@deksden-com/dd-flow-cli 0.9.0-beta.8 → 0.9.0-beta.9

package/CHANGELOG.md

@@ -1,5 +1,12 @@
 # @deksden-com/dd-flow-cli
 
+## 0.9.0-beta.9
+
+### Patch Changes
+
+- f156ad8: Use harness-qualified native child capacity for PLAN-REVIEW, CODE and
+  CODE-REVIEW fan-out instead of creating flow-owned capacity probes.
+
 ## 0.9.0-beta.8
 
 ### Patch Changes

package/dist/build-info.json

@@ -1,11 +1,11 @@
 {
   "cli_package": "@deksden-com/dd-flow-cli",
-  "cli_version": "0.9.0-beta.8",
-  "cli_commit": "98af28f2e50816d1042ee2a75d2f8133b655ffcb",
-  "built_at": "2026-09-04T09:50:57.779Z",
+  "cli_version": "0.9.0-beta.9",
+  "cli_commit": "7c6e37d05c8039349dee26fb7446abd6dcaef136",
+  "built_at": "2026-09-04T15:11:45.164Z",
   "built_with_canon": {
     "version": "4.0.1",
-    "commit": "7bd6761e6bd4217ceec08e2f9e98387b60545661",
+    "commit": "71c62fccbfa32c61ae77874ff42c8f9fb19e9f81",
     "flow_contract": "dd-flow-canonical-2026-08",
     "repo_root": "/Users/deksden/Documents/_Projects/dd-memorybank",
     "memorybank_root": "/Users/deksden/Documents/_Projects/dd-memorybank/.memory-bank",

package/dist/services/vnext-fanout.js

@@ -7,12 +7,6 @@ import { getFlowRunVariables } from "./runs.js";
 import { listWorks } from "./work-registry.js";
 import { vnextStageDirectory } from "../domain/stage-catalog.js";
 export const subagentCapacityKey = "runtime.subagent.available_slots";
-export const capacityProbe = {
-    fanout_size: 15,
-    probe_hold_seconds: 60,
-    cleanup_deadline_seconds: 180,
-    completion_token: "AGENT-NN"
-};
 export function fanoutState(input) {
     if (!input.hasWork)
         return input.dispatch === "none" ? "coordinator_required" : "dispatch_required";
@@ -70,8 +64,7 @@ export function getVnextFanoutStatus(context, input) {
             state: fanoutState({ dispatch: descriptor.dispatch, hasWork, capacityRequired: descriptor.capacity_required, capacityKnown, created: counts.created ?? 0, running: counts.running ?? 0 }),
             capacity: {
                 run_key: subagentCapacityKey,
-                available_slots: capacityKnown ? available : null,
-                ...(descriptor.capacity_required && !capacityKnown ? { probe: capacityProbe } : {})
+                available_slots: capacityKnown ? available : null
             },
             works: { ...counts, ready: ready.works.map((work) => ({ work_id: work.work_id, task: work.task, start_command: work.start_command })) }
         }

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

@@ -16,13 +16,10 @@ 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]));
@@ -126,7 +123,7 @@ 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));
@@ -162,14 +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);
@@ -184,7 +181,7 @@ export function finishVnextPlanReview(context, input) {
         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, 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 = ? AND task <> 'Capacity probe: return ready and finish this Work.'", [project.id, run.id, parent.work_id]);
+    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;
@@ -267,7 +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 | 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");
@@ -327,15 +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"));
@@ -454,14 +451,14 @@ 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 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);

package/dist/services/vnext-plan.js

@@ -81,9 +81,9 @@ export function startVnextPlan(context, input) {
     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 policyMergeAliases = codeCheckProfile?.mandatory_by_gate.merge ?? [];
     const mergeContract = mergeRequired

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@deksden-com/dd-flow-cli",
-  "version": "0.9.0-beta.8",
+  "version": "0.9.0-beta.9",
   "description": "Mechanical runtime CLI for dd-flow workflows.",
   "type": "module",
   "bin": {