: "harness_qualification", reused: true } };
         setRuntimeRunVariable(context, { projectRoot, runId: run.id, key: reviewCapacityKey, value: input.availableSlots });
-        appendFlowRunTimelineEvent(context, project.id, run.id, { type: "subagent_capacity_refreshed", previous_available_slots: prior, available_slots: input.availableSlots, fanout_size: capacityProbeFanoutSize });
-        return { ok: true, run_id: run.id, capacity: { available_slots: input.availableSlots, previous_available_slots: prior, source: "bounded_probe", refreshed: true } };
+        appendFlowRunTimelineEvent(context, project.id, run.id, { type: "subagent_capacity_refreshed", previous_available_slots: prior, available_slots: input.availableSlots, source: "harness_qualification" });
+        return { ok: true, run_id: run.id, capacity: { available_slots: input.availableSlots, previous_available_slots: prior, source: "harness_qualification", refreshed: true } };
     }
     setRuntimeRunVariable(context, { projectRoot, runId: run.id, key: reviewCapacityKey, value: input.availableSlots });
-    appendFlowRunTimelineEvent(context, project.id, run.id, { type: "subagent_capacity_observed", available_slots: input.availableSlots, fanout_size: capacityProbeFanoutSize });
-    return { ok: true, run_id: run.id, capacity: { available_slots: input.availableSlots, source: "one_shot_fanout", fanout_size: capacityProbeFanoutSize } };
+    appendFlowRunTimelineEvent(context, project.id, run.id, { type: "subagent_capacity_observed", available_slots: input.availableSlots, source: "harness_qualification" });
+    return { ok: true, run_id: run.id, capacity: { available_slots: input.availableSlots, source: "harness_qualification" } };
 }
 export function finishVnextPlanReview(context, input) {
     const projectRoot = resolveProjectRoot(input.projectRoot);
@@ -182,22 +180,26 @@ export function finishVnextPlanReview(context, input) {
     if (!parent || parent.status !== "running")
         throw new AppError("invalid_work_state", "PLAN-REVIEW has no running parent Work", 2);
     const decisionPath = path.resolve(input.decisionFile ?? path.join(root, "decision.json"));
-    const decision = readDecision(context, projectRoot, run.id, decisionPath);
-    const children = 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.'", [project.id, run.id, parent.work_id]);
+    const decision = readDecision(context, projectRoot, run.id, home, decisionPath);
+    const children = context.db.all("SELECT work_id, task, status, result, created_at FROM works WHERE project_id = ? AND run_id = ? AND parent_work_id = ?", [project.id, run.id, parent.work_id]);
     const contextFile = readJson(path.join(root, "work-context.json"));
     const groups = contextFile.groups ?? [];
     const reviewedPlanRevision = contextFile.system?.plan_revision;
     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 +211,21 @@ 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 +239,7 @@ 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 +264,7 @@ 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 reports qualified_capacity_required, stop. The external harness controller qualifies the selected profile outside this RUN and records the resulting integer with ${capacityRecordCommand(context, input.run.id, input.projectRoot, "<qualified-native-child-count>")}. PLAN-REVIEW never launches a capacity probe.`, "After dispatch, launch at most the qualified 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");
@@ -318,15 +324,15 @@ function plannedReviewWaves(groups, capacity) {
     }
     return waves;
 }
-export function capacityProbeResponse(context, input) {
-    return { ok: true, run_id: input.runId, stage, outcome: "capacity_probe_required", capacity_probe: { ...capacityProbe, controller_instruction: "Launch exactly one concurrent batch of 15 independent leaf probes. Use all-settled behavior: every initial launch is attempted once, and a launch rejection is expected evidence of the current limit. Do not retry, replace, or add probes. Count the initial launches that returned a live session handle. Successful probes must not call tools, read files, create children or run dd-flow; they wait 60 seconds then return exactly AGENT-NN. Wait only for cleanup, at most 180 seconds from first launch; terminate unfinished probes, then close/delete every finished probe session that the harness permits. Record only the initial successful-launch count after cleanup." }, next_action: "run_one_shot_capacity_probe_then_retry_dispatch", next: { record_command: capacityRecordCommand(context, input.runId, input.projectRoot, "<successful-initial-launches>"), retry_command: dispatchCommand(context, input.runId, input.projectRoot) } };
+export function capacityQualificationResponse(context, input) {
+    return { ok: true, run_id: input.runId, stage, outcome: "qualified_capacity_required", next_action: "qualify_harness_profile_then_record_capacity", next: { record_command: capacityRecordCommand(context, input.runId, input.projectRoot, "<qualified-native-child-count>"), retry_command: dispatchCommand(context, input.runId, input.projectRoot) } };
 }
 function ensureReviewCapacity(context, input) {
     const variables = getFlowRunVariables(context, { projectRoot: input.projectRoot, runId: input.runId });
     const known = variables.variables[reviewCapacityKey];
     if (typeof known === "number" && Number.isInteger(known) && known >= 0)
         return { pending: false, availableSlots: known };
-    return { pending: true, response: capacityProbeResponse(context, { projectRoot: input.projectRoot, runId: input.runId }) };
+    return { pending: true, response: capacityQualificationResponse(context, { projectRoot: input.projectRoot, runId: input.runId }) };
 }
 function protocolIdsForReview(home) {
     const report = readJson(path.join(home, "02-protocolize", "stage-report.json"));
@@ -376,7 +382,7 @@ catch {
 function rootWorkId(context, projectId, runId) { const root = 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", [projectId, runId]); if (!root)
     throw new AppError("runtime_missing", "vNext RUN has no root Work", 1); return root.work_id; }
 function registerCode(context, projectId, runId, batch) { return addWorkBatch(context, { parentWorkId: rootWorkId(context, projectId, runId), file: batch }); }
-function readDecision(context, projectRoot, runId, file) { validateSchema({ schemaName: "plan-review-decision", file, projectRoot, runId, ddFlowHome: context.ddFlowHome }); return readJson(file); }
+function readDecision(context, projectRoot, runId, runRoot, file) { validateSchema({ schemaName: "plan-review-decision", file, projectRoot, runId, runRoot, ddFlowHome: context.ddFlowHome }); return readJson(file); }
 function preserveDecisionReceipt(root, source) {
     const receipt = path.join(root, "decision.receipt.json");
     const bytes = fs.readFileSync(source);
@@ -419,6 +425,14 @@ 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 });
@@ -437,15 +451,15 @@ function finishTerminalReview(context, input) {
     const now = context.now();
     const childStatus = input.decision.outcome === "cancelled" ? "cancelled" : "failed";
     // A terminal decision is structured cancellation of the review subtree. It
-    // must settle probes as well as semantic reviewers before its parent Work
-    // can settle; otherwise the RUN contains abandoned active Work/Session links.
+    // must settle semantic reviewers before its parent Work can settle; otherwise
+    // the RUN contains abandoned active Work/Session links.
     settleReviewChildren(context, { projectId: input.projectId, runId: input.run.id, parentWorkId: input.parentWorkId, status: childStatus, reason: `PLAN-REVIEW terminal decision: ${input.decision.summary}`, now });
     settleReviewParent(context, { projectRoot: input.projectRoot, projectId: input.projectId, runId: input.run.id, parentWorkId: input.parentWorkId, status: childStatus, result: input.decision.summary, reason: `plan_review_${input.decision.outcome}`, resultPath: path.join(input.root, "stage-report.json") });
     const rootWork = rootWorkId(context, input.projectId, input.run.id);
     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 settledChildren = context.db.all("SELECT work_id, task, status, result, created_at FROM works WHERE project_id = ? AND run_id = ? AND parent_work_id = ?", [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 ?? 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";
@@ -483,10 +497,10 @@ function codeCommand(context, runId, projectRoot) { return `${flowCommand(contex
 function dispatchCommand(context, runId, projectRoot) { return `${flowCommand(context)} plan-review dispatch ${runId} --project-root ${JSON.stringify(projectRoot)} --json`; }
 function finishCommand(context, runId, projectRoot, decision) { return `${flowCommand(context)} stage finish ${runId} --stage plan-review --project-root ${JSON.stringify(projectRoot)} --decision-file ${JSON.stringify(decision)} --json`; }
 function capacityRecordCommand(context, runId, projectRoot, availableSlots) { return `${flowCommand(context)} run capacity record ${runId} --available-slots ${availableSlots} --project-root ${JSON.stringify(projectRoot)} --json`; }
-function requireRun(context, projectRoot, runId) { const project = requireProjectByRoot(context, projectRoot); const run = context.db.get("SELECT id, project_id, workspace_root, run_home_path, index_json FROM runs WHERE project_id = ? AND id = ?", [project.id, runId]); if (!run)
+function requireRun(context, projectRoot, runId) { const project = requireProjectByRoot(context, projectRoot); const run = context.db.get("SELECT id, project_id, workspace_root, run_root, index_json FROM runs WHERE project_id = ? AND id = ?", [project.id, runId]); if (!run)
     throw new AppError("not_found", "RUN is not registered", 1); return run; }
-function requireHome(run) { if (!run.run_home_path)
-    throw new AppError("runtime_missing", "RUN workspace is unavailable", 1); return run.run_home_path; }
+function requireHome(run) { if (!run.run_root)
+    throw new AppError("runtime_missing", "RUN artifact root is unavailable", 1); return run.run_root; }
 function read(file) { if (!fs.existsSync(file))
     throw new AppError("not_found", "Required vNext prompt is missing", 1, { file }); return fs.readFileSync(file, "utf8"); }
 function readJson(file) { try {
@@ -497,3 +511,4 @@ catch {
 } }
 function writeJson(file, value) { const temporary = `${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) => `${path.join(root, id, "aspect-map.json")}`);
     const owned = protocolOwnership(home, protocols);
@@ -71,17 +74,24 @@ export function startVnextPlan(context, input) {
     const pauseCommand = stagePauseCommand(context, { runId: run.id, stage: "plan", workId: planWorkId, projectRoot });
     const pauseCommandTemplate = stagePauseCommandTemplate(pauseCommand);
     const validationCommands = protocols.flatMap((_, index) => [
-        `${flowCommand(context)} schema validate --schema vnext-protocol-plan --file ${JSON.stringify(planPaths[index])} --project-root ${JSON.stringify(run.workspace_root)} --json`,
-        `${flowCommand(context)} schema validate --schema plan-aspect-map --file ${JSON.stringify(mapPaths[index])} --project-root ${JSON.stringify(run.workspace_root)} --json`
+        `${flowCommand(context)} schema validate --schema vnext-protocol-plan --file ${JSON.stringify(planPaths[index])} --project-root ${JSON.stringify(run.workspace_root)} --run ${run.id} --json`,
+        `${flowCommand(context)} schema validate --schema plan-aspect-map --file ${JSON.stringify(mapPaths[index])} --project-root ${JSON.stringify(run.workspace_root)} --run ${run.id} --json`
     ]);
     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.";
+        ? `- The qualified reviewer capacity is ${measuredCapacity}. This is a runtime fact for later PLAN-REVIEW dispatch; do not repeat qualification or invent a different value.`
+        : "- Reviewer capacity is not qualified yet. PLAN must not qualify or launch reviewers; an external harness controller supplies it before fan-out.";
+    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 uses externally qualified capacity to schedule 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) => `- ${command}`), "Finish PLAN only after all questions are resolved and both validation commands pass:", finishCommand, "The response returns the only PLAN-REVIEW start command. Follow it; do not start CODE directly.", "</execution_commands>", "", "<stage_instructions>", template, "</stage_instructions>", ""].join("\n");
     const artifactMaterialization = { status: "materialized", completeness: "partially_filled", plan_paths: planPaths, aspect_map_paths: mapPaths, cli_owned_plan_fields: ["schema_id", "plan_id", "protocol_id", "revision", "source_refs"], cli_owned_aspect_map_fields: ["schema_id", "protocol_id", "plan_id", "plan_revision", "catalog_ref", "aspects[].aspect_id"], validation_commands: validationCommands };
     const promptPath = path.join(root, "stage-prompt.md");
     fs.writeFileSync(promptPath, prompt);
@@ -126,7 +136,7 @@ export function finishVnextPlan(context, input) {
     const reviewCommand = `${flowCommand(context)} stage start ${run.id} --stage plan-review --project-root ${JSON.stringify(projectRoot)} --json`;
     const report = { schema_id: "dd-flow/stage-report@1", run_id: run.id, stage: "plan", generated_at: now, verdict: "done", semantic: { result: `Accepted ${protocols.length} executable PLAN artifact${protocols.length === 1 ? "" : "s"}.`, acceptance: protocols, changed_files: [...planFiles.map((file) => path.relative(projectRoot, file)), ...mapFiles.map((file) => runRef(run.id, home, file)), runRef(run.id, home, batch)], checks: ["protocol-plan schema", "aspect-map schema", "cross-artifact references", "generated CODE batch"], evidence: [runRef(run.id, home, path.join(root, "stage-report.json"))], next_action: "start_plan_review", plans: planFiles.map((file) => path.relative(projectRoot, file)), aspect_maps: mapFiles.map((file) => runRef(run.id, home, file)), code_work_batch: runRef(run.id, home, batch), batch_checksum: batchChecksum }, mechanical: { started_at: stageStartedAt(home, now), finished_at: now, wall_clock_ms: Math.max(0, Date.parse(now) - Date.parse(stageStartedAt(home, now))), git: gitFacts(run.workspace_root), session_stats_command: `${flowCommand(context)} stat run sessions ls --run ${run.id} --project-root ${JSON.stringify(projectRoot)} --json`, usage_stats_command: `${flowCommand(context)} stat usage --run ${run.id} --project-root ${JSON.stringify(projectRoot)} --json`, next_command: reviewCommand }, 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" } };
     const reportJson = writeStageReport(root, report).json;
-    validateSchema({ schemaName: "stage-report", file: reportJson, projectRoot, ddFlowHome: context.ddFlowHome, runId: run.id });
+    validateSchema({ schemaName: "stage-report", file: reportJson, projectRoot, ddFlowHome: context.ddFlowHome, runId: run.id, runRoot: home });
     completeFlowRunStage(context, { projectRoot, runId: run.id, stage: "plan", status: "done", data: "stage-report.json", dataSchemaId: "dd-flow/stage-report@1", report: "stage-report.md", stageReport: "stage-report.html" });
     advanceFlowRun(context, { projectRoot, runId: run.id, status: "running", verdict: "planned", nextAction: "start_plan_review" });
     appendFlowRunTimelineEvent(context, project.id, run.id, { type: "plan_accepted", work_id: work.work_id, protocols, id: workSession.id, next_stage: "plan-review" });
@@ -149,7 +159,7 @@ export function validateVnextPlanArtifacts(context, input) {
     for (const [index, file] of planFiles.entries()) {
         const protocolId = input.protocols[index];
         try {
-            validateSchema({ schemaName: "vnext-protocol-plan", file, projectRoot: workspaceRoot, ddFlowHome: context.ddFlowHome, runId: input.runId });
+            validateSchema({ schemaName: "vnext-protocol-plan", file, projectRoot: workspaceRoot, ddFlowHome: context.ddFlowHome, runId: input.runId, runRoot: input.home });
             const value = readPlan(file);
             assertPlanIdentity(value, planIdentity(input.home, input.runId, protocolId, ownership.get(protocolId) ?? []), file);
             validatePlanSemantics(file, new Set(ownership.get(protocolId) ?? []), obligations);
@@ -170,20 +180,29 @@ export function validateVnextPlanArtifacts(context, input) {
     for (const file of mapFiles) {
         try {
             normalizeAspectMapRefs(file, workspaceRoot, input.home, input.runId);
-            validateSchema({ schemaName: "plan-aspect-map", file, projectRoot: workspaceRoot, ddFlowHome: context.ddFlowHome });
+            validateSchema({ schemaName: "plan-aspect-map", file, projectRoot: workspaceRoot, ddFlowHome: context.ddFlowHome, runId: input.runId, runRoot: input.home });
             validateAspectMap(file, input.protocols, workspaceRoot);
         }
         catch (error) {
             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-${crypto.randomUUID()}`;
         try {
             const projection = projectCodeWorkBatch({ home: input.home, workspaceRoot, runId: input.runId, plans, protocols: input.protocols, ...(frozenDocumentBaselines ? { frozenDocumentBaselines } : {}) });
             validateProjectedPaths(projection, workspaceRoot, input.home, input.runId);
             fs.writeFileSync(temporaryBatch, `${JSON.stringify(projection, null, 2)}\n`);
-            validateSchema({ schemaName: "code-work-batch", file: temporaryBatch, projectRoot: workspaceRoot, ddFlowHome: context.ddFlowHome, runId: input.runId });
+            validateSchema({ schemaName: "code-work-batch", file: temporaryBatch, projectRoot: workspaceRoot, ddFlowHome: context.ddFlowHome, runId: input.runId, runRoot: input.home });
             validateWorkBatchFile(temporaryBatch);
             if (input.publishBatch !== false) {
                 fs.renameSync(temporaryBatch, batch);
@@ -212,10 +231,10 @@ export function validateVnextCodeHandoff(context, input) {
 const planTask = "Produce accepted plan.json and aspect-map.json artifacts.";
 function planExample(protocolId) { return { schema_id: "dd-flow/protocol-plan@6", plan_id: "PLAN-001", protocol_id: protocolId, revision: 1, title: "Example", summary: "A compact executable plan.", source_refs: [{ kind: "specify", id: "SPECIFY", path: "run://RUN-000/01-specify/specify.json", requirement_ids: ["R-001", "AC-001"] }], goal: { outcome: "Deliver the accepted behavior.", constraints: ["Keep the accepted scope."], non_goals: [] }, assessment: { scope_breadth: { level: "narrow", surfaces: ["one surface"], reason: "One vertical slice." }, solution_novelty: { level: "established", surfaces: ["existing pattern"], reason: "Reuse project practice." }, solution_uncertainty: { level: "low", surfaces: ["known behavior"], reason: "No open technical question." }, failure_impact: { level: "low", surfaces: ["local feature"], reason: "Reversible local change." }, selected_depth: "compact_plan", depth_trigger: "none" }, decisions: [], document_updates: [], checks: [{ id: "CHK-P1-TEST", command: "pnpm test", purpose: "Proves the changed behavior.", run_at: "work", availability: "available" }], items: [{ id: "P1", title: "Implement behavior", summary: "Change the owning surface.", details: "Follow the accepted requirement and project conventions.", depends_on: [], requirement_refs: ["R-001", "AC-001"], semantic_spine: { user_outcome: "The requested behavior is available.", component_responsibility: "Own the behavior.", must_preserve: ["Existing behavior."], non_goals: [], acceptance_contribution: "Makes AC-001 observable." }, execution_context: { required_read: ["apps/api/src/example.ts"], discovery_boundary: ["Related tests only."], planned_write_areas: ["apps/api/src/"], stop_conditions: ["Stop if accepted scope conflicts with current truth."] }, verification: { check_refs: ["CHK-P1-TEST"] } }], acceptance: [{ criterion_id: "AC-001", plan_item_ids: ["P1"], changed_surfaces: ["apps/api/src/example.ts"], path: "Exercise the accepted user path.", environment: "Local test environment.", fixtures: [], cleanup: "No persistent fixture.", check_refs: ["CHK-P1-TEST"], expected_evidence: ["Focused check passes."], proof_limits: ["Manual production evidence is not claimed."], gate: "work" }] }; }
 function aspectMapExample(protocolId) { return { $schema: "plan-aspect-map.schema.json", schema_id: "dd-flow/plan-aspect-map@3", protocol_id: protocolId, plan_id: "PLAN-001", plan_revision: 1, catalog_ref: { path: ".memory-bank/dd-flow/mb-sdlc/plan-aspects/aspects" }, routing: { initial_state: "orchestrator_local", selected_route: "local_compact", reason: "One genuinely small semantic unit.", groups: [] }, review_groups: [], aspects: [{ aspect_id: "example_aspect", applicability: "not_applicable", reason: "Only an example; use the supplied real catalog.", planned_artifact_refs: [] }] }; }
-function requireRun(context, root, id) { const project = requireProjectByRoot(context, root); const run = context.db.get("SELECT id, project_id, workspace_root, run_home_path FROM runs WHERE project_id = ? AND id = ?", [project.id, id]); if (!run)
+function requireRun(context, root, id) { const project = requireProjectByRoot(context, root); const run = context.db.get("SELECT id, project_id, workspace_root, run_root FROM runs WHERE project_id = ? AND id = ?", [project.id, id]); if (!run)
     throw new AppError("not_found", "RUN is not registered", 1); return run; }
-function requireHome(run) { if (!run.run_home_path)
-    throw new AppError("runtime_missing", "RUN workspace is unavailable", 1); return run.run_home_path; }
+function requireHome(run) { if (!run.run_root)
+    throw new AppError("runtime_missing", "RUN artifact root is unavailable", 1); return run.run_root; }
 function protocolIds(home) {
     const report = JSON.parse(fs.readFileSync(path.join(home, "02-protocolize", "stage-report.json"), "utf8"));
     const ids = report.semantic?.acceptance;
@@ -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 });
+    }
+}

package/dist/services/vnext-protocolize.js

@@ -135,7 +135,7 @@ export function finishVnextProtocolize(context, input) {
     const root = path.join(requireRunHome(run), "02-protocolize");
     const resultFile = path.resolve(input.resultFile);
     inside(root, resultFile);
-    validateSchema({ schemaName: "vnext-protocolize-result", file: resultFile, projectRoot, ddFlowHome: context.ddFlowHome, runId: run.id });
+    validateSchema({ schemaName: "vnext-protocolize-result", file: resultFile, projectRoot, ddFlowHome: context.ddFlowHome, runId: run.id, runRoot: requireRunHome(run) });
     const result = readResult(resultFile);
     if (result.outcome !== "protocolized")
         throw new AppError("validation", "PROTOCOLIZE may finish only as protocolized; use stage pause for every user question", 2);
@@ -162,7 +162,7 @@ export function finishVnextProtocolize(context, input) {
     const reportMarkdown = path.join(root, "stage-report.md");
     const reportHtml = path.join(root, "stage-report.html");
     writeStageReport(root, report);
-    validateSchema({ schemaName: "stage-report", file: reportJson, projectRoot, ddFlowHome: context.ddFlowHome, runId: run.id });
+    validateSchema({ schemaName: "stage-report", file: reportJson, projectRoot, ddFlowHome: context.ddFlowHome, runId: run.id, runRoot: requireRunHome(run) });
     completeFlowRunStage(context, { projectRoot, runId: run.id, stage, status: "done", data: "protocolize-result.json", dataSchemaId: "dd-flow/vnext-protocolize-result@3", report: "stage-report.md", stageReport: "stage-report.html" });
     appendFlowRunTimelineEvent(context, project.id, run.id, { type: "protocol_documents_materialized", work_id: work.work_id, protocol_ids: protocolIds, stage });
     advanceFlowRun(context, { projectRoot, runId: run.id, status: "running", verdict: "protocolized", nextAction: "start_plan" });
@@ -318,7 +318,7 @@ function readResult(file) {
 function acceptedSpecifyObligations(context, projectRoot, runId, specifyPath) {
     if (!fs.existsSync(specifyPath))
         throw new AppError("not_found", "PROTOCOLIZE requires accepted specify.json", 1, { path: specifyPath });
-    validateSchema({ schemaName: "vnext-specify", file: specifyPath, projectRoot, ddFlowHome: context.ddFlowHome, runId });
+    validateSchema({ schemaName: "vnext-specify", file: specifyPath, projectRoot, ddFlowHome: context.ddFlowHome, runId, runRoot: requireRunHome(requireRun(context, projectRoot, runId)) });
     const specify = readVnextSpecifyResult(specifyPath);
     return [
         ...specify.requirements.map((obligation) => ({ ...obligation, kind: "requirement" })),
@@ -498,10 +498,10 @@ function featureIndex(epicRoot, featureSlug) {
 }
 function requireWork(context, id) { const work = context.db.get("SELECT * FROM works WHERE work_id = ?", [id]); if (!work)
     throw new AppError("not_found", "Work is not registered", 1, { work_id: id }); return work; }
-function requireRun(context, projectRoot, id) { const project = requireProjectByRoot(context, projectRoot); const run = context.db.get("SELECT id, short_id, slug, project_id, workspace_root, run_home_path FROM runs WHERE project_id = ? AND id = ?", [project.id, id]); if (!run)
+function requireRun(context, projectRoot, id) { const project = requireProjectByRoot(context, projectRoot); const run = context.db.get("SELECT id, short_id, slug, project_id, workspace_root, run_root FROM runs WHERE project_id = ? AND id = ?", [project.id, id]); if (!run)
     throw new AppError("not_found", "RUN is not registered", 1, { run_id: id }); return run; }
-function requireRunHome(run) { if (!run.run_home_path)
-    throw new AppError("runtime_missing", "RUN has no portable workspace", 1); return run.run_home_path; }
+function requireRunHome(run) { if (!run.run_root)
+    throw new AppError("runtime_missing", "RUN artifact root is unavailable", 1); return run.run_root; }
 function inside(root, file) { const relative = path.relative(root, file); if (relative.startsWith("..") || path.isAbsolute(relative))
     throw new AppError("path_escape", "Result must be inside the protocolize workspace", 2); }
 function writeJson(file, value) { const tmp = `${file}.${crypto.randomUUID()}.tmp`; fs.writeFileSync(tmp, `${JSON.stringify(value, null, 2)}\n`); fs.renameSync(tmp, file); }

package/dist/services/vnext-specify.js

@@ -141,7 +141,7 @@ export function submitVnextSpecify(context, input) {
         throw new AppError("not_found", "--result-file must point to an existing file inside the SPECIFY workspace", 1, { result_file: input.resultFile });
     }
     try {
-        validateSchema({ schemaName: "vnext-specify", file: candidateFile, projectRoot, ddFlowHome: context.ddFlowHome, runId: run.id });
+        validateSchema({ schemaName: "vnext-specify", file: candidateFile, projectRoot, ddFlowHome: context.ddFlowHome, runId: run.id, runRoot: runHome });
         const result = readVnextSpecifyResult(candidateFile);
         validateObligations(result, candidateFile);
         const normalizedResult = `${JSON.stringify(result, null, 2)}\n`;
@@ -169,7 +169,7 @@ export function submitVnextSpecify(context, input) {
         const htmlPath = path.join(stageRoot, "stage-report.html");
         const report = buildStageReport({ run, work, workSession, outcome, result, resultMarkdown: renderedMarkdown, now, stageRoot, resultFile });
         writeStageReport(stageRoot, report);
-        validateSchema({ schemaName: "stage-report", file: reportPath, projectRoot, ddFlowHome: context.ddFlowHome, runId: run.id });
+        validateSchema({ schemaName: "stage-report", file: reportPath, projectRoot, ddFlowHome: context.ddFlowHome, runId: run.id, runRoot: runHome });
         completeFlowRunStage(context, {
             projectRoot,
             runId: run.id,
@@ -268,7 +268,7 @@ function readIntake(input) {
 }
 function requireRun(context, projectRoot, runId) {
     const project = requireProjectByRoot(context, projectRoot);
-    const run = context.db.get("SELECT id, project_id, project_root, workspace_root, run_home_path FROM runs WHERE project_id = ? AND id = ?", [project.id, runId]);
+    const run = context.db.get("SELECT id, project_id, project_root, workspace_root, run_root FROM runs WHERE project_id = ? AND id = ?", [project.id, runId]);
     if (!run)
         throw new AppError("not_found", "RUN is not registered", 1, { run_id: runId });
     return run;
@@ -280,9 +280,9 @@ function runStatus(context, projectId, runId) {
     return run.status;
 }
 function requiredRunHome(run) {
-    if (!run.run_home_path)
-        throw new AppError("runtime_missing", "RUN has no portable workspace", 1, { run_id: run.id });
-    return run.run_home_path;
+    if (!run.run_root)
+        throw new AppError("runtime_missing", "RUN has no portable artifact root", 1, { run_id: run.id });
+    return run.run_root;
 }
 function requireWork(context, workId) {
     const work = context.db.get("SELECT * FROM works WHERE work_id = ?", [workId]);