lee-spec-kit 0.9.3 → 0.9.5

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "lee-spec-kit",
-  "version": "0.9.3",
+  "version": "0.9.5",
   "description": "Document-centered harness engineering toolkit for AI agent development",
   "type": "module",
   "bin": {

package/templates/en/common/README.md

@@ -117,6 +117,10 @@ When requirements/scope change, the “what to update” must be explicit in doc
 ## CLI Config (`.lee-spec-kit.json`)
 
 When you run `lee-spec-kit init`, it creates `.lee-spec-kit.json` in the docs root (default: `docs/`).
+Interactive init lets you keep the recommended defaults or customize task
+implementation delegation, Plan/Task/Feature reviews, and Local integration.
+Non-interactive setup exposes the same choices through `--task-agent`,
+`--reviews`, `--max-review-rounds`, and `--completion-strategy`.
 
 - Used by `lee-spec-kit feature`, `config`, `update`, `detect`, and workflow validators to resolve docs location / project type / language.
 - `docsRepo`, `pushDocs`, `docsRemote` are metadata for the CLI-managed **Docs Push policy** (the CLI does not auto-push).
@@ -130,11 +134,24 @@ When you run `lee-spec-kit init`, it creates `.lee-spec-kit.json` in the docs ro
 - `docsRepo` ("embedded" | "standalone"): How docs are managed
 - `pushDocs` (boolean, optional): Only written when `docsRepo: "standalone"` (whether to push to remote)
 - `docsRemote` (string, optional): Only written when `pushDocs: true` (remote repo URL)
-- `workflow.prePrReview.reviewer` (object): Pre-PR subagent execution settings
+- `workflow.agentExecution.task` (object): task implementation delegation settings
+  - `enabled`: delegates each `task_execute` action to a subagent; new and updated projects default to `true`
   - `type`: currently only `"subagent"` is supported
   - `model`: `"inherit"` or a model name supported by the runtime
   - `reasoningEffort`: `low | medium | high | xhigh | max | ultra`
   - `onUnavailable`: `inherit | error` when the requested model is unavailable
+  - `workflow-stage` returns a stable task ID, working/docs directories, and a machine-readable `workerContract`. The worker executes directly without rerunning `workflow-stage` or delegating again.
+  - The implementation subagent can edit project code and run task-scoped checks. The main agent retains docs, task state, commits, approvals, and remote actions; official hooks reject commits while `task_execute` is active.
+- `workflow.agentReview.maxRounds` (positive integer): maximum automatic finding-remediation passes shared by Plan/task/Feature reviews before `workflow-stage` returns `review_escalation`; the initial review is not counted and the default is `1`
+- `workflow.agentReview.plan` / `workflow.agentReview.task` / `workflow.agentReview.feature` (object): Plan/task/Feature independent review settings
+  - `enabled`: enables that review gate; new projects default Plan and Feature to `true`, and task to `false`
+  - `evidenceMode`: `path_required | any`
+  - `reviewer`: fresh read-only subagent execution settings
+  - `type`: currently only `"subagent"` is supported
+  - `model`: `"inherit"` or a model name supported by the runtime
+  - `reasoningEffort`: `low | medium | high | xhigh | max | ultra`
+  - `onUnavailable`: `inherit | error` when the requested model is unavailable
+  - Plan review runs before Plan approval and binds its evidence to the exact `specHash` and `planHash` returned by `workflow-stage`; later spec/plan content changes require a fresh review
 - `workflow.baseBranch` (string): branch that receives a completed local Feature
 - `workflow.completionStrategy` (`"local-ff" | "local-squash" | "none"`): fast-forward, create one verified squash commit, or explicitly finish without integration
 - `workflow.deleteFeatureBranchAfterMerge` (boolean): delete the integrated local Feature branch after cleanup; remote branches are never deleted
@@ -171,14 +188,47 @@ When you run `lee-spec-kit init`, it creates `.lee-spec-kit.json` in the docs ro
     "deleteFeatureBranchAfterMerge": true,
     "featureChecks": [],
     "postMergeChecks": [],
-    "prePrReview": {
-      "evidenceMode": "path_required",
-      "reviewer": {
+    "agentExecution": {
+      "task": {
+        "enabled": true,
         "type": "subagent",
         "model": "inherit",
         "reasoningEffort": "high",
         "onUnavailable": "inherit"
       }
+    },
+    "agentReview": {
+      "maxRounds": 1,
+      "plan": {
+        "enabled": true,
+        "evidenceMode": "path_required",
+        "reviewer": {
+          "type": "subagent",
+          "model": "inherit",
+          "reasoningEffort": "high",
+          "onUnavailable": "inherit"
+        }
+      },
+      "task": {
+        "enabled": false,
+        "evidenceMode": "path_required",
+        "reviewer": {
+          "type": "subagent",
+          "model": "inherit",
+          "reasoningEffort": "high",
+          "onUnavailable": "inherit"
+        }
+      },
+      "feature": {
+        "enabled": true,
+        "evidenceMode": "path_required",
+        "reviewer": {
+          "type": "subagent",
+          "model": "inherit",
+          "reasoningEffort": "high",
+          "onUnavailable": "inherit"
+        }
+      }
     }
   },
   "allowedDocsEntries": {
@@ -192,7 +242,7 @@ When you run `lee-spec-kit init`, it creates `.lee-spec-kit.json` in the docs ro
 }
 ```
 
-New local projects use `local-ff`. Choose `local-squash` to create one base-branch commit while preserving the source Feature tip under an internal `refs/lee-spec-kit/integrations/*` ref for task-checkpoint evidence. During `update`, an existing local project with no explicit `completionStrategy` receives `none` so an upgrade does not unexpectedly merge its current branch. Set it to `local-ff` or `local-squash` deliberately when ready.
+New and updated projects delegate task implementation and enable Plan review by default; set `workflow.agentExecution.task.enabled` or `workflow.agentReview.plan.enabled` to `false` to opt out. New local projects use `local-ff` with Feature agent review enabled. Set `workflow.agentReview.task.enabled` to `true` when every task also needs independent review. Finding remediation defaults to `1`: findings from the first review are applied once and reviewed again, and another `changes_requested` decision returns `review_escalation`. Increase `workflow.agentReview.maxRounds` before authorizing another remediation and fresh review. Choose `local-squash` to create one base-branch commit while preserving the source Feature tip under an internal `refs/lee-spec-kit/integrations/*` ref for task-checkpoint evidence. During `update`, an existing local project with no explicit `completionStrategy` receives `none` so an upgrade does not unexpectedly merge its current branch. Set it to `local-ff` or `local-squash` deliberately when ready.
 
 ```json
 {

package/templates/en/common/agents/agents.md

@@ -52,9 +52,14 @@ This document defines workflow policy, not a custom runtime loop.
 
 - lee-spec-kit owns docs structure, workflow stages, and validators.
 - Codex owns the execution loop, tool usage, and hook lifecycle.
-- Do not start implementation unless `workflow-stage --json` reports `stage === "implementation"` and `implementationAllowed === true`.
-- When `workflow-stage --json` returns `nextAction.category: pre_pr_review` with `executor: subagent`, run a fresh, read-only subagent review using the returned `model`, `reasoningEffort`, and `onUnavailable` policy. Do not select or require a named review skill.
-- The Pre-PR review subagent returns findings without modifying code. The main agent remediates findings and records reviewer metadata and the final decision as evidence.
+- Modify implementation code only when `implementationAllowed === true`. Normal task work uses `stage === "implementation"`; review fixes use `task_review_fix` or `feature_review_fix`, and verification fixes use `feature_remediation`.
+- When `nextAction.category` is `plan_review` with `executor: subagent`, delegate a fresh read-only review of `spec.md` and `plan.md` for the returned `specHash` and `planHash`. The main agent records the returned `reviewRound`, Plan Review evidence, decision, reviewer metadata, and both hashes. Any later spec/plan content change invalidates that review.
+- When `nextAction.category` is `task_execute` with `executor: subagent`, mark that one task active, then delegate its implementation and task-scoped checks to a fresh subagent in the returned `workingDirectory` with the returned model, reasoning effort, unavailability policy, and exact `workerContract`. No named execution skill is required.
+- The implementation worker executes directly, follows the approved Verification Contract, and does not add unplanned durable tests. It must not run `workflow-stage` or spawn another subagent. It may edit project code and run scoped checks, but it must not edit lee-spec-kit docs, change task state, commit, request approvals, or perform remote/destructive actions. The main agent inspects the result and owns docs synchronization, task transitions, commits, and workflow continuation; official hooks block commits while `task_execute` remains active.
+- When `nextAction.category` is `task_review` with `executor: subagent`, delegate a fresh read-only review for the returned task ID and SHA/tree range, then record the returned `reviewRound`.
+- When `nextAction.category` is `pre_pr_review` with `executor: subagent`, run a fresh read-only Feature review using the returned model, reasoning effort, `reviewRound`, and SHA/tree range. Do not select or require a named review skill.
+- Review subagents return findings without modifying code. The main agent remediates findings and records reviewer metadata, reviewed scope, evidence, decision, and exact hash/SHA/tree target metadata.
+- `workflow.agentReview.maxRounds` counts automatic finding-remediation passes, not the initial review. When `nextAction.category` is `review_escalation`, those passes have been exhausted. Do not remediate or delegate another review. Keep implementation blocked, present the exact returned action options, and require `maxRounds` to be increased before another remediation pass.
 - Treat spec/plan/tasks approval, issue creation, and branch creation as hard gates before implementation.
 - In standalone mode, do not hand-write `git worktree add`; run the exact `nextAction.command` from `workflow-stage` so the managed workspace path, stale directory cleanup, and `.env`/`.env.*` copy step stay consistent.
 - In local mode, do not stop after implementation approval. Follow the exact `local verify`, `local merge`, and `local cleanup` commands returned by `workflow-stage` until verified integration and cleanup produce `done`. A `feature_remediation` stage explicitly permits fixes in the Feature worktree.

package/templates/en/common/agents/skills/create-feature.md

@@ -20,10 +20,12 @@ This guide defines how to start or continue a feature in the Codex-native lee-sp
 - Docs are the SSOT. Follow the active feature docs directly.
 - Progress through the documented stages directly:
   - `spec.md` defines scope and review state
-  - `plan.md` defines the implementation approach
+  - `plan.md` defines the implementation approach and Verification Contract
   - `tasks.md` drives execution order
   - `issue.md` / `pr.md` are part of the stage gate once the feature reaches GitHub workflow stages
 - Do not begin implementation just because `tasks.md` exists. Implementation starts only when `workflow-stage --json` allows it.
+- When Plan review is enabled, move `plan.md` to Review, delegate the returned fresh read-only `plan_review`, and record its `reviewRound`, evidence, decision, reviewer metadata, `specHash`, and `planHash`. Only an approved review of the current hashes can reach Plan approval. The reviewer challenges NONE/UPDATE/ADD decisions, requirement coverage, independent oracles, stable observation boundaries, realistic failure/rollback cases, exclusions, and focused/full verification scope without editing docs.
+- `workflow.agentReview.maxRounds` counts Plan finding-remediation passes, not the initial review. If `workflow-stage` returns `review_escalation`, stop automatic Plan remediation and present its exact action options. Another remediation requires increasing the setting first.
 - When scope or behavior changes, update the active feature docs in the same turn before continuing.
 - Ask for approval at documented review checkpoints and before remote or destructive actions.
 - Use `npx lee-spec-kit commit-audit --json` before `git commit` when docs-path validation matters.

package/templates/en/common/agents/skills/create-pr.md

@@ -25,15 +25,18 @@ When `workflow-stage --json` returns `nextAction.executor: subagent`, delegate t
 2. Inspect regression, exception handling, critical/security risks, side effects, user flow impact, and release readiness.
 3. Check maintainability: split oversized functions/files when needed, reuse/integrate existing code where appropriate, and remove obsolete code.
 4. Judge whether the implementation actually fits the feature intent and scope documented in `spec.md` / `plan.md` / `tasks.md`.
-5. When `workflow.prePrReview.evidenceMode=path_required` (default), generate a real review artifact such as `review-trace.json` before approval. In `evidenceMode=any`, direct record mode without a separate artifact is also allowed unless execution evidence is explicitly enforced.
+5. When `workflow.agentReview.feature.evidenceMode=path_required` (default), generate a real review artifact such as `review-trace.json` before approval. In `evidenceMode=any`, direct record mode without a separate artifact is also allowed unless execution evidence is explicitly enforced.
 6. `Pre-PR Evidence` should follow the configured evidence policy. In `path_required`, it must point to a real existing path.
 7. Record `Summary`, `Feature Intent Summary`, `Implementation Fit`, `Missing Cases`, `Spec Alignment Checked`, `Finding Count`, `Blocking Findings`, `Findings`, and `Residual Risks` with non-placeholder content.
 8. Use `commandsExecuted` only for optional audit/targeted verification that you actually chose to run during review.
 9. In code-review stage, keep `PR Review Evidence/Decision` aligned with `decisions.md` by adding a `PR Review Log` section with `Summary` and `Decision`.
 10. `Pre-PR Decision` must use `decision: approve|changes_requested|blocked ...` (or `결정: ...`).
-11. Ensure the final decision is `approve` before moving to PR creation.
+11. Record `Pre-PR Review Round` from the returned `reviewRound`.
+12. Confirm `Pre-PR Reviewed Head` and `Pre-PR Reviewed Tree` match the `targetSha` and `targetTree` returned by `workflow-stage`.
+13. `workflow.agentReview.maxRounds` counts Feature finding-remediation passes, not the initial review. If `workflow-stage` returns `review_escalation`, stop automatic remediation and present its exact action options. Another remediation requires increasing the setting first.
+14. Ensure the final decision is `approve` before moving to PR creation.
 
-The review artifact must record the actual `executor`, `model`, `reasoningEffort`, reviewed commit/diff scope, findings, and final decision. The review subagent must not modify code; the main agent owns finding remediation and documentation updates.
+The review artifact must record the actual `executor`, `model`, `reasoningEffort`, reviewed commit/diff scope, target SHA/tree, findings, and final decision. The review subagent must not modify code; the main agent owns finding remediation and documentation updates.
 
 ---
 

package/templates/en/common/agents/skills/execute-task.md

@@ -13,11 +13,18 @@ Use the active feature folder as the execution SSOT.
   - continue the single `[DOING]` task, or
   - move the next highest-priority `[TODO]` task to `[DOING]`
 - Work one task at a time. Do not batch-complete multiple tasks in one pass.
+- When `nextAction.executor === "subagent"`, delegate that task's implementation and task-scoped verification to a fresh subagent in `nextAction.workingDirectory` using the returned `model`, `reasoningEffort`, `onUnavailable`, and exact `workerContract`. If a requested model is unavailable, `inherit` means retry with the current model inherited and `error` means stop and report the failure.
+- The implementation worker executes the assigned task directly. It must not run `workflow-stage`, delegate again, edit lee-spec-kit docs, change task state, commit, request approval, or perform remote/destructive actions. It may edit project code and run task-scoped checks only.
+- The implementation worker follows the approved `plan.md` Verification Contract: `NONE` adds no durable test, `UPDATE` minimally changes the owning existing test, and `ADD` adds only contract-linked tests. If the approved decision is insufficient, report the gap to the main agent instead of expanding test scope.
+- Legacy task lines without an explicit ID receive a stable synthetic `taskId` from `workflow-stage`; use that returned ID without rewriting the legacy task solely to add an ID.
 
 ## 2. Execute and record
 
 - Keep `tasks.md` aligned with reality:
   - do not mark `[DONE]` without real completion and verification
+  - when `workflow.agentReview.task.enabled=true`, move completed implementation to `[REVIEW]` instead of `[DONE]`, create the checkpoint commit, and run the independent review
+  - record the returned `reviewRound`, and move `[REVIEW]` to `[DONE]` only after the task reviewer approves the current SHA/tree
+  - `workflow.agentReview.maxRounds` counts task finding-remediation passes, not the initial review; at `review_escalation`, stop automatic remediation and present the exact action options, and increase the setting before another remediation
   - update `Acceptance` and `Checklist` in the same edit when closing a task
   - if a completed task needs follow-up, add a new task instead of rewriting history
 - If you need to add a new task, append a complete task block in `tasks.md` with a concrete title, `Acceptance`, `Checklist`, and `NON-PRD` or existing `PRD-*` tag.
@@ -33,6 +40,7 @@ Use the active feature folder as the execution SSOT.
 ## 4. Commit and stop guardrails
 
 - Before `git commit`, use `npx lee-spec-kit commit-audit --json` when docs-path validation matters.
+- The main agent, never the implementation subagent, owns docs/project commits and the task checkpoint.
 - Before stopping, use `npx lee-spec-kit workflow-audit --json` if code or feature docs changed.
 - Keep one row per test command in the `tasks.md` test log and update that row on reruns instead of appending duplicates.
 
@@ -40,7 +48,7 @@ Use the active feature folder as the execution SSOT.
 
 - Ask for approval only at documented review checkpoints and before remote or destructive actions.
 - Before issue creation, PR creation, push, merge, or similar remote actions, share the exact artifact or plan first.
-- Codex may delegate implementation work, but docs updates, approval handling, and remote actions stay with the main session.
+- Delegate Plan/task/Feature review to the fresh read-only subagent and exact hash/SHA/tree target returned by `workflow-stage`; keep docs updates, finding remediation, approval handling, and remote actions in the main session.
 
 ## Strict Rules
 

package/templates/en/common/features/README.md

@@ -57,9 +57,13 @@ npx lee-spec-kit workflow-stage <feature-ref> --json
 
 Use the returned `stage`, `nextAction`, and `implementationAllowed` values as the current workflow state.
 
+With Plan review enabled, planning follows `plan Review → fresh read-only Plan review → plan approval`. The review is bound to the returned `specHash` and `planHash`; changing either document's content invalidates the prior evidence. The reviewer checks the Verification Contract and test decisions without editing docs.
+
 The three final completion checkboxes in `tasks.md` carry `lee-spec-kit:completion:*` HTML markers. You may customize their visible wording, but preserve the marker on each checkbox line; `workflow-stage` uses the marker as the machine-readable identity and falls back to the legacy canonical wording for older projects.
 
-In a local workflow with `completionStrategy: "local-ff"` or `"local-squash"`, completion is `implementation_approve → feature_verify → local_merge → local_cleanup → done`. Failed checks enter `feature_remediation` with implementation enabled. `local-ff` moves only the verified Feature SHA; `local-squash` requires the integration tree to match the verified Feature tree. Both require cleanup before `done`. After cleanup, a Feature remains `done` while its recorded integration commit is still an ancestor of the current base, even when later Features advance that base.
+With Feature agent review enabled, local completion is `feature review → implementation_approve → feature_verify → local_merge → local_cleanup → done`. With task review enabled, every task follows `DOING → REVIEW → task review → DONE`. Failed checks enter `feature_remediation` with implementation enabled. `local-ff` moves only the verified Feature SHA; `local-squash` requires the integration tree to match the verified Feature tree. Both require cleanup before `done`. After cleanup, a Feature remains `done` while its recorded integration commit is still an ancestor of the current base, even when later Features advance that base.
+
+When `workflow.agentExecution.task.enabled=true`, each `task_execute` action carries the configured subagent model, reasoning effort, stable task ID, implementation working directory, and a machine-readable `workerContract`. The worker executes directly without calling `workflow-stage` or delegating again. It follows the approved Verification Contract, does not add unplanned durable tests, edits project code, and runs task-scoped checks only; the main agent owns docs synchronization, task transitions, commits, approvals, and remote actions. Official hooks block commits until the main agent advances the workflow to `task_commit`.
 
 A remediation commit invalidates verification and the prior local-merge confirmation. Refresh review evidence for the changed diff when Pre-PR review is enabled, verify the new tip, and obtain local-merge approval again.
 
@@ -124,20 +128,24 @@ Keeping the shared artifact for history is fine, but when it conflicts with feat
 | Scope | Field | Values |
 | --- | --- | --- |
 | Document status | `Status` in `spec.md`/`plan.md`, `Doc Status` in `tasks.md` | `Draft` \| `Review` \| `Approved` |
+| Plan review status | `Plan Review` in `plan.md` | `Pending` \| `Running` \| `Done` |
+| Plan review evidence/decision | `Plan Review Evidence` / `Plan Review Decision` | evidence path and `decision: approve\|changes_requested\|blocked ...` |
+| Plan review target | `Plan Reviewed Spec Hash` / `Plan Reviewed Plan Hash` | current content hashes returned by `workflow-stage` |
 | Issue doc status | `Status` in `issue.md` | `Draft` \| `Ready` |
 | PR doc status | `Status` in `pr.md` | `Draft` \| `Ready` |
 | PR review status | `PR Status` in `tasks.md` | `Review` \| `Approved` |
 | Pre-PR review status | `Pre-PR Review` in `tasks.md` | `Pending` \| `Done` |
 | Pre-PR review evidence | `Pre-PR Evidence` in `tasks.md` | evidence link/log/doc path |
 | Pre-PR review decision | `Pre-PR Decision` in `tasks.md` | `decision: approve\|changes_requested\|blocked ...` |
+| Pre-PR review target | `Pre-PR Reviewed Head` / `Pre-PR Reviewed Tree` | current SHA/tree returned by `workflow-stage` |
 | PR review evidence | `PR Review Evidence` in `tasks.md` | evidence link/log/doc path |
 | PR review decision | `PR Review Decision` in `tasks.md` | `decision: ...` (or `결정: ...`) |
 
 ---
 
-## Pre-PR Subagent Checklist
+## Agent Review Checklist
 
-Delegate every Pre-PR review to a fresh, read-only subagent using the model and reasoning effort returned by `workflow-stage --json`. The subagent follows `agents/skills/create-pr.md` (`Pre-PR Baseline Checklist`); the main agent owns finding remediation and evidence recording.
+Delegate Plan, task, and Feature reviews to fresh, read-only subagents using the model, reasoning effort, and exact target returned by `workflow-stage --json`. Plan review covers current spec/plan content hashes, task review covers its checkpoint range, and Feature review covers the base-to-Feature-tip diff. The subagent returns defect-focused findings without modifying code or docs; the main agent owns remediation and evidence recording. No named review skill is required.
 
 ---
 
@@ -146,7 +154,7 @@ Delegate every Pre-PR review to a fresh, read-only subagent using the model and
 | File           | Role                      | When to Write       |
 | -------------- | ------------------------- | ------------------- |
 | `spec.md`      | **What and Why**          | Feature definition  |
-| `plan.md`      | **How** (technical)       | After spec approval |
+| `plan.md`      | **How** + Verification Contract | After spec approval |
 | `tasks.md`     | Specific work items       | After plan approval |
 | `issue.md`     | Issue draft + issue state (`Draft/Ready`) | Before/when creating issue |
 | `pr.md`        | PR draft + PR state (`Draft/Ready`) | Before/when creating PR |

package/templates/en/common/features/feature-base/plan.md

@@ -12,6 +12,18 @@
 - **Created**: {YYYY-MM-DD}
 - **Status**: -
   - Values: Draft | Review | Approved
+- **Plan Review**: Pending
+  - Values: Pending | Running | Done
+- **Plan Review Evidence**: -
+  - Example: `docs/features/F001-foo/decisions.md` or another existing review artifact under the docs root
+- **Plan Review Decision**: -
+  - Format: `decision: approve|changes_requested|blocked ...`
+- **Plan Review Round**: -
+  - Positive integer returned by `workflow-stage --json`; the first review is `1`
+- **Plan Reviewed Spec Hash**: -
+  - Exact `specHash` returned by `workflow-stage --json`
+- **Plan Reviewed Plan Hash**: -
+  - Exact `planHash` returned by `workflow-stage --json`
 
 ---
 
@@ -37,11 +49,38 @@ src/
 
 ---
 
-## Test Strategy
+## Verification Contract
 
-- **Unit Tests**:
-- **Integration Tests**:
-- **E2E Tests**:
+### Change Classification
+
+- **Type**: COPY | REFACTOR | BUG_FIX | NEW_BEHAVIOR | HIGH_RISK
+- **Risk**: LOW | MEDIUM | HIGH
+
+### Observable Contract
+
+- **Supported behavior**:
+- **Preconditions**:
+- **Success guarantees**:
+- **Important failure guarantees**:
+- **Intentionally unsupported cases**:
+
+### Test Decisions
+
+| Contract / Requirement | Decision | Test Level | Realistic Regression Protected | Independent Oracle |
+| ---------------------- | -------- | ---------- | ------------------------------ | ------------------ |
+| (AC/FR reference)      | NONE \| UPDATE \| ADD | Unit \| Integration \| E2E \| Non-test | (failure this prevents) | (spec/released behavior/external reference) |
+
+### Excluded Tests
+
+- (duplicates, implementation details, unsupported synthetic inputs, or framework behavior intentionally not tested)
+
+### Verification Execution
+
+- **During implementation**:
+- **Before task completion**:
+- **Before Feature completion**:
+- **Manual/UI verification**:
+- **Full suite required**: Yes | No — (reason)
 
 ---
 

package/templates/en/common/features/feature-base/tasks.md

@@ -2,13 +2,16 @@
 
 ## Task Rules
 
-- **Status**: `[TODO]` → `[DOING]` → `[DONE]`
+- **Status**: normally `[TODO]` → `[DOING]` → `[DONE]`; when `workflow.agentReview.task.enabled=true`, use `[DOING]` → `[REVIEW]` → `[DONE]`
+- **Implementation delegation**: when `workflow.agentExecution.task.enabled=true`, the configured worker follows the returned `workerContract`, executes directly without running `workflow-stage` or delegating again, and may edit project code plus run task-scoped checks only; the main agent retains this document, task state, commits, approvals, and remote actions.
 - **Task communication / confirmation**:
   - `[TODO] → [DOING]`: share the task title first, then update the task state in `tasks.md`
-  - `[DOING] → [DONE]`: share the result and verification first, then update `Acceptance` and `Checklist` in the same edit
+  - `[DOING] → [REVIEW]/[DONE]`: share the result and verification first, then update `Acceptance` and `Checklist` in the same edit
+  - `[REVIEW] → [DONE]`: record the returned review round, fresh task review evidence, decision, reviewed head, and reviewed tree before completion
   - Ask for approval before changing task state only when the task crosses a documented review checkpoint or before remote/destructive actions.
   - Do not invent a standalone `OK` approval step when the workflow does not require one.
   - Do not mark `[DONE]` while any item in that task's `Checklist` remains unchecked.
+  - `workflow.agentReview.maxRounds` counts automatic finding-remediation passes, not the initial review. Stop at `review_escalation` after those passes are exhausted; another remediation requires increasing the setting first.
 - **PRD mapping (recommended)**: add an existing PRD requirement ID tag like `[PRD-FR-001]` or `[PRD-SCOPE-V1-DESKTOP-EDITOR]` to each task line, or tag non-PRD tasks as `[NON-PRD]`.
   - Do not invent PRD IDs in `tasks.md`. Only reference IDs that already exist in `docs/prd` or the upstream requirements doc.
   - If this is a legacy feature without PRD IDs yet, backfill IDs in the source requirements doc first, then align `spec.md` `PRD Refs` and task tags together.
@@ -39,7 +42,12 @@
 - **Pre-PR Decision**: -
   - Format: `decision: approve|changes_requested|blocked ...` (or `결정: ...`)
   - PR creation requires final decision `approve`
-  - Follow `agents/skills/create-pr.md` (`Pre-PR Baseline Checklist`) as the default baseline
+- **Pre-PR Review Round**: -
+  - Positive integer returned by `workflow-stage --json`; the first review is `1`
+- **Pre-PR Reviewed Head**: -
+  - Project code commit SHA reviewed by the Feature reviewer
+- **Pre-PR Reviewed Tree**: -
+  - Project code tree SHA reviewed by the Feature reviewer
 - **PR Review**: -
   - Values: Pending | Running | Done
   - Mark `Running` when PR review handoff starts; use `Done` only if your team explicitly tracks review completion here
@@ -59,6 +67,11 @@
     - (verification condition)
   - Checklist:
     - [ ] (subtask)
+  - Review Evidence: -
+  - Review Decision: -
+  - Review Round: -
+  - Reviewed Head: -
+  - Reviewed Tree: -
 ```
 
 > `PRD-FR-001` in the example is just one valid `PRD-*` key. If the key is not defined in the PRD source yet, do not add it to tasks first.

package/templates/ko/common/README.md

@@ -117,6 +117,9 @@ npx lee-spec-kit docs get agents --json
 ## CLI 설정 파일 (`.lee-spec-kit.json`)
 
 `lee-spec-kit init`을 실행하면 문서 루트(기본: `docs/`)에 `.lee-spec-kit.json`이 생성됩니다.
+대화형 init에서는 권장 설정을 그대로 쓰거나 Task 구현 위임, Plan/Task/Feature 검수,
+Local 통합 방식을 직접 선택할 수 있습니다. 비대화형 실행에서는 `--task-agent`,
+`--reviews`, `--max-review-rounds`, `--completion-strategy`로 같은 값을 지정합니다.
 
 - `lee-spec-kit feature`, `config`, `update`, `detect`, workflow validator에서 문서 위치/프로젝트 타입/언어를 해석하는 용도로 사용됩니다.
 - `docsRepo`, `pushDocs`, `docsRemote`는 CLI 관리 **Docs Push 정책**을 위한 메타데이터입니다. (자동 push는 하지 않습니다)
@@ -130,11 +133,24 @@ npx lee-spec-kit docs get agents --json
 - `docsRepo` ("embedded" | "standalone"): Docs 관리 방식
 - `pushDocs` (boolean, optional): `docsRepo: "standalone"`일 때만 생성 (원격 push 여부)
 - `docsRemote` (string, optional): `pushDocs: true`일 때만 생성 (원격 레포 URL)
-- `workflow.prePrReview.reviewer` (object): Pre-PR 서브에이전트 실행 설정
+- `workflow.agentExecution.task` (object): 태스크 구현 위임 설정
+  - `enabled`: 각 `task_execute`를 서브에이전트에게 위임할지 여부. 새 프로젝트와 업데이트된 프로젝트의 기본값은 `true`
   - `type`: 현재 `"subagent"`만 지원
   - `model`: `"inherit"` 또는 런타임이 지원하는 모델명
   - `reasoningEffort`: `low | medium | high | xhigh | max | ultra`
   - `onUnavailable`: 지정 모델을 사용할 수 없을 때 `inherit | error`
+  - `workflow-stage`는 안정적인 태스크 ID, 작업/docs 경로, machine-readable `workerContract`를 반환합니다. worker는 `workflow-stage` 재호출이나 재위임 없이 직접 실행합니다.
+  - 구현 서브에이전트는 프로젝트 코드와 태스크 범위 검사를 담당하고, 메인 에이전트는 문서, 태스크 상태, 커밋, 승인, 원격 작업을 유지합니다. 공식 hook은 `task_execute` 중 커밋을 거부합니다.
+- `workflow.agentReview.maxRounds` (양의 정수): Plan/태스크/Feature 리뷰에 공통 적용되는 리뷰 지적 자동 반영 최대 횟수. 최초 리뷰는 세지 않으며 기본값은 `1`
+- `workflow.agentReview.plan` / `workflow.agentReview.task` / `workflow.agentReview.feature` (object): Plan/태스크/Feature 독립 리뷰 설정
+  - `enabled`: 해당 리뷰 게이트 활성화 여부. 새 프로젝트는 Plan과 Feature `true`, task `false`
+  - `evidenceMode`: `path_required | any`
+  - `reviewer`: fresh 읽기 전용 서브에이전트 실행 설정
+  - `type`: 현재 `"subagent"`만 지원
+  - `model`: `"inherit"` 또는 런타임이 지원하는 모델명
+  - `reasoningEffort`: `low | medium | high | xhigh | max | ultra`
+  - `onUnavailable`: 지정 모델을 사용할 수 없을 때 `inherit | error`
+  - Plan 검수는 Plan 승인 전에 실행되며 `workflow-stage`가 반환한 정확한 `specHash`와 `planHash`에 evidence를 묶습니다. 이후 spec/plan 내용이 바뀌면 fresh 검수가 필요합니다.
 - `workflow.baseBranch` (string): 완료된 local Feature를 통합할 기준 브랜치
 - `workflow.completionStrategy` (`"local-ff" | "local-squash" | "none"`): fast-forward, 검증된 단일 squash commit 생성, 또는 명시적으로 통합 없이 종료
 - `workflow.deleteFeatureBranchAfterMerge` (boolean): cleanup 후 통합된 local Feature 브랜치 삭제 여부. 원격 브랜치는 삭제하지 않음
@@ -171,14 +187,47 @@ npx lee-spec-kit docs get agents --json
     "deleteFeatureBranchAfterMerge": true,
     "featureChecks": [],
     "postMergeChecks": [],
-    "prePrReview": {
-      "evidenceMode": "path_required",
-      "reviewer": {
+    "agentExecution": {
+      "task": {
+        "enabled": true,
         "type": "subagent",
         "model": "inherit",
         "reasoningEffort": "high",
         "onUnavailable": "inherit"
       }
+    },
+    "agentReview": {
+      "maxRounds": 1,
+      "plan": {
+        "enabled": true,
+        "evidenceMode": "path_required",
+        "reviewer": {
+          "type": "subagent",
+          "model": "inherit",
+          "reasoningEffort": "high",
+          "onUnavailable": "inherit"
+        }
+      },
+      "task": {
+        "enabled": false,
+        "evidenceMode": "path_required",
+        "reviewer": {
+          "type": "subagent",
+          "model": "inherit",
+          "reasoningEffort": "high",
+          "onUnavailable": "inherit"
+        }
+      },
+      "feature": {
+        "enabled": true,
+        "evidenceMode": "path_required",
+        "reviewer": {
+          "type": "subagent",
+          "model": "inherit",
+          "reasoningEffort": "high",
+          "onUnavailable": "inherit"
+        }
+      }
     }
   },
   "allowedDocsEntries": {
@@ -192,7 +241,7 @@ npx lee-spec-kit docs get agents --json
 }
 ```
 
-새 local 프로젝트는 `local-ff`를 사용합니다. base branch에 하나의 commit만 남기려면 `local-squash`를 선택하세요. 이때 task checkpoint 증거를 위해 원본 Feature tip을 내부 `refs/lee-spec-kit/integrations/*` ref로 보존합니다. 기존 local 프로젝트에 명시적 `completionStrategy`가 없으면 `update`가 `none`을 넣어 업그레이드 도중 현재 브랜치를 갑자기 병합하지 않습니다. 준비가 끝난 뒤 `local-ff` 또는 `local-squash`로 명시적으로 전환하세요.
+새 프로젝트와 업데이트된 프로젝트는 기본적으로 태스크 구현을 위임하고 Plan 검수를 활성화합니다. 직접 구현하거나 Plan 검수를 끄려면 각각 `workflow.agentExecution.task.enabled` 또는 `workflow.agentReview.plan.enabled`를 `false`로 설정하세요. 새 local 프로젝트는 `local-ff`와 Feature agent review를 사용합니다. 태스크마다 리뷰하려면 `workflow.agentReview.task.enabled`를 `true`로 켜세요. 리뷰 지적 자동 반영 기본값은 `1`입니다. 즉 첫 리뷰 지적은 한 번 반영하고 다시 리뷰하며, 또 `changes_requested`가 나오면 `review_escalation`으로 전환합니다. 추가 반영과 fresh 리뷰를 허용하려면 먼저 `workflow.agentReview.maxRounds`를 늘리세요. base branch에 하나의 commit만 남기려면 `local-squash`를 선택하세요. 이때 task checkpoint 증거를 위해 원본 Feature tip을 내부 `refs/lee-spec-kit/integrations/*` ref로 보존합니다. 기존 local 프로젝트에 명시적 `completionStrategy`가 없으면 `update`가 `none`을 넣어 업그레이드 도중 현재 브랜치를 갑자기 병합하지 않습니다. 준비가 끝난 뒤 `local-ff` 또는 `local-squash`로 명시적으로 전환하세요.
 
 ```json
 {

package/templates/ko/common/agents/agents.md

@@ -52,9 +52,14 @@
 
 - lee-spec-kit은 문서 구조, workflow 단계, validator를 담당합니다.
 - Codex는 실행 루프, 도구 사용, hook lifecycle을 담당합니다.
-- `workflow-stage --json`가 `stage === "implementation"`이고 `implementationAllowed === true`를 반환하기 전에는 구현을 시작하지 않습니다.
-- `workflow-stage --json`의 `nextAction.category`가 `pre_pr_review`이고 `executor`가 `subagent`이면, 반환된 `model`, `reasoningEffort`, `onUnavailable` 정책으로 fresh context의 읽기 전용 서브에이전트 리뷰를 실행합니다. 리뷰 스킬 이름을 선택하거나 요구하지 않습니다.
-- Pre-PR 리뷰 서브에이전트는 finding만 반환하며 코드를 수정하지 않습니다. 메인 에이전트가 finding을 반영하고 reviewer metadata와 최종 decision을 evidence에 기록합니다.
+- `implementationAllowed === true`일 때만 구현 코드를 수정합니다. 일반 태스크 구현은 `stage === "implementation"`에서, 리뷰 수정은 `task_review_fix` 또는 `feature_review_fix`에서, 검증 수정은 `feature_remediation`에서만 수행합니다.
+- `nextAction.category`가 `plan_review`이고 `executor`가 `subagent`이면 반환된 `specHash`와 `planHash`를 대상으로 fresh 읽기 전용 서브에이전트가 `spec.md`와 `plan.md`를 검수합니다. 메인 에이전트가 반환된 `reviewRound`, Plan 검수 evidence, decision, reviewer metadata와 두 hash를 기록하며 이후 spec/plan 내용 변경은 기존 검수를 무효화합니다.
+- `nextAction.category`가 `task_execute`이고 `executor`가 `subagent`이면 해당 태스크 하나를 활성화한 뒤, 반환된 `workingDirectory`에서 fresh 서브에이전트에게 반환된 모델·추론도·unavailability 정책과 정확한 `workerContract`로 구현과 태스크 범위 검증을 위임합니다. 특정 이름의 실행 스킬은 요구하지 않습니다.
+- 구현 worker는 승인된 Verification Contract를 따르고 계획되지 않은 영구 테스트를 추가하지 않으며 직접 실행합니다. `workflow-stage`를 호출하거나 다른 서브에이전트를 생성하지 않습니다. 프로젝트 코드 수정과 범위 내 검사는 수행할 수 있지만 lee-spec-kit 문서 수정, 태스크 상태 변경, 커밋, 승인 요청, 원격/파괴적 작업은 하지 않습니다. 메인 에이전트가 결과를 확인하고 문서 동기화, 태스크 전환, 커밋, 후속 workflow를 소유하며, 공식 hook은 `task_execute`가 활성화된 동안 커밋을 차단합니다.
+- `nextAction.category`가 `task_review`이고 `executor`가 `subagent`이면 반환된 task ID와 SHA/tree 범위를 fresh context의 읽기 전용 서브에이전트가 리뷰하고 반환된 `reviewRound`를 기록합니다.
+- `nextAction.category`가 `pre_pr_review`이고 `executor`가 `subagent`이면 반환된 모델·추론도·`reviewRound`·SHA/tree 범위로 fresh context의 읽기 전용 Feature 리뷰를 실행합니다. 리뷰 스킬 이름을 선택하거나 요구하지 않습니다.
+- 리뷰 서브에이전트는 finding만 반환하고 코드를 수정하지 않습니다. 메인 에이전트가 finding을 반영하고 reviewer metadata, reviewed scope, evidence, decision, 정확한 hash/SHA/tree target metadata를 기록합니다.
+- `workflow.agentReview.maxRounds`는 최초 리뷰 횟수가 아니라 리뷰 지적 자동 반영 횟수입니다. `nextAction.category`가 `review_escalation`이면 그 횟수를 모두 사용한 상태입니다. 자동 수정이나 추가 리뷰 위임을 하지 말고 구현을 차단한 채 반환된 선택지를 그대로 제시합니다. 추가 반영은 먼저 `maxRounds`를 늘려야 합니다.
 - spec / plan / tasks 승인, issue 생성, branch 생성은 구현 전 하드 게이트로 취급합니다.
 - standalone 모드에서는 `git worktree add`를 직접 만들지 말고 `workflow-stage`의 정확한 `nextAction.command`를 실행해 managed workspace 경로, stale 디렉터리 정리, `.env`/`.env.*` 복사 단계가 일관되게 유지되도록 합니다.
 - local 모드에서는 구현 승인 직후 종료하지 않습니다. `workflow-stage`가 반환하는 정확한 `local verify`, `local merge`, `local cleanup` 명령을 따라 검증·통합·정리가 확인되어 `done`이 될 때까지 진행합니다. `feature_remediation` 단계에서는 Feature worktree 수정이 명시적으로 허용됩니다.

package/templates/ko/common/agents/skills/create-feature.md

@@ -20,10 +20,12 @@
 - 문서가 SSOT입니다. 활성 feature 문서를 직접 따라갑니다.
 - 문서 단계는 직접 따라갑니다:
   - `spec.md`는 범위와 리뷰 상태를 정의합니다
-  - `plan.md`는 구현 접근을 정의합니다
+  - `plan.md`는 구현 접근과 Verification Contract를 정의합니다
   - `tasks.md`는 실제 실행 순서를 정의합니다
   - `issue.md`, `pr.md`는 GitHub 단계에 들어가면 stage gate의 일부로 사용합니다
 - `tasks.md`가 있다고 바로 구현하지 않습니다. 구현은 `workflow-stage --json`가 허용할 때만 시작합니다.
+- Plan 검수가 활성화되어 있으면 `plan.md`를 Review로 바꾸고 반환된 fresh 읽기 전용 `plan_review`를 위임한 뒤 `reviewRound`, evidence, decision, reviewer metadata, `specHash`, `planHash`를 기록합니다. 현재 hash에 대한 approve 검수만 Plan 승인으로 이어질 수 있습니다. reviewer는 문서를 수정하지 않고 NONE/UPDATE/ADD 결정, 요구사항 커버리지, 독립적인 Oracle, 안정적인 관찰 경계, 현실적인 실패/롤백, 제외 범위, focused/full 검증 범위를 점검합니다.
+- `workflow.agentReview.maxRounds`는 최초 Plan 리뷰가 아니라 지적 자동 반영 횟수입니다. `workflow-stage`가 `review_escalation`을 반환하면 Plan 자동 수정을 멈추고 정확한 선택지를 사용자에게 제시합니다. 추가 반영은 먼저 설정값을 늘려야 합니다.
 - 범위나 동작이 바뀌면 같은 턴 안에서 활성 feature 문서를 같이 업데이트합니다.
 - 사용자 승인은 문서화된 review checkpoint와 원격/파괴적 작업 전에만 요청합니다.
 - docs 경로 검사가 중요하면 `git commit` 전에 `npx lee-spec-kit commit-audit --json`를 사용합니다.

package/templates/ko/common/agents/skills/create-pr.md

@@ -25,15 +25,18 @@ Pre-PR 리뷰에서 서브에이전트가 항상 수행하는 최소 기준입
 2. 회귀/예외 처리, 크리티컬·보안 리스크, 사이드 이펙트, 사용자 흐름 영향, 배포 준비도를 점검합니다.
 3. 유지보수성을 점검합니다: 큰 함수/파일은 필요 시 분리하고, 기존 코드 재사용·통합 가능성을 확인하며, 불필요해진 코드를 정리합니다.
 4. 현재 구현이 `spec.md` / `plan.md` / `tasks.md`에 기록된 feature 의도와 범위에 실제로 맞는지 평가합니다.
-5. `workflow.prePrReview.evidenceMode=path_required`(기본)일 때는 승인 전에 `review-trace.json` 같은 실제 리뷰 산출물을 남깁니다. `evidenceMode=any`에서는 실행 증거 강제가 없는 한 별도 산출물 없이 직접 기록하는 경로도 허용됩니다.
+5. `workflow.agentReview.feature.evidenceMode=path_required`(기본)일 때는 승인 전에 `review-trace.json` 같은 실제 리뷰 산출물을 남깁니다. `evidenceMode=any`에서는 실행 증거 강제가 없는 한 별도 산출물 없이 직접 기록하는 경로도 허용됩니다.
 6. `PR 전 리뷰 Evidence`는 설정된 evidence 정책을 따라야 합니다. `path_required`일 때는 실제 존재하는 문서 경로를 사용합니다.
 7. `Summary`, `Feature Intent Summary`, `Implementation Fit`, `Missing Cases`, `Spec Alignment Checked`, `Finding Count`, `Blocking Findings`, `Findings`, `Residual Risks`를 placeholder 없이 기록합니다.
 8. 리뷰 중에 audit/타깃 검증 명령을 실제로 실행했다면 그때만 `commandsExecuted`에 기록합니다.
 9. 코드리뷰 단계에서도 `PR 리뷰 Evidence/Decision`과 `decisions.md`를 동기화하고 `PR Review Log`(또는 `PR 리뷰 로그`)의 `Summary`/`Decision`을 기록합니다.
 10. `PR 전 리뷰 Decision`은 `결정: approve|changes_requested|blocked ...` (또는 `decision: ...`) 형식을 사용합니다.
-11. PR 생성 단계로 이동하기 전 최종 Decision이 `approve`인지 확인합니다.
+11. 반환된 `reviewRound`를 `PR 전 리뷰 Round`에 기록합니다.
+12. `PR 전 리뷰 Head`와 `PR 전 리뷰 Tree`가 `workflow-stage`의 `targetSha`와 `targetTree`와 일치하는지 확인합니다.
+13. `workflow.agentReview.maxRounds`는 최초 Feature 리뷰가 아니라 지적 자동 반영 횟수입니다. `workflow-stage`가 `review_escalation`을 반환하면 자동 수정을 멈추고 정확한 선택지를 사용자에게 제시합니다. 추가 반영은 먼저 설정값을 늘려야 합니다.
+14. PR 생성 단계로 이동하기 전 최종 Decision이 `approve`인지 확인합니다.
 
-리뷰 산출물에는 실제 사용한 `executor`, `model`, `reasoningEffort`, 검토한 commit/diff 범위, finding과 최종 decision을 기록합니다. 리뷰 서브에이전트는 코드를 수정하지 않으며, finding 반영과 문서 갱신은 메인 에이전트가 담당합니다.
+리뷰 산출물에는 실제 사용한 `executor`, `model`, `reasoningEffort`, 검토한 commit/diff 범위, target SHA/tree, finding과 최종 decision을 기록합니다. 리뷰 서브에이전트는 코드를 수정하지 않으며, finding 반영과 문서 갱신은 메인 에이전트가 담당합니다.
 
 ---
 

package/templates/ko/common/agents/skills/execute-task.md

@@ -13,11 +13,18 @@
   - 이미 `[DOING]`인 태스크가 하나 있으면 그것을 이어서 수행하고
   - 없으면 가장 우선순위가 높은 `[TODO]` 태스크를 `[DOING]`으로 바꿉니다
 - 한 번에 하나의 태스크만 진행합니다.
+- `nextAction.executor === "subagent"`이면 반환된 `workingDirectory`에서 해당 태스크의 구현과 태스크 범위 검증을 fresh 서브에이전트에게 반환된 `model`, `reasoningEffort`, `onUnavailable` 설정과 정확한 `workerContract`로 위임합니다. 지정 모델을 사용할 수 없을 때 `inherit`은 현재 모델을 상속해 다시 위임하고, `error`는 중단 후 실패를 보고한다는 뜻입니다.
+- 구현 worker는 할당된 태스크를 직접 실행합니다. `workflow-stage` 재호출, 재위임, lee-spec-kit 문서 수정, 태스크 상태 변경, 커밋, 승인 요청, 원격/파괴적 작업은 하지 않습니다. 프로젝트 코드 수정과 태스크 범위 검사만 수행합니다.
+- 구현 worker는 승인된 `plan.md` Verification Contract를 따릅니다. `NONE`이면 영구 테스트를 추가하지 않고, `UPDATE`이면 계약을 소유한 기존 테스트만 최소 수정하며, `ADD`이면 계약에 연결된 테스트만 추가합니다. 승인된 결정이 충분하지 않다면 테스트 범위를 임의로 넓히지 말고 메인 에이전트에 보고합니다.
+- 명시적 ID가 없는 레거시 태스크는 `workflow-stage`가 안정적인 synthetic `taskId`를 반환합니다. ID 추가만을 위해 레거시 태스크 문서를 다시 쓰지 말고 반환된 ID를 사용합니다.
 
 ## 2. 실행과 기록
 
 - `tasks.md`를 현실과 맞게 유지합니다:
   - 실제 완료/검증 없이 `[DONE]`로 바꾸지 않습니다
+  - `workflow.agentReview.task.enabled=true`이면 구현/검증 완료 후 `[DONE]` 대신 `[REVIEW]`로 바꾸고 checkpoint commit을 만든 뒤 독립 리뷰를 진행합니다
+  - 반환된 `reviewRound`를 기록하고, task review가 현재 SHA/tree를 `approve`한 뒤에만 `[REVIEW]`를 `[DONE]`으로 바꿉니다
+  - `workflow.agentReview.maxRounds`는 최초 task 리뷰가 아니라 지적 자동 반영 횟수입니다. `review_escalation`에서는 자동 수정을 멈추고 정확한 선택지를 사용자에게 제시하며, 추가 반영 전에 설정값을 늘려야 합니다
   - 태스크를 닫을 때는 같은 수정에서 `Acceptance`와 `Checklist`도 함께 갱신합니다
   - 완료된 태스크에 후속 작업이 생기면 히스토리를 고치지 말고 새 태스크를 추가합니다
 - 새 태스크를 추가해야 한다면 `tasks.md`에 구체적인 제목, `Acceptance`, `Checklist`, 그리고 `NON-PRD` 또는 기존 `PRD-*` 태그가 있는 완전한 태스크 블록을 추가하세요.
@@ -33,6 +40,7 @@
 ## 4. 커밋과 종료 가드레일
 
 - docs 경로 검사가 중요하면 `git commit` 전에 `npx lee-spec-kit commit-audit --json`를 사용합니다.
+- 문서/프로젝트 커밋과 태스크 checkpoint는 구현 서브에이전트가 아니라 메인 에이전트가 소유합니다.
 - 코드나 feature 문서를 바꿨다면 종료 전에 `npx lee-spec-kit workflow-audit --json`로 동기화 상태를 확인합니다.
 - `tasks.md` 테스트 로그는 명령어당 1개 행만 유지하고, 재실행 시 기존 행을 갱신합니다.
 
@@ -40,7 +48,7 @@
 
 - 사용자 승인은 문서화된 review checkpoint와 원격/파괴적 작업 전에만 요청합니다.
 - issue 생성, PR 생성, push, merge 같은 원격 작업 전에는 올릴 artifact나 계획을 먼저 공유합니다.
-- 구현 자체는 Codex가 필요하면 위임할 수 있지만, 문서 갱신, 승인 처리, 원격 작업은 메인 세션에서 유지합니다.
+- Plan/task/Feature review는 `workflow-stage`가 반환한 모델·추론도와 정확한 hash/SHA/tree target 설정의 fresh 읽기 전용 서브에이전트에게 맡기고, 문서 갱신, finding 반영, 승인 처리, 원격 작업은 메인 세션에서 유지합니다.
 
 ## 절대 규칙