cclaw-cli 0.51.22 → 0.51.24

package/dist/content/subagents.js

@@ -7,6 +7,8 @@ import { conversationLanguagePolicyMarkdown } from "./language-policy.js";
  */
 const SUBAGENT_AGENT_NAMES = [
     "planner",
+    "product-manager",
+    "critic",
     "reviewer",
     "security-reviewer",
     "test-author",
@@ -83,16 +85,26 @@ can enforce phase-appropriate write boundaries. Use separate workers only when t
 |---|---|---|---|---|
 | Claude | \`native\` | Task (named subagent_type) | AskUserQuestion | \`npx cclaw-cli doctor\` |
 | Cursor | \`generic-dispatch\` | Task (generic subagent_type: explore/generalPurpose/…) | AskQuestion | \`npx cclaw-cli doctor\` |
-| OpenCode | \`role-switch\` | plugin dispatch _or_ in-session role-switch | \`question\` (permission-gated; \`permission.question: "allow"\`) | \`npx cclaw-cli doctor\` |
-| Codex | \`role-switch\` | in-session role-switch (mandatory evidenceRefs) | \`request_user_input\` (experimental; Plan / Collaboration mode) | \`npx cclaw-cli doctor\` |
+| OpenCode | \`native\` | generated \`.opencode/agents/<agent>.md\` subagents via Task / \`@agent\` mention | \`question\` (permission-gated; \`permission.question: "allow"\`) | \`npx cclaw-cli doctor\` |
+| Codex | \`native\` | generated \`.codex/agents/<agent>.toml\` custom agents via native parallel subagent spawning | \`request_user_input\` (experimental; Plan / Collaboration mode) | \`npx cclaw-cli doctor\` |
 
 **Dispatch rules driven by \`subagentFallback\`:**
 
 - \`native\` — use the harness's own named subagent primitive; delegation entry uses \`fulfillmentMode: "isolated"\`.
 - \`generic-dispatch\` — map each cclaw agent onto the generic dispatcher with a role prompt; delegation entry uses \`fulfillmentMode: "generic-dispatch"\`.
-- \`role-switch\` — announce the role in-session, perform the work, append a delegation row with \`fulfillmentMode: "role-switch"\` and ≥1 \`evidenceRef\`. Without evidenceRefs the \`delegation:mandatory:current_stage\` check reports \`missingEvidence\` and blocks stage completion.
+- \`role-switch\` — degraded fallback only when the active runtime cannot expose its declared dispatch surface. Announce the role in-session, perform the work, append a delegation row with \`fulfillmentMode: "role-switch"\` and ≥1 \`evidenceRef\`. Without evidenceRefs the \`delegation:mandatory:current_stage\` check reports \`missingEvidence\` and blocks stage completion.
 
-The only time a \`harness_limitation\` waiver fires automatically is when every installed harness declares \`subagentFallback: "waiver"\`. cclaw 0.33 no longer maps Codex onto auto-waiver — the agent must role-switch with evidence.
+### Native dispatch contract
+
+Use real harness subagents for OpenCode and Codex:
+
+1. OpenCode: invoke the generated \`.opencode/agents/<agent>.md\` subagent via Task or \`@<agent>\`. Built-in \`general\` / \`explore\` remain fallback subagent types for ad hoc tasks, but cclaw's core roles are generated by name.
+2. Codex: ask Codex to spawn the generated \`.codex/agents/<agent>.toml\` custom agent(s) by name; for review-style independent lanes, request parallel spawning and wait for all results before reconciliation.
+3. Claude: use the native named Task subagent. Cursor: map the cclaw role onto the generic Task/Subagent surface with a self-contained prompt.
+4. Produce stage output in the current artifact, with anchors suitable for \`evidenceRefs\`.
+5. Append delegation ledger rows with \`stage\`, \`agent\`, \`mode\`, \`status: "completed"\`, and \`fulfillmentMode\` matching the dispatch mode (\`"isolated"\` for Claude/OpenCode/Codex, \`"generic-dispatch"\` for Cursor).
+
+The only time a \`harness_limitation\` waiver fires automatically is when every installed harness declares \`subagentFallback: "waiver"\`. Do not map Codex or OpenCode onto auto-waiver or default role-switch; they have true subagent surfaces.
 
 ### Model routing
 
@@ -120,9 +132,9 @@ Concrete per-stage rules so the controller does not have to guess which tier fit
 
 | Stage | Deep slot | Balanced slot(s) | Fast fan-out | Trigger to escalate |
 |---|---|---|---|---|
-| brainstorm | planner (only if ambiguity spans >1 module) | — | run in-thread research playbooks | promote to \`balanced\` reviewer once direction locks |
-| scope | planner (always) | — | run \`research/git-history.md\` in-thread when churn is high | promote to \`balanced\` planner if scope touches external contracts |
-| design | planner (always) | security-reviewer (if trust boundary touched) | run \`research/framework-docs-lookup.md\` + \`research/best-practices-lookup.md\` in-thread | escalate one specialist to \`deep\` only if a failure mode is Critical-severity |
+| brainstorm | planner (only if ambiguity spans >1 module) | product-manager / critic when product value or premise is uncertain | run in-thread research playbooks | promote to \`balanced\` critic if the do-nothing path may beat the idea |
+| scope | planner (always) | product-manager / critic when mode changes user value or boundaries are soft | run \`research/git-history.md\` in-thread when churn is high | promote to \`balanced\` critic if scope mode is disputed |
+| design | planner (always) | critic, security-reviewer, test-author when alternatives/trust/testability apply | run \`research/framework-docs-lookup.md\` + \`research/best-practices-lookup.md\` in-thread | escalate one specialist to \`deep\` only if a failure mode is Critical-severity |
 | spec | — | reviewer (if spec > 200 lines or multiple ACs) | — | escalate to \`deep\` only for spec ↔ design contradictions |
 | plan | planner (solo, always) | — | — | never fan out at plan stage; one owner for dependency graph |
 | tdd | — | ${formatAgentList(stageSummary("tdd").primaryAgents)} (per slice, carrying RED/GREEN/REFACTOR evidence) · reviewer (slice-local only when sliceReview triggers) | doc-updater (API surface changes) | escalate to \`deep\` only when a RED test cannot be expressed (design leak) |
@@ -591,6 +603,56 @@ Output format (mandatory):
 - Close with RISK_SUMMARY and SHIP_BLOCKERS (explicit list, possibly empty).
 \`\`\`
 
+`;
+}
+function productManagerEnhancedBody() {
+    return `
+
+## Task Tool Delegation
+
+Use this payload when product discovery needs an isolated lens:
+
+\`\`\`
+You are a product-manager subagent.
+
+DISCOVERY GOAL: {problem/value decision to clarify}
+CONTEXT: {existing artifact excerpts, user segment, constraints}
+DEPTH: {lite|standard|deep}
+
+Required output:
+- PERSONA_JTBD: persona, job, pain/trigger
+- VALUE_HYPOTHESIS: expected value and success metric
+- EVIDENCE_SIGNAL: strongest evidence, weakest assumption
+- WHY_NOW_AND_DO_NOTHING: why now plus consequence of no action
+- NON_GOALS: explicit exclusions
+- SCOPE_HANDOFF: one recommendation for hold/selective/expand/reduce
+\`\`\`
+
+`;
+}
+function criticEnhancedBody() {
+    return `
+
+## Task Tool Delegation
+
+Use this payload when a premise, scope mode, or engineering path needs adversarial pressure:
+
+\`\`\`
+You are a critic subagent.
+
+DECISION_UNDER_REVIEW: {direction/scope/design choice}
+CONTEXT: {artifact excerpts, constraints, known risks}
+DEPTH: {lite|standard|deep}
+
+Required output:
+- PREMISE_ATTACK: what could make this decision wrong
+- CHEAPER_ALTERNATIVE: smaller or more reversible option
+- SHADOW_ALTERNATIVE: viable competing path
+- SWITCH_TRIGGER: signal that should change the decision
+- FAILURE_RESCUE: likely failure and rescue/degraded behavior
+- VERIFICATION_EVIDENCE: evidence needed before locking
+\`\`\`
+
 `;
 }
 function reviewerEnhancedBody() {
@@ -679,6 +741,10 @@ export function enhancedAgentBody(agentName) {
     switch (agentName) {
         case "planner":
             return plannerEnhancedBody();
+        case "product-manager":
+            return productManagerEnhancedBody();
+        case "critic":
+            return criticEnhancedBody();
         case "reviewer":
             return reviewerEnhancedBody();
         case "security-reviewer":

package/dist/content/templates.js

@@ -29,10 +29,28 @@ export const ARTIFACT_TEMPLATES = {
 ### Discovered context
 - (paths, prior artifacts, seeds, prompt fragments — referenced by downstream stages, or \`- None.\`)
 
-## Problem
-- **What we're solving:**
-- **Success criteria:**
-- **Constraints:**
+## Problem Decision Record
+- **Depth:** lite | standard | deep
+- **Frame type:** product | technical-maintenance
+
+### Product framing (use when applicable)
+- **Persona / user:**
+- **Job to be done:**
+- **Pain / trigger:**
+- **Value hypothesis:**
+- **Evidence / signal:**
+- **Success metric:**
+- **Why now:**
+- **Do-nothing consequence:**
+- **Non-goals:**
+
+### Technical-maintenance framing (use when product framing is not applicable)
+- **Affected operator/developer:**
+- **Current failure mode:**
+- **Expected operational improvement:**
+- **Verification signal:**
+- **Do-nothing cost:**
+- **Non-goals:**
 
 ## Premise Check
 - **Right problem?** (yes/no + one-line justification — take a position)
@@ -43,11 +61,10 @@ export const ARTIFACT_TEMPLATES = {
 - *How might we …?* — one line naming the user, the desired outcome, and the binding constraint.
 
 ## Sharpening Questions
+> Ask one decision-changing question at a time. For concrete early exits, record \`None - early exit\` with rationale.
 | # | Question | Answer / Assumption | Decision impact |
 |---|---|---|---|
 | 1 |  |  |  |
-| 2 |  |  |  |
-| 3 |  |  |  |
 
 ## Clarifying Questions
 | # | Question | Answer | Decision impact |
@@ -55,7 +72,7 @@ export const ARTIFACT_TEMPLATES = {
 | 1 |  |  |  |
 
 ## Approach Tier
-- Tier: Lightweight | Standard | Deep
+- Tier: lite | standard | deep
 - Why this tier:
 
 ## Short-Circuit Decision
@@ -80,7 +97,7 @@ export const ARTIFACT_TEMPLATES = {
 - **Approach:**
 - **Rationale:** Trace this to the prior Approach Reaction.
 - **Approval:** pending
-- **Next-stage handoff:** On standard track, hand this to \`scope\`; on medium track, hand this directly to \`spec\` with explicit requirements/constraints.
+- **Next-stage handoff:** On standard track, hand this to \`scope\`; on medium track, hand this directly to \`spec\`. Include upstream decisions used, drift, confidence, unresolved questions, risk hints, and non-goals.
 
 ## Not Doing
 - (3-5 things this brainstorm is *not* committing to — distinct from \`Deferred\`. These will not appear in scope unless the user explicitly opts in.)
@@ -165,8 +182,21 @@ ${SEED_SHELF_SECTION}
 | HOUR 4-5 (integration) |  |  |  |
 | HOUR 6+ (polish/tests) |  |  |  |
 
+## Scope Contract
+- **Selected mode:** HOLD SCOPE | SELECTIVE EXPANSION | SCOPE EXPANSION | SCOPE REDUCTION
+- **In scope:**
+- **Out of scope:**
+- **Requirements:**
+- **Locked decisions:**
+- **Discretion areas:**
+- **Deferred ideas:**
+- **Accepted reference ideas:**
+- **Rejected reference ideas:**
+- **Success definition:**
+- **Design handoff:**
+
 ## Scope Mode
-- [ ] SCOPE EXPANSION — dream bigger; user explicitly opts into the larger product slice.
+- [ ] SCOPE EXPANSION — explore ambitious alternatives; user explicitly opts into the larger product slice.
 - [ ] SELECTIVE EXPANSION — hold baseline scope and cherry-pick one high-leverage addition.
 - [ ] HOLD SCOPE — preserve the approved brainstorm direction with maximum rigor.
 - [ ] SCOPE REDUCTION — strip to the smallest useful wedge when risk/blast radius is too high.
@@ -174,9 +204,24 @@ ${SEED_SHELF_SECTION}
 ## Mode-Specific Analysis
 | Selected mode | Rationale | Depth |
 |---|---|---|
-|  |  | default / deep |
+|  |  | lite / standard / deep |
+
+> Default path: one selected-mode row plus rationale. Deep/high-risk scope may expand with optional evidence headings below.
+
+## Landscape Check
+- Optional for EXPAND/SELECTIVE/deep; omit for compact HOLD SCOPE.
+
+## Taste Calibration
+- Optional quality-bar references from in-repo modules/files.
 
-> Default path: one selected-mode row plus rationale. Deep/high-risk scope may expand below with mode-specific analysis.
+## Reference Pull
+- Optional evidence from \`/Users/zuevrs/Downloads/references\`; list accepted/rejected ideas or \`Not needed - compact scope\`.
+
+## Ambitious Alternatives
+- Optional for SCOPE EXPANSION/SELECTIVE; list larger alternatives and disposition.
+
+## Ruthless Minimum Slice
+- Optional for SCOPE REDUCTION/high-risk scope; define the smallest useful wedge.
 
 ## Requirements (stable IDs)
 | ID | Requirement (observable outcome) | Priority | Source (origin doc / prompt line) |
@@ -241,6 +286,9 @@ ${SEED_SHELF_SECTION}
 
 ## Scope Summary
 - Selected mode: (one of \`SCOPE EXPANSION\` | \`SELECTIVE EXPANSION\` | \`HOLD SCOPE\` | \`SCOPE REDUCTION\`)
+- Confidence: high | medium | low
+- Drift from brainstorm: None / <specific drift>
+- Unresolved questions: None / <questions>
 - Strongest challenges resolved:
 - Recommended path:
 - Accepted scope:
@@ -291,7 +339,7 @@ ${SEED_SHELF_SECTION}
 
 ## Compact-First Scaffold
 - Default to the compact design spine unless risk requires Standard/Deep add-ons.
-- Compact required spine: Codebase Investigation, Architecture Boundaries, Architecture Diagram, Data Flow, Failure Mode Table, Test Strategy, and Completion Dashboard.
+- Compact required spine: Upstream Handoff, Codebase Investigation, Engineering Lock, Architecture Boundaries, Architecture Diagram, Data Flow, Failure Mode Table, Test Strategy, Spec Handoff, and Completion Dashboard.
 - Mark optional Standard/Deep sections as \`Omitted - compact design\` when they do not apply; do not expand the scaffold just to fill empty tables.
 
 ## Upstream Handoff
@@ -302,9 +350,14 @@ ${SEED_SHELF_SECTION}
 - Drift from upstream (or \`None\`):
 
 ## Codebase Investigation
-| File | Current responsibility | Patterns discovered |
-|---|---|---|
-|  |  |  |
+| File | Current responsibility | Patterns discovered | Existing fit / reuse candidate |
+|---|---|---|---|
+|  |  |  |  |
+
+## Engineering Lock
+| Decision area | Chosen path | Shadow alternative | Switch trigger | Failure/rescue/degraded behavior | Verification evidence | Confidence |
+|---|---|---|---|---|---|---|
+|  |  |  |  |  |  |  |
 
 ## Search Before Building
 | Layer | Label | What to reuse first |
@@ -336,9 +389,9 @@ ${MARKDOWN_CODE_FENCE}
 ## Data-Flow Shadow Paths
 - Standard/Deep add-on; omit when compact design does not need a shadow path.
 <!-- diagram: data-flow-shadow-paths -->
-| Path | Trigger | Fallback/Degrade behavior |
-|---|---|---|
-|  |  |  |
+| Chosen path | Shadow alternative | Switch trigger | Failure/rescue/degraded behavior | Verification evidence |
+|---|---|---|---|---|
+|  |  |  |  |  |
 
 ## Error Flow Diagram
 - Standard/Deep add-on; omit when the Failure Mode Table is sufficient.
@@ -387,6 +440,8 @@ ${MARKDOWN_CODE_FENCE}
 |  |  |  |  |
 
 ## Data Flow
+- Data/state flow:
+- Critical path:
 - Happy path:
 - Nil/empty input path:
 - Upstream error path:
@@ -431,6 +486,23 @@ ${MARKDOWN_CODE_FENCE}
 |---|---|---|
 |  |  |  |
 
+## Rejected Alternatives
+| Alternative | Why rejected | Revival signal |
+|---|---|---|
+|  |  |  |
+
+## Design Decisions
+| Decision Ref | Requirement / LD refs | Decision | Spec impact |
+|---|---|---|---|
+| DD-1 |  |  |  |
+
+## Spec Handoff
+- Requirements to carry forward:
+- Design decisions to encode:
+- Risks and rescue paths:
+- Test/performance expectations:
+- Unresolved questions (or \`None\`):
+
 ## Outside Voice Findings
 | ID | Dimension | Finding | Disposition | Rationale |
 |---|---|---|---|---|
@@ -645,7 +717,7 @@ Execution rule: complete and verify each batch before starting the next batch.
 # TDD Artifact
 
 ## Upstream Handoff
-- Source artifacts: \`04-spec.md\`; \`05-plan.md\` when present. Quick track uses spec acceptance items / bug reproduction slices instead of nonexistent plan tasks.
+- Source artifacts: \`04-spec.md\` plus the active track's upstream source item (plan slice on standard/medium, spec acceptance item or bug reproduction slice on quick).
 - Decisions carried forward:
 - Constraints carried forward:
 - Open questions:
@@ -672,11 +744,11 @@ Execution rule: complete and verify each batch before starting the next batch.
 | S-1 |  |  |  |
 
 ## Acceptance Mapping
-| Slice | Plan task ID or quick source | Spec criterion ID |
+| Slice | Source item ID | Spec criterion ID |
 |---|---|---|
-| S-1 | T-1 / QS-1 | AC-1 |
+| S-1 | SRC-1 | AC-1 |
 
-> On quick track, map to the \`Quick Reproduction Contract\` bug slice or spec acceptance item. Do not invent a plan task just to satisfy this table.
+> Map each slice to the active track's source item: plan slice on standard/medium, or the \`Quick Reproduction Contract\` bug slice / spec acceptance item on quick.
 
 ## Failure Analysis
 | Slice | Expected missing behavior | Actual failure reason |
@@ -693,7 +765,7 @@ Execution rule: complete and verify each batch before starting the next batch.
 - Behavior preserved:
 
 ## Traceability
-- Plan task IDs:
+- Source item IDs:
 - Spec criterion IDs:
 
 
@@ -729,22 +801,46 @@ Execution rule: complete and verify each batch before starting the next batch.
 # Review Artifact
 
 ## Upstream Handoff
-- Source artifacts: \`04-spec.md\`, \`06-tdd.md\`; \`05-plan.md\` only when present. Quick track reviews spec acceptance items / bug reproduction slices without requiring plan-task coverage.
+- Source artifacts: \`04-spec.md\`, \`06-tdd.md\`, plus the active track's upstream source item when available.
 - Decisions carried forward:
 - Constraints carried forward:
 - Open questions:
 - Drift from upstream (or \`None\`):
 
+## Review Evidence Scope
+- Base/head:
+- Files inspected:
+- Changed-file coverage summary:
+- Diagnostics run:
+- Omitted files with explicit reason:
+- Reviewer delegation evidence:
+- Security-reviewer delegation evidence:
+
+## Changed-File Coverage
+| File | Coverage status | Evidence / no-impact reason |
+|---|---|---|
+|  | inspected / broader-module / omitted-no-impact |  |
+
 ## Layer 1 Verdict
 | Criterion | Verdict | Evidence |
 |---|---|---|
 | AC-1 | PASS/FAIL |  |
 
 ## Layer 2 Findings
-| ID | Severity | Category | Description | Status |
-|---|---|---|---|---|
-| R-1 | Critical/Important/Suggestion | correctness/security/performance/architecture/external-safety |  | open/resolved |
-- NO_CHANGE_ATTESTATION: <required when Category=security has no entries; explain why no security-relevant changes were detected>
+| ID | Severity | Category | File:line / no-line reason | Description | Status |
+|---|---|---|---|---|---|
+| R-1 | Critical/Important/Suggestion | correctness/security/performance/architecture/external-safety | path:line |  | open/resolved |
+- NO_FINDINGS_ATTESTATION: <required when no findings are reported; cite inspected coverage>
+
+## Security Sweep Attestation
+- Result: findings | NO_CHANGE_ATTESTATION | NO_SECURITY_IMPACT
+- Inspected surfaces:
+- Rationale:
+
+## Dependency & Version Audit
+- Relevant: yes/no
+- Manifests/lockfiles/generated clients/CI/runtime config/external APIs inspected:
+- Result / no-impact rationale:
 
 ## Incoming Feedback Queue
 | ID | Source | Severity | File:line | Request | Status | Evidence |
@@ -766,15 +862,15 @@ Execution rule: complete and verify each batch before starting the next batch.
 
 ## Completeness Snapshot
 - AC coverage: <N>/<M> (<percent>%)
-- Task coverage (tasks backed by ≥1 test slice): <N>/<M> or \`N/A - quick track has no plan artifact\`
+- Source item coverage (source items backed by ≥1 test slice): <N>/<M> or \`N/A - direct spec/reproduction coverage\`
 - Slice coverage (slices linked to ≥1 AC or bug reproduction slice): <N>/<M>
 - Adversarial review: not triggered | pass | fail
 - Overall: complete | concerns | blocked
 
 ## Trace Matrix Check
-- Command: \`cclaw internal trace-matrix\` when plan artifacts exist or the active track enforces it; quick track may record direct AC/reproduction-slice coverage instead.
+- Command: \`cclaw internal trace-matrix\` when the active track enforces it; otherwise record direct AC/reproduction-slice coverage.
 - Orphaned criteria: 0
-- Orphaned tasks: 0 or \`N/A - quick track\`
+- Orphaned source items: 0 or \`N/A - direct spec/reproduction coverage\`
 - Orphaned tests: 0
 - Evidence ref:
 

package/dist/content/track-render-context.js

@@ -34,8 +34,15 @@ export function renderTrackTerminology(value, context) {
     }
     return value
         .replace(/\btask from the plan\b/giu, `${context.traceabilitySourceNoun} from the spec`)
+        .replace(/\bplan confirmation\b/giu, "spec approval")
+        .replace(/\bplan approval\b/giu, "spec approval")
+        .replace(/\bapproved plan slice\b/giu, `approved ${context.traceabilitySliceNoun}`)
+        .replace(/\bplanned slice\b/giu, context.traceabilitySliceNoun)
         .replace(/\bplan task ID\b/giu, context.traceabilityIdNoun)
         .replace(/\bplan task\b/giu, context.traceabilitySourceNoun)
+        .replace(/\bplan-task\b/giu, "acceptance-criterion")
+        .replace(/\btask coverage\b/giu, "source item coverage")
+        .replace(/\borphaned tasks\b/giu, "orphaned source items")
         .replace(/\bplan row\b/giu, "acceptance row")
         .replace(/\btraceable to plan slice\b/giu, `traceable to ${context.traceabilitySliceNoun}`)
         .replace(/\bplan slice\b/giu, context.traceabilitySliceNoun)

package/dist/delegation.d.ts

@@ -6,8 +6,8 @@ export type DelegationStatus = "scheduled" | "completed" | "failed" | "waived";
  * How a delegation was actually fulfilled. Advisory — mirrors the harness
  * `subagentFallback` that was in effect when the entry was recorded.
  *
- * - `isolated`         — Claude-style isolated subagent worker.
- * - `generic-dispatch` — Cursor-style Task dispatch mapped to a named role.
+ * - `isolated`         — native isolated subagent worker (Claude/OpenCode/Codex).
+ * - `generic-dispatch` — generic Task/Subagent dispatch mapped to a named role.
  * - `role-switch`      — performed in-session with explicit role announce.
  * - `harness-waiver`   — auto-waived due to missing dispatch capability.
  */

package/dist/delegation.js

@@ -18,6 +18,13 @@ function delegationLockPath(projectRoot) {
 function createSpanId() {
     return `dspan-${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 10)}`;
 }
+function activeHarnessSubagentFallback() {
+    const activeHarness = process.env.CCLAW_ACTIVE_HARNESS;
+    if (!activeHarness)
+        return undefined;
+    return HARNESS_ADAPTERS[activeHarness]
+        ?.capabilities.subagentFallback;
+}
 async function resolveReviewDiffBase(projectRoot) {
     let head = "";
     try {
@@ -229,10 +236,7 @@ export async function appendDelegation(projectRoot, entry) {
             stamped.evidenceRefs = [];
         }
         if (stamped.status === "completed" && stamped.fulfillmentMode === undefined) {
-            const activeFallback = process.env.CCLAW_ACTIVE_HARNESS
-                ? HARNESS_ADAPTERS[process.env.CCLAW_ACTIVE_HARNESS]
-                    ?.capabilities.subagentFallback
-                : undefined;
+            const activeFallback = activeHarnessSubagentFallback();
             if (activeFallback) {
                 stamped.fulfillmentMode = expectedFulfillmentMode([activeFallback]);
             }
@@ -291,8 +295,9 @@ export async function checkMandatoryDelegations(projectRoot, stage, options = {}
     const missingEvidence = [];
     const config = await readConfig(projectRoot).catch(() => null);
     const harnesses = config?.harnesses ?? [];
-    const fallbacks = harnesses.map((h) => HARNESS_ADAPTERS[h].capabilities.subagentFallback);
-    const expectedMode = expectedFulfillmentMode(fallbacks);
+    const configuredFallbacks = harnesses.map((h) => HARNESS_ADAPTERS[h].capabilities.subagentFallback);
+    const activeFallback = activeHarnessSubagentFallback();
+    const expectedMode = expectedFulfillmentMode(activeFallback ? [activeFallback] : configuredFallbacks);
     for (const agent of mandatory) {
         const rows = forRun.filter((e) => e.agent === agent);
         const completedRows = rows.filter((e) => e.status === "completed");
@@ -307,9 +312,11 @@ export async function checkMandatoryDelegations(projectRoot, stage, options = {}
         if (hasWaived) {
             waived.push(agent);
         }
-        // Evidence is required for any non-isolated completion mode. Legacy rows
-        // without fulfillmentMode are inferred to `isolated` during parse.
-        const evidenceRequired = completedRows.some((e) => (e.fulfillmentMode ?? "isolated") !== "isolated");
+        // Evidence is required for non-isolated completions and for explicit
+        // degraded role-switch rows. Native OpenCode/Codex/Claude isolated
+        // dispatch is accepted as true subagent work; role-switch remains a
+        // fallback that must point at artifact evidence.
+        const evidenceRequired = expectedMode !== "isolated" || completedRows.some((e) => (e.fulfillmentMode ?? "isolated") !== "isolated");
         if (hasCompleted &&
             evidenceRequired &&
             !completedRows.some((e) => Array.isArray(e.evidenceRefs) && e.evidenceRefs.length > 0)) {

package/dist/doctor-registry.js

@@ -39,7 +39,7 @@ const RULES = [
         }
     },
     {
-        test: /^(dir:|command:|utility_command:|skill:|utility_skill:|agent:|harness_tool_ref:|harness_ref:|stage_examples_ref:|doctor_ref:)/,
+        test: /^(dir:|command:|utility_command:|stage_command:|skill:|utility_skill:|agent:|harness_tool_ref:|harness_ref:|stage_examples_ref:|doctor_ref:)/,
         metadata: {
             severity: "error",
             summary: "Generated runtime surface presence check.",