cool-workflow 0.1.81 → 0.1.82

package/dist/worker-isolation.js

@@ -34,6 +34,7 @@ const telemetry_ledger_1 = require("./telemetry-ledger");
 const coordinator_1 = require("./coordinator");
 const helpers_1 = require("./worker-isolation/helpers");
 const paths_1 = require("./worker-isolation/paths");
+const validation_1 = require("./validation");
 exports.WORKER_ISOLATION_SCHEMA_VERSION = 1;
 function allocateWorkerScope(run, task, options = {}) {
     ensureWorkerState(run);
@@ -269,11 +270,31 @@ function getWorkerScope(run, workerId) {
     const file = node_path_1.default.join(workerRoot(run), (0, state_1.safeFileName)(workerId), paths_1.WORKER_SCOPE_FILE);
     if (!node_fs_1.default.existsSync(file))
         return undefined;
-    const scope = JSON.parse(node_fs_1.default.readFileSync(file, "utf8"));
+    const scope = (0, validation_1.validateWorkerScope)(JSON.parse(node_fs_1.default.readFileSync(file, "utf8")));
     upsertWorkerScope(run, scope);
     return scope;
 }
 function recordWorkerOutput(run, workerId, resultPath, options = {}) {
+    // Accept-path orchestrator. The recorded order + side effects of these ordered
+    // steps are load-bearing (replay determinism + the hash-chained audit/telemetry
+    // ledgers cross-link by parent event ids), so each helper runs exactly where it
+    // did before and mutates the shared `accept` context in place. Do NOT reorder.
+    const accept = validateWorkerResult(run, workerId, resultPath, options);
+    const delegation = attestWorkerDelegation(accept);
+    acceptWorkerResult(accept, delegation);
+    recordWorkerDelegationLedger(accept, delegation);
+    runWorkerVerify(accept);
+    recordWorkerCompletion(accept, delegation);
+    fanOutWorkerOutput(accept);
+    if (options.persist !== false)
+        (0, state_1.saveCheckpoint)(run);
+    return accept.output;
+}
+/** Step 1 — validateResult: resolve scope/task, enforce the sandbox boundary, the
+ *  result-file existence, the envelope contract, and (opt-in) resolvable evidence.
+ *  Fail-closed: any guard records a worker failure and throws BEFORE accept-side
+ *  state mutation. Returns the partially-filled accept context on success. */
+function validateWorkerResult(run, workerId, resultPath, options) {
     const scope = requireWorkerScope(run, workerId);
     const task = requireWorkerTask(run, scope);
     const absoluteResultPath = node_path_1.default.resolve(resultPath);
@@ -316,6 +337,30 @@ function recordWorkerOutput(run, workerId, resultPath, options = {}) {
             throw new Error(error.message);
         }
     }
+    return {
+        run,
+        workerId,
+        options,
+        scope,
+        task,
+        absoluteResultPath,
+        rawResult,
+        parsedResult,
+        destination: "",
+        pathAuditId: "",
+        acceptedAuditId: "",
+        resultNode: undefined,
+        verifierNodeId: "",
+        verifierStatus: "",
+        output: undefined
+    };
+}
+/** Step 2 — attestSandbox/attestDelegation: verify the agent's signed telemetry
+ *  BEFORE recording it, enforce the opt-in require-attested-telemetry gate (still
+ *  fail-closed, pre-mutation), and build the agent-hop provenance. Non-agent hops
+ *  return an empty delegation. */
+function attestWorkerDelegation(accept) {
+    const { run, workerId, options, task, absoluteResultPath, rawResult } = accept;
     // Agent Delegation Drive (v0.1.38): if this worker's result.md was produced by an
     // EXTERNAL agent, record the agent-hop attestation AS PROVENANCE — the agent
     // (kind:process) handle, the agent-REPORTED model (never CW_AGENT_MODEL), the
@@ -355,6 +400,16 @@ function recordWorkerOutput(run, workerId, resultPath, options = {}) {
             ...(telemetry ? { usageAttestation: telemetry.status, usageAttestationReason: telemetry.reason } : {})
         }
         : undefined;
+    return { agentDelegation, telemetry };
+}
+/** Step 3 — recordStateNode/audit: the irreversible accept. Records the allowed
+ *  path decision, copies the result into the run results dir, completes the task,
+ *  builds + appends the result node, emits the accepted audit event, re-normalizes
+ *  node evidence against both audit ids, and surfaces the empty-capture warning.
+ *  Writes destination/pathAuditId/acceptedAuditId/resultNode back into `accept`. */
+function acceptWorkerResult(accept, delegation) {
+    const { run, workerId, scope, task, absoluteResultPath, parsedResult } = accept;
+    const { agentDelegation } = delegation;
     const pathAudit = (0, trust_audit_1.recordSandboxPathDecision)(run, {
         workerId,
         taskId: task.id,
@@ -443,58 +498,75 @@ function recordWorkerOutput(run, workerId, resultPath, options = {}) {
             metadata: { reason: "no findings or evidence captured from result.md", resultPath: destination }
         });
     }
+    accept.destination = destination;
+    accept.pathAuditId = pathAudit.id;
+    accept.acceptedAuditId = acceptedAudit.id;
+    accept.resultNode = resultNode;
+}
+/** Step 4 — recordTelemetryLedger: the agent-hop attestation. Binds the telemetry
+ *  verdict into the append-only hash-chained ledger BEFORE the audit event (so the
+ *  event can cross-link the record hash), then emits the worker.agent-delegation
+ *  audit event. No-op for non-agent hops. */
+function recordWorkerDelegationLedger(accept, delegation) {
+    const { agentDelegation } = delegation;
     // The agent-hop attestation event — hung off worker.output, alongside
     // worker.backend. Recorded in trust-audit/provenance, NEVER in node evidence.
-    if (agentDelegation) {
-        // Track 1 (tamper-evidence): bind this verdict into the append-only,
-        // hash-chained telemetry ledger BEFORE the audit event, so the event can
-        // cross-link the record hash. Editing the recorded verdict/usage later breaks
-        // the chain (verifyTelemetryLedger). Only when a verdict was computed.
-        const ledgerRecord = agentDelegation.usageAttestation
-            ? (0, telemetry_ledger_1.appendTelemetryAttestation)(run, {
-                workerId,
-                taskId: task.id,
-                promptDigest: agentDelegation.promptDigest,
-                reportedUsage: agentDelegation.reportedUsage,
-                usageSignature: agentDelegation.usageSignature,
-                attestation: agentDelegation.usageAttestation,
-                attestationReason: agentDelegation.usageAttestationReason
-            })
-            : undefined;
-        (0, trust_audit_1.recordTrustAuditEvent)(run, {
-            kind: "worker.agent-delegation",
-            decision: "recorded",
-            source: "host-attested",
+    if (!agentDelegation)
+        return;
+    const { run, workerId, scope, task, resultNode, acceptedAuditId } = accept;
+    // Track 1 (tamper-evidence): bind this verdict into the append-only,
+    // hash-chained telemetry ledger BEFORE the audit event, so the event can
+    // cross-link the record hash. Editing the recorded verdict/usage later breaks
+    // the chain (verifyTelemetryLedger). Only when a verdict was computed.
+    const ledgerRecord = agentDelegation.usageAttestation
+        ? (0, telemetry_ledger_1.appendTelemetryAttestation)(run, {
             workerId,
             taskId: task.id,
-            nodeId: resultNode.id,
-            sandboxProfileId: scope.sandboxProfileId,
-            policySnapshot: scope.sandboxPolicy,
-            parentEventIds: [acceptedAudit.id],
-            metadata: {
-                backendId: agentDelegation.backendId,
-                handleKind: agentDelegation.handle.kind,
-                handleRef: agentDelegation.handle.ref,
-                model: agentDelegation.model,
-                promptDigest: agentDelegation.promptDigest,
-                resultDigest: agentDelegation.resultDigest,
-                command: agentDelegation.command,
-                args: agentDelegation.args,
-                exitCode: agentDelegation.exitCode,
-                // Track 1: the telemetry verdict travels with the agent-hop event so the
-                // audit report can surface `unattested` usage loudly. Absent ⇒ no usage.
-                ...(agentDelegation.usageAttestation
-                    ? {
-                        telemetryAttestation: agentDelegation.usageAttestation,
-                        ...(agentDelegation.usageAttestationReason ? { telemetryAttestationReason: agentDelegation.usageAttestationReason } : {}),
-                        ...(agentDelegation.reportedUsage ? { reportedUsage: agentDelegation.reportedUsage } : {}),
-                        // Cross-link to the hash-chained ledger entry (tamper-evidence).
-                        ...(ledgerRecord ? { telemetryRecordId: ledgerRecord.recordId, telemetryRecordHash: ledgerRecord.recordHash, telemetryPrevHash: ledgerRecord.prevHash } : {})
-                    }
-                    : {})
-            }
-        });
-    }
+            promptDigest: agentDelegation.promptDigest,
+            reportedUsage: agentDelegation.reportedUsage,
+            usageSignature: agentDelegation.usageSignature,
+            attestation: agentDelegation.usageAttestation,
+            attestationReason: agentDelegation.usageAttestationReason
+        })
+        : undefined;
+    (0, trust_audit_1.recordTrustAuditEvent)(run, {
+        kind: "worker.agent-delegation",
+        decision: "recorded",
+        source: "host-attested",
+        workerId,
+        taskId: task.id,
+        nodeId: resultNode.id,
+        sandboxProfileId: scope.sandboxProfileId,
+        policySnapshot: scope.sandboxPolicy,
+        parentEventIds: [acceptedAuditId],
+        metadata: {
+            backendId: agentDelegation.backendId,
+            handleKind: agentDelegation.handle.kind,
+            handleRef: agentDelegation.handle.ref,
+            model: agentDelegation.model,
+            promptDigest: agentDelegation.promptDigest,
+            resultDigest: agentDelegation.resultDigest,
+            command: agentDelegation.command,
+            args: agentDelegation.args,
+            exitCode: agentDelegation.exitCode,
+            // Track 1: the telemetry verdict travels with the agent-hop event so the
+            // audit report can surface `unattested` usage loudly. Absent ⇒ no usage.
+            ...(agentDelegation.usageAttestation
+                ? {
+                    telemetryAttestation: agentDelegation.usageAttestation,
+                    ...(agentDelegation.usageAttestationReason ? { telemetryAttestationReason: agentDelegation.usageAttestationReason } : {}),
+                    ...(agentDelegation.reportedUsage ? { reportedUsage: agentDelegation.reportedUsage } : {}),
+                    // Cross-link to the hash-chained ledger entry (tamper-evidence).
+                    ...(ledgerRecord ? { telemetryRecordId: ledgerRecord.recordId, telemetryRecordHash: ledgerRecord.recordHash, telemetryPrevHash: ledgerRecord.prevHash } : {})
+                }
+                : {})
+        }
+    });
+}
+/** Step 5 — runVerify: drive the verify pipeline stage off the accepted result node
+ *  and record the verifier node id on the task + accept context. */
+function runWorkerVerify(accept) {
+    const { run, workerId, scope, task, parsedResult, destination, resultNode } = accept;
     const verifierResult = (0, pipeline_runner_1.createPipelineRunner)({ persist: false }).runPipelineStage(run, "verify", resultNode.id, {
         outputNodeId: `${run.id}:verifier:${task.id}`,
         outputStatus: "verified",
@@ -507,14 +579,24 @@ function recordWorkerOutput(run, workerId, resultPath, options = {}) {
         metadata: { taskId: task.id, workerId, resultNodeId: resultNode.id, sandboxProfileId: scope.sandboxProfileId }
     });
     task.verifierNodeId = verifierResult.outputNodeId;
+    accept.verifierNodeId = verifierResult.outputNodeId;
+    // Carry the verify verdict for the scope-status transition in recordWorkerCompletion.
+    accept.verifierStatus = verifierResult.status;
+}
+/** Step 6 — recordStateNode (worker record): assemble the worker output record +
+ *  host-attested usage record, then persist the worker scope with the verify-derived
+ *  status, output digest/size, and (when present) usage. */
+function recordWorkerCompletion(accept, delegation) {
+    const { run, workerId, scope, task, absoluteResultPath, rawResult, resultNode, verifierNodeId, verifierStatus, pathAuditId, acceptedAuditId } = accept;
+    const { agentDelegation, telemetry } = delegation;
     const output = {
         workerId,
         taskId: task.id,
         resultPath: absoluteResultPath,
         recordedAt: new Date().toISOString(),
         stateNodeId: resultNode.id,
-        verifierNodeId: verifierResult.outputNodeId,
-        auditEventIds: [pathAudit.id, acceptedAudit.id]
+        verifierNodeId,
+        auditEventIds: [pathAuditId, acceptedAuditId]
     };
     // Host-attested usage rides on the worker record. Recorded when the agent
     // REPORTED a model OR token usage — `unreported`/absent stays ABSENT (never
@@ -537,7 +619,7 @@ function recordWorkerOutput(run, workerId, resultPath, options = {}) {
     updateWorkerScope(run, {
         ...scope,
         updatedAt: new Date().toISOString(),
-        status: verifierResult.status === "advanced" ? "verified" : "completed",
+        status: verifierStatus === "advanced" ? "verified" : "completed",
         resultNodeId: resultNode.id,
         output,
         // Output integrity (v0.1.63): SHA256 digest + file size
@@ -545,20 +627,23 @@ function recordWorkerOutput(run, workerId, resultPath, options = {}) {
         outputSizeBytes: Buffer.byteLength(rawResult, "utf8"),
         ...(usageRecord ? { usage: usageRecord } : {})
     });
-    const blackboardLinks = publishWorkerOutputToBlackboard(run, scope, task, parsedResult.summary, destination, absoluteResultPath, resultNode.evidence, acceptedAudit.id);
+    accept.output = output;
+}
+/** Step 7 — fanOut: publish the accepted output to the blackboard and record the
+ *  multi-agent worker output (linking the blackboard message/artifact refs). */
+function fanOutWorkerOutput(accept) {
+    const { run, workerId, scope, task, absoluteResultPath, parsedResult, destination, resultNode, verifierNodeId, acceptedAuditId } = accept;
+    const blackboardLinks = publishWorkerOutputToBlackboard(run, scope, task, parsedResult.summary, destination, absoluteResultPath, resultNode.evidence, acceptedAuditId);
     (0, multi_agent_1.recordMultiAgentWorkerOutput)(run, {
         workerId,
         taskId: task.id,
         resultNodeId: resultNode.id,
-        verifierNodeId: verifierResult.outputNodeId,
+        verifierNodeId,
         evidence: resultNode.evidence,
         artifactPaths: [destination, absoluteResultPath],
         blackboardMessageIds: blackboardLinks.messageIds,
         blackboardArtifactRefIds: blackboardLinks.artifactRefIds
     });
-    if (options.persist !== false)
-        (0, state_1.saveCheckpoint)(run);
-    return output;
 }
 function recordWorkerFailure(run, workerId, error, options = {}) {
     const scope = requireWorkerScope(run, workerId);
@@ -786,7 +871,7 @@ function loadWorkerScopesFromDisk(run) {
         .filter((entry) => entry.isDirectory())
         .map((entry) => node_path_1.default.join(workerRoot(run), entry.name, paths_1.WORKER_SCOPE_FILE))
         .filter((file) => node_fs_1.default.existsSync(file))
-        .map((file) => JSON.parse(node_fs_1.default.readFileSync(file, "utf8")));
+        .map((file) => (0, validation_1.validateWorkerScope)(JSON.parse(node_fs_1.default.readFileSync(file, "utf8"))));
 }
 function requireWorkerScope(run, workerId) {
     const scope = getWorkerScope(run, workerId);

package/docs/agent-delegation-drive.7.md

@@ -261,3 +261,4 @@ Adds the opt-in fast architecture-review lane: scoped JSONL source contexts, dif
 ## Resumable Drive & Resume Routing (v0.1.81)
 
 Adds `run resume <id> --drive/--once` alongside `quickstart --resume`: a stopped pipeline resumes in-place, advancing to completion (`--drive`) or one deterministic step (`--once`) over the same plan->dispatch->agent-fulfill->accept->commit lifecycle, echoing `resumedFrom: <id>`. Fixes the `run resume --drive` CLI routing so the drive flag reaches the resumed run instead of being read as an app name. Replay determinism and the agent evidence triple are unchanged.
+_No behavioral change in v0.1.82 (drive/quickstart resolve the run repo via an explicit base directory rather than process.chdir; delegation behavior is unchanged)._

package/docs/cli-mcp-parity.7.md

@@ -389,3 +389,4 @@ Adds the opt-in fast architecture-review lane: scoped JSONL source contexts, dif
 ## Re-Prove Verbs on Both Surfaces (v0.1.81)
 
 v0.1.81 grows the parity surface with two new both-surface, fail-closed verbs declared once in the capability registry: `cw audit verify` / `cw_audit_verify` re-proves the trust-audit chain and exits non-zero on any unverified or corrupt chain, and `cw run inspect-archive` / `cw_run_inspect_archive` is a read-only archive integrity check. Each `cw <cmd> --json` is schema-identical to its `cw_<tool>` and validated by the same parity gate.
+_No changes in v0.1.82._

package/docs/contract-migration-tooling.7.md

@@ -129,3 +129,4 @@ Migration DAG with reversible edges (v0.1.45), capability auto-discovery (v0.1.4
 Adds the opt-in fast architecture-review lane: scoped JSONL source contexts, diff-aware exports, reusable Map and Assess results, measurable wrapper metrics, actionable background full-review handoff, and userland model policy flags for routing fast/strong workers without changing the full review contract.
 
 _No changes to the contract-migration subsystem in v0.1.81._
+_No changes in v0.1.82._

package/docs/control-plane-scheduling.7.md

@@ -116,3 +116,4 @@ Migration DAG with reversible edges (v0.1.45), capability auto-discovery (v0.1.4
 Adds the opt-in fast architecture-review lane: scoped JSONL source contexts, diff-aware exports, reusable Map and Assess results, measurable wrapper metrics, actionable background full-review handoff, and userland model policy flags for routing fast/strong workers without changing the full review contract.
 
 _No changes to the control-plane scheduling surface in v0.1.81._
+_No behavioral change in v0.1.82 (schedule/run ids are now deterministic-but-unique via a monotonic counter + pid instead of Math.random; ids stay collision-free)._

package/docs/durable-state-and-locking.7.md

@@ -115,3 +115,4 @@ Adds the opt-in fast architecture-review lane: scoped JSONL source contexts, dif
 ## Deterministic Tombstone Hash (v0.1.81)
 
 The reclamation tombstone's freed-manifest is now path-sorted before it feeds `tombstoneHash`, so the same freed set always yields the same hash regardless of filesystem enumeration order. This removes a non-determinism from the write-ahead chain (v0.1.39/v0.1.40), keeping the per-run tombstone hash-chain replayable and stable across hosts. Atomicity, locking, and the durable re-point seam are unchanged. v0.1.81 also adds import-time refusal (`CW_REQUIRE_ARCHIVE_INTEGRITY=1`) and restore-time trust-audit re-proving — see run-registry-control-plane(7).
+_No changes in v0.1.82._

package/docs/evidence-adoption-reasoning-chain.7.md

@@ -276,3 +276,4 @@ Migration DAG with reversible edges (v0.1.45), capability auto-discovery (v0.1.4
 Adds the opt-in fast architecture-review lane: scoped JSONL source contexts, diff-aware exports, reusable Map and Assess results, measurable wrapper metrics, actionable background full-review handoff, and userland model policy flags for routing fast/strong workers without changing the full review contract.
 
 _No changes to the Evidence Adoption Reasoning Chain surface in v0.1.81. The v0.1.81 trust-audit `computeEventHash` fix hardens the underlying audit records this chain links to by reference, but the derived reasoning view, its commands, and its eval gates are unchanged._
+_No changes in v0.1.82._

package/docs/execution-backends.7.md

@@ -306,3 +306,4 @@ Migration DAG with reversible edges (v0.1.45), capability auto-discovery (v0.1.4
 Adds the opt-in fast architecture-review lane: scoped JSONL source contexts, diff-aware exports, reusable Map and Assess results, measurable wrapper metrics, actionable background full-review handoff, and userland model policy flags for routing fast/strong workers without changing the full review contract.
 
 _No changes to the execution-backends surface in v0.1.81._
+_No behavioral change in v0.1.82 (the delegated child programs moved from inline `node -e` strings to `scripts/children/`; spawn argv and shell:false behavior are byte-identical)._

package/docs/multi-agent-cli-mcp-surface.7.md

@@ -273,3 +273,4 @@ Adds the opt-in fast architecture-review lane: scoped JSONL source contexts, dif
 ## New Both-Surface Verbs (v0.1.81)
 
 v0.1.81 adds `audit verify` (`cw_audit_verify`) and `run inspect-archive` (`cw_run_inspect_archive`) — both declared once in the capability registry and exposed identically on the CLI and MCP, fail-closed (non-zero exit / `ok:false` on an unverified chain or a tampered archive).
+_No changes in v0.1.82._

package/docs/multi-agent-eval-replay-harness.7.md

@@ -308,3 +308,4 @@ Migration DAG with reversible edges (v0.1.45), capability auto-discovery (v0.1.4
 Adds the opt-in fast architecture-review lane: scoped JSONL source contexts, diff-aware exports, reusable Map and Assess results, measurable wrapper metrics, actionable background full-review handoff, and userland model policy flags for routing fast/strong workers without changing the full review contract.
 
 _No changes to the multi-agent eval/replay harness in v0.1.81 (the multi-agent-eval module was carved into behavior-preserving siblings; replay output is byte-identical)._
+_v0.1.82 — replay now RE-DERIVES the projection from the raw captured state instead of copying the baseline, so a nondeterministic projection is caught instead of silently passing; a new regression smoke (including an intrinsic-nondeterminism case) proves the moat has teeth._

package/docs/multi-agent-operator-ux.7.md

@@ -320,3 +320,4 @@ Migration DAG with reversible edges (v0.1.45), capability auto-discovery (v0.1.4
 Adds the opt-in fast architecture-review lane: scoped JSONL source contexts, diff-aware exports, reusable Map and Assess results, measurable wrapper metrics, actionable background full-review handoff, and userland model policy flags for routing fast/strong workers without changing the full review contract.
 
 _No changes to the multi-agent operator UX surface in v0.1.81 (the operator-ux module was carved into behavior-preserving siblings; output is byte-identical)._
+_No changes in v0.1.82._

package/docs/node-snapshot-diff-replay.7.md

@@ -141,3 +141,4 @@ Migration DAG with reversible edges (v0.1.45), capability auto-discovery (v0.1.4
 Adds the opt-in fast architecture-review lane: scoped JSONL source contexts, diff-aware exports, reusable Map and Assess results, measurable wrapper metrics, actionable background full-review handoff, and userland model policy flags for routing fast/strong workers without changing the full review contract.
 
 _No changes to node-snapshot diff/replay in v0.1.81._
+_No behavioral change in v0.1.82 (the node projection field set was unified into one source of truth; snapshot/replay digests are byte-identical)._

package/docs/observability-cost-accounting.7.md

@@ -200,3 +200,4 @@ Migration DAG with reversible edges (v0.1.45), capability auto-discovery (v0.1.4
 Adds the opt-in fast architecture-review lane: scoped JSONL source contexts, diff-aware exports, reusable Map and Assess results, measurable wrapper metrics, actionable background full-review handoff, and userland model policy flags for routing fast/strong workers without changing the full review contract.
 
 _No changes to the observability + cost-accounting surface in v0.1.81 (the observability module was carved into behavior-preserving siblings; output is byte-identical)._
+_No changes in v0.1.82._

package/docs/project-index.md

@@ -1,15 +1,15 @@
 # Cool Workflow Project Index
 
-Generated from the current repository code on 2026-06-14 by `npm run sync:project-index`.
+Generated from the current repository code on 2026-06-15 by `npm run sync:project-index`.
 
 ## Snapshot
 
 - Package: `cool-workflow`
-- Version: `0.1.81`
-- Source modules: `58`
+- Version: `0.1.82`
+- Source modules: `61`
 - Workflow apps: `7`
 - Docs: `49`
-- Smoke tests: `86`
+- Smoke tests: `87`
 - Repository: https://github.com/coo1white/cool-workflow
 
 ## Architecture
@@ -90,9 +90,11 @@ multi-agent host -> topology -> blackboard/coordinator
 - [evidence-grounding.ts](../src/evidence-grounding.ts)
 - [evidence-reasoning.ts](../src/evidence-reasoning.ts)
 - [execution-backend.ts](../src/execution-backend.ts)
+- [gates.ts](../src/gates.ts)
 - [multi-agent-eval.ts](../src/multi-agent-eval.ts)
 - [multi-agent-operator-ux.ts](../src/multi-agent-operator-ux.ts)
 - [multi-agent-trust.ts](../src/multi-agent-trust.ts)
+- [node-projection.ts](../src/node-projection.ts)
 - [node-snapshot.ts](../src/node-snapshot.ts)
 - [observability.ts](../src/observability.ts)
 - [reclamation.ts](../src/reclamation.ts)
@@ -107,6 +109,7 @@ multi-agent host -> topology -> blackboard/coordinator
 - [telemetry-attestation.ts](../src/telemetry-attestation.ts)
 - [telemetry-demo.ts](../src/telemetry-demo.ts)
 - [telemetry-ledger.ts](../src/telemetry-ledger.ts)
+- [validation.ts](../src/validation.ts)
 - [verifier-registry.ts](../src/verifier-registry.ts)
 - [workbench-host.ts](../src/workbench-host.ts)
 - [workbench.ts](../src/workbench.ts)
@@ -210,6 +213,7 @@ Smoke tests mirror the public contracts. The high-signal suites are:
 - [h7-custom-profile-persist-smoke.js](../test/h7-custom-profile-persist-smoke.js)
 - [mcp-app-surface-smoke.js](../test/mcp-app-surface-smoke.js)
 - [multi-agent-cli-mcp-surface-smoke.js](../test/multi-agent-cli-mcp-surface-smoke.js)
+- [multi-agent-eval-determinism-regression-smoke.js](../test/multi-agent-eval-determinism-regression-smoke.js)
 - [multi-agent-eval-replay-harness-smoke.js](../test/multi-agent-eval-replay-harness-smoke.js)
 - [multi-agent-eval-replay-smoke.js](../test/multi-agent-eval-replay-smoke.js)
 - [multi-agent-operator-ux-smoke.js](../test/multi-agent-operator-ux-smoke.js)

package/docs/real-execution-backends.7.md

@@ -148,3 +148,4 @@ Migration DAG with reversible edges (v0.1.45), capability auto-discovery (v0.1.4
 Adds the opt-in fast architecture-review lane: scoped JSONL source contexts, diff-aware exports, reusable Map and Assess results, measurable wrapper metrics, actionable background full-review handoff, and userland model policy flags for routing fast/strong workers without changing the full review contract.
 
 _No changes to the real execution backends in v0.1.81._
+_No behavioral change in v0.1.82 (delegated child programs extracted to `scripts/children/`; spawn behavior byte-identical)._

package/docs/release-and-migration.7.md

@@ -288,3 +288,4 @@ Adds the opt-in fast architecture-review lane: scoped JSONL source contexts, dif
 ## Migration Compatibility (v0.1.81)
 
 v0.1.81 is additive: every change is a new flag/verb/env (`audit verify`, `run inspect-archive`, `verify-import --strict`, `CW_REQUIRE_ARCHIVE_INTEGRITY`, `quickstart --resume`, `run resume --drive`) or an internal behavior-preserving carve. Run-state schema, existing outputs, files, and exit codes are byte-identical, so runs and archives from prior versions load and verify unchanged. No migration action is required.
+_No changes in v0.1.82._

package/docs/release-tooling.7.md

@@ -133,8 +133,10 @@ inherits the agent's own credentials, and imports no model SDK — the red line.
 ```bash
 # check only (gate + independent review, no mutation):
 node plugins/cool-workflow/scripts/release-flow.js --check
-# cut a tag once review is green:
-node plugins/cool-workflow/scripts/release-flow.js --cut --version 0.1.77 [--push]
+# cut a tag once review is green (when --push, also creates the GitHub Release):
+node plugins/cool-workflow/scripts/release-flow.js --cut --version 0.1.77 [--push] [--no-release]
+# backfill / re-create the GitHub Release for an already-pushed tag (no gate/cut):
+node plugins/cool-workflow/scripts/release-flow.js --release --version 0.1.77 [--soft]
 ```
 
 The per-platform difference is config, not code — set the reviewer agent:
@@ -152,6 +154,34 @@ also get generated MCP manifests (`.gemini-plugin/`, `.opencode-plugin/`) so the
 `cw_*` tools are available as MCP tools in those hosts. The verdict path
 (`.cw-release/review-<sha>.verdict`) and the tag-push CI backstop are unchanged.
 
+### GitHub Release finishing step
+
+A `--cut --push` finishes by creating the **GitHub Release** for the tag, and
+`--release --version x.y.z` creates-or-skips one for an already-pushed tag
+(backfill). The notes body is assembled from the `## x.y.z` CHANGELOG section as
+shipped at the tag, the independent reviewer's one-line capability, and a
+"Provenance & audit" footer linking the reviewed commit, the **committed** reviewer
+verdict, the full diff, and the provenance-attested npm version.
+
+This step is **distribution upside, not a correctness gate**: the load-bearing
+artifacts (the tag and the provenance-attested npm publish) already exist when it
+runs. Therefore `gh` is **not** part of the node/git portability floor — it runs
+ONLY in the human `--cut --push` / `--release` paths (never a gate or CI path) and
+is **idempotent** (skips if the Release already exists).
+
+Failure semantics differ by mode: in the **`--cut --push` finishing step** an
+absent/unauthenticated/erroring `gh` **skips with a stderr note and never fails the
+cut** (the tag and npm publish stand) — opt out entirely with `--no-release`. In the
+explicit **`--release` backfill** it **fails closed** (exit 1) with guidance, since
+the operator asked for it directly; add `--soft` to downgrade `--release` to the same
+best-effort skip-not-fail behavior.
+
+Honesty: the notes claim the gated flow ("independent release-reviewer, verdict
+above") **only when a committed `APPROVED` verdict is found at the tag**; backfilling
+an ungated tag emits a neutral caveat instead and warns on stderr — the notes never
+assert a review that isn't there. Test seam: `CW_RELEASE_FLOW_GH_CMD` swaps the `gh`
+binary for a stub (spawned `shell:false`) so the smoke exercises it offline.
+
 0.1.51
 
 0.1.76
@@ -165,3 +195,4 @@ also get generated MCP manifests (`.gemini-plugin/`, `.opencode-plugin/`) so the
 Adds the opt-in fast architecture-review lane: scoped JSONL source contexts, diff-aware exports, reusable Map and Assess results, measurable wrapper metrics, actionable background full-review handoff, and userland model policy flags for routing fast/strong workers without changing the full review contract.
 
 _No changes to the release-flow tooling in v0.1.81; this release was cut through the existing gate->review->tag flow._
+_No changes to the release-tooling contract in v0.1.82._

package/docs/run-registry-control-plane.7.md

@@ -399,3 +399,4 @@ Adds the opt-in fast architecture-review lane: scoped JSONL source contexts, dif
 ## Resume Drive, Inspect-Archive & Restore Re-Prove (v0.1.81)
 
 v0.1.81 adds `run resume <id> --drive/--once` (continue an interrupted run via the agent-drive loop; default resume stays read-only and byte-identical), `run inspect-archive PATH` (read-only archive integrity check that names any offending file without importing), and restore-time hardening: `verify-import` now re-proves the trust-audit chain on restore and gains `--strict`, and `CW_REQUIRE_ARCHIVE_INTEGRITY=1` refuses a stripped-integrity archive before any write.
+_No behavioral change in v0.1.82 (run resolution now threads an explicit base directory via CoolWorkflowRunner.withBaseDir instead of mutating process.cwd; the resolved run is unchanged)._

package/docs/run-retention-reclamation.7.md

@@ -199,3 +199,4 @@ Adds the opt-in fast architecture-review lane: scoped JSONL source contexts, dif
 ## Deterministic Freed Manifest (v0.1.81)
 
 The freed manifest is path-sorted before it feeds `tombstoneHash`, so reclamation's write-ahead tombstone hash-chain is reproducible across hosts regardless of filesystem enumeration order. Reclaimed tiers, the re-point seam, and the default (reclaim-nothing) policy are unchanged.
+_No changes in v0.1.82._

package/docs/state-explosion-management.7.md

@@ -277,3 +277,4 @@ Migration DAG with reversible edges (v0.1.45), capability auto-discovery (v0.1.4
 Adds the opt-in fast architecture-review lane: scoped JSONL source contexts, diff-aware exports, reusable Map and Assess results, measurable wrapper metrics, actionable background full-review handoff, and userland model policy flags for routing fast/strong workers without changing the full review contract.
 
 _No changes to the state-explosion management surface in v0.1.81 (the module was carved into behavior-preserving siblings; output is byte-identical)._
+_No changes in v0.1.82._

package/docs/team-collaboration.7.md

@@ -213,3 +213,4 @@ Migration DAG with reversible edges (v0.1.45), capability auto-discovery (v0.1.4
 Adds the opt-in fast architecture-review lane: scoped JSONL source contexts, diff-aware exports, reusable Map and Assess results, measurable wrapper metrics, actionable background full-review handoff, and userland model policy flags for routing fast/strong workers without changing the full review contract.
 
 _No changes to the team-collaboration surface in v0.1.81._
+_No changes in v0.1.82._

package/docs/web-desktop-workbench.7.md

@@ -221,3 +221,4 @@ Migration DAG with reversible edges (v0.1.45), capability auto-discovery (v0.1.4
 Adds the opt-in fast architecture-review lane: scoped JSONL source contexts, diff-aware exports, reusable Map and Assess results, measurable wrapper metrics, actionable background full-review handoff, and userland model policy flags for routing fast/strong workers without changing the full review contract.
 
 _No changes to the Web / Desktop Workbench in v0.1.81._
+_No changes in v0.1.82._

package/manifest/plugin.manifest.json

@@ -2,7 +2,7 @@
   "_comment": "SINGLE SOURCE OF TRUTH for every vendor manifest. Edit THIS file, then run `npm run gen:manifests`. Do NOT hand-edit the generated vendor manifests (.claude-plugin/, .codex-plugin/, .agents/, .mcp.json) — `npm run gen:manifests -- --check` (run by release:check) will fail if they drift from this source.",
   "identity": {
     "name": "cool-workflow",
-    "version": "0.1.81",
+    "version": "0.1.82",
     "license": "BSD-2-Clause",
     "homepage": "https://github.com/coo1white/cool-workflow",
     "author": {

package/manifest/source-context-profiles.json

@@ -25,7 +25,7 @@
     },
     "runtime": {
       "description": "Runtime-kernel source context for state, orchestration, scheduling, execution, and shared types.",
-      "maxLines": 40000,
+      "maxLines": 42000,
       "include": [
         "plugins/cool-workflow/src/**",
         "plugins/cool-workflow/package.json",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "cool-workflow",
-  "version": "0.1.81",
+  "version": "0.1.82",
   "bin": {
     "cool-workflow": "scripts/cw.js",
     "cw": "scripts/cw.js"

package/scripts/canonical-apps.js

@@ -83,7 +83,7 @@ const canonicalApps = [
       "--source",
       "plugins/cool-workflow/docs/workflow-app-framework.7.md",
       "--scope",
-      "Cool Workflow v0.1.81",
+      "Cool Workflow v0.1.82",
       "--freshness",
       "as of release preparation"
     ]
@@ -117,14 +117,14 @@ function main() {
     assert.ok(summary, `${app.id} must appear in app list`);
     assert.equal(summary.sourceKind, "app-directory");
     assert.equal(summary.legacy, false);
-    assert.equal(summary.version, "0.1.81");
+    assert.equal(summary.version, "0.1.82");
 
     const validation = runJson(["app", "validate", manifestPath]);
     assert.equal(validation.valid, true, `${app.id} manifest must validate`);
 
     const shown = runJson(["app", "show", app.id]);
     assert.equal(shown.app.id, app.id);
-    assert.equal(shown.app.version, "0.1.81");
+    assert.equal(shown.app.version, "0.1.82");
     assert.ok(shown.app.metadata.canonical, `${app.id} must be marked canonical`);
     assert.ok(shown.app.sandboxProfiles.length > 0, `${app.id} must declare sandbox profiles`);
     assertTaskIdsUnique(shown);
@@ -135,7 +135,7 @@ function main() {
     const plan = runJson(["plan", app.id, ...app.args(workspace)]);
     const state = JSON.parse(fs.readFileSync(plan.statePath, "utf8"));
     assert.equal(state.workflow.app.id, app.id);
-    assert.equal(state.workflow.app.version, "0.1.81");
+    assert.equal(state.workflow.app.version, "0.1.82");
     assert.equal(state.workflow.app.metadata.canonical, true);
     assert.ok(state.tasks.some((task) => task.requiresEvidence), `${app.id} plan must include evidence gates`);
     assert.ok(state.tasks.every((task) => task.sandboxProfileId), `${app.id} plan must include sandbox hints`);