uctm 1.5.4 → 2.0.0

package/README.md

@@ -70,14 +70,14 @@ The agents analyze your request, plan the work, and execute through isolated sub
 [game-dev] Build a brick breaker game in HTML
 ```
 
-* **Requirement Analysis**: The agent analyzes your requirement and asks for approval. Review `works/WORK-NN/Requirement.md` and type **"approve"** to proceed.
+* **Requirement Analysis**: Main Claude spawns a single `orchestrator` agent (`mode=gated` by default), which nests a `specifier` call to analyze your requirement. Orchestrator pauses at **[GATE-1]** and returns a summary for approval. Review `works/WORK-NN/Requirement.md` and type **"approve"** to proceed.
 ```
 {Requirement specification content}
-If you approve Requirement.md, I will call the Planner to create PLAN.md + TASK decomposition.
+If you approve Requirement.md, orchestrator will nest-spawn the Planner to create PLAN.md + TASK decomposition.
 Let me know if you want to modify anything.
 ```
 
-* **WORK Execution Plan**: The agent builds an execution plan and asks again. Review `works/WORK-NN/PLAN.md` and `TASK-NN.md`, then type **"approve"** to proceed.
+* **WORK Execution Plan**: Orchestrator nests a `planner` call to build an execution plan, then pauses again at **[GATE-2]**. Review `works/WORK-NN/PLAN.md` and `TASK-NN.md`, then type **"approve"** to proceed.
 ```
 WORK-31 Development Approval Request
 
@@ -86,8 +86,6 @@ WORK-31 Development Approval Request
   ┌─────────┬─────────────────────────┐
   │  Item   │        Details          │
   ├─────────┼─────────────────────────┤
-  │ Mode    │ full                    │
-  ├─────────┼─────────────────────────┤
   │ TASKs   │ 6 (TASK-00 ~ TASK-05)  │
   └─────────┴─────────────────────────┘
 
@@ -99,11 +97,11 @@ WORK-31 Development Approval Request
      └─────────────────────────────────→ TASK-05 ($$$$$$$$$)
 
   - TASK-01/02 parallel, TASK-03/04 parallel, TASK-05 final integration
-  - If approved, scheduler → builder → verifier → committer pipeline will execute.
+  - If approved, orchestrator schedules the TASK DAG itself and nest-spawns builder → verifier → committer for each TASK — no further approval gates in this phase.
 
   Proceed?
 ```
-* Per-TASK: build → verify → commit repeats automatically for each TASK.
+* Per-TASK: build → verify → commit repeats automatically for each TASK, all inside the same nested orchestrator run.
 ```
 ● TASK-05 committed. Updating PROGRESS.md and finalizing WORK-31.
 ```
@@ -118,7 +116,7 @@ WORK-31 Development Approval Request
 ```
 [WORK start] Change the submit button label to "Send" — auto
 ```
-Add "auto" to skip all approval steps and run the entire process automatically.
+Add "auto" to run orchestrator in `mode=auto` — it completes the entire WORK in one nested spawn with **zero approval gates**, and records every judgment call it had to make on your behalf in `works/WORK-NN/DECISIONS.md` and the final report's `## 자동 결정 사항` section.
 
 ### 2. Token Economy
 
@@ -127,11 +125,10 @@ I'm cost-conscious (honestly). So this agent applies four token-saving strategie
 **(1) Serena MCP for codebase analysis.**
 The agent prioritizes [Serena MCP](https://github.com/oraios/serena) for code exploration — reading symbols instead of entire files. (Huge thanks to the Serena team.)
 
-**(2) Three execution modes + spawn consolidation to minimize subagent overhead.** The WORK-PIPELINE has 6 agent stages running sequentially. For a single-TASK WORK, that's 6 subagent sessions — each consuming tokens just to boot up. Wasteful. So the specifier agent decides the execution mode based on complexity: **direct** mode uses only 3 agent calls (specifier → builder → verifier+committer), skipping planner and scheduler. On top of this, related agents are combined into a single spawn: specifier+planner run together, and verifier+committer run together — reducing total spawns by **~30%** (6 TASKs: 20 → 14 spawns). See [Three Execution Modes](#concept-three-execution-modes).
+**(2) A single nested orchestrator instead of per-stage round-trips.** The WORK-PIPELINE has up to 6 agent stages. Instead of Main Claude calling each stage one at a time, Main Claude spawns a single `orchestrator` agent **once**, which nests specifier → planner → builder → verifier → committer as sub-spawns of its own (Claude Code sub-agent nesting, depth 2). The orchestrator schedules the TASK DAG itself instead of round-tripping through Main Claude for every stage. See [Concept: Orchestrator Modes](#concept-orchestrator-modes-gated-vs-auto).
 
-**(3) Structured XML communication.** Subagents can't nest — Main Claude orchestrates everything.
-* When one agent finishes and the next agent starts, Main Claude sits in between, causing data to be transmitted twice. This communication is a blob of text —
-* The receiving side has to parse it again. So we standardized the communication format as XML.
+**(3) Structured XML communication.** Even with nesting, every hop between agents is still a text boundary — gate summaries and hand-offs are still just blobs of text.
+* Whichever side receives it has to parse it again. So we standardized the communication format as XML.
 * Every bit helps.
 * (This also made agent log monitoring much easier.) See [Structured Agent Communication](#structured-agent-communication).
 
@@ -164,23 +161,15 @@ Six subagents work across any project and any language, automatically handling *
 
 ## Usage
 
-### Trivial Fix (direct mode)
+### Small Fix
 
 ```
 > [bugfix] Fix typo in login error message
 ```
 
-Main Claude calls specifier, which selects `execution-mode: direct` and returns a dispatch XML. Main Claude then calls builder (implements the change) and committer (commits). Creates WORK-NN directory + PLAN + result.md + commit.
-
-### Quick Task (pipeline mode)
-
-```
-> [bugfix] Fix the login button not responding on mobile
-```
-
-Main Claude calls specifier, which selects `execution-mode: pipeline` and creates PLAN. Then Main Claude calls builder → verifier → committer in sequence.
+Main Claude spawns `orchestrator` once (`mode=gated` by default). Orchestrator nests specifier → **[GATE-1]** → planner → **[GATE-2]** → builder → verifier → committer. Creates WORK-NN directory + PLAN + result.md + commit — all inside the same orchestrator run. Every WORK takes this path, so even a one-line change is planned and recorded.
 
-### Complex Feature (WORK)
+### Feature (WORK)
 
 #### 1. Create WORK (Planning)
 
@@ -188,7 +177,7 @@ Main Claude calls specifier, which selects `execution-mode: pipeline` and create
 > [new-feature] Build a user authentication feature. Plan it.
 ```
 
-The planner analyzes the project and creates WORK-01:
+Orchestrator nests specifier → planner, which analyzes the project and creates WORK-01:
 
 ```
 WORK-01: User Authentication
@@ -208,7 +197,7 @@ WORK-01: User Authentication
 > Run WORK-01 pipeline
 ```
 
-The scheduler executes WORK-01's TASKs in dependency order.
+Orchestrator's internal DAG scheduling (STEP C) executes WORK-01's TASKs in dependency order, nesting builder → verifier → committer for each TASK — this phase has no approval gates.
 
 #### 3. Add to Existing WORK
 
@@ -235,6 +224,8 @@ WORK Status
 > Resume WORK-02
 ```
 
+"automatically" triggers `mode=auto` — orchestrator completes in a single nested spawn without gates. "Resume" reattaches to the parked orchestrator via `SendMessage` (context preserved), or falls back to a fresh nested re-spawn reconstructed from `work_{WORK}.log` if the handle was lost.
+
 #### 6. Run a Specific TASK
 
 Skip to a specific TASK within a WORK (e.g., retry after a failure):
@@ -243,7 +234,7 @@ Skip to a specific TASK within a WORK (e.g., retry after a failure):
 > Run WORK-02: TASK-02
 ```
 
-The scheduler returns the next TASK, then Main Claude calls builder → verifier → committer in sequence.
+Orchestrator's DAG scheduling (STEP C) identifies the next READY TASK, then nests builder → verifier → committer in sequence.
 
 #### 7. Force WORK Creation (Skip Complexity Check)
 
@@ -255,7 +246,7 @@ Use the `[new-work]` tag to always create a new WORK regardless of complexity:
 
 #### 8. Handle Failure / Retry
 
-If a TASK fails during the pipeline, the scheduler retries up to 3 times automatically.
+If a TASK fails during the pipeline, orchestrator re-dispatches builder up to 3 times automatically (STEP C). If all 3 attempts fail, it escalates via `<needs-decision>` — a `<gate type="decision">` in `mode=gated`, or an auto-decision recorded in `DECISIONS.md` in `mode=auto`.
 If it still fails, you can inspect the result file and retry manually:
 
 ```
@@ -284,7 +275,7 @@ If WORK-02 is `IN_PROGRESS`, the specifier asks:
 > What's the status of WORK-03: TASK-02?
 ```
 
-The scheduler reads `PROGRESS.md` and `result.md` files to report current state.
+Orchestrator reads `PROGRESS.md` and `result.md` files to report current state.
 
 ---
 
@@ -337,10 +328,11 @@ uctm update --lang en
 
 ```bash
 git clone https://github.com/UCJung/uc-taskmanager-claude-agent.git /tmp/uc-tm
-mkdir -p .claude/agents
-cp /tmp/uc-tm/npm/agents/*.md .claude/agents/
+mkdir -p .claude/agents .claude/references
+cp /tmp/uc-tm/npm/agents/*.md .claude/agents/          # 6 agents (orchestrator, specifier, planner, builder, verifier, committer)
+cp /tmp/uc-tm/npm/references/*.md .claude/references/   # 7 reference files
 rm -rf /tmp/uc-tm
-git add .claude/agents/ && git commit -m "chore: add uc-taskmanager agents"
+git add .claude/agents/ .claude/references/ && git commit -m "chore: add uc-taskmanager agents"
 ```
 
 ### Local Plugin Test
@@ -355,50 +347,48 @@ claude --plugin-dir ./
 ```bash
 claude
 > /agents
-# specifier, planner, scheduler, builder, verifier, committer → confirm all 6
+# orchestrator, specifier, planner, builder, verifier, committer → confirm all 6
 ```
 
 ---
 
-## Concept: Three Execution Modes
+## Concept: Orchestrator Modes (Gated vs Auto)
 
-Main Claude detects the `[]` tag and calls the **specifier** subagent, which selects one of three `execution-mode` values:
+Main Claude detects the `[]` tag and spawns a single **orchestrator** subagent, passing `mode=gated` (default) or `mode=auto` (when the request contains "auto"/"자동으로"). Orchestrator nests every other agent itself — Main Claude never calls specifier/planner/builder/verifier/committer directly:
 
 ```
-User Request → Main Claude (orchestrator)
-                    │
-                    ▼
-              ┌───────────┐
-              │ specifier │ (called by Main Claude)
-              └─────┬─────┘
+User Request → Main Claude
                     │
-              execution-mode returned
-                    │
-      ├─ direct  (no build/test required)
-      │   → specifier returns dispatch XML → Main Claude calls builder → [verifier+committer]
-      │
-      ├─ pipeline  (build/test required, single domain, sequential)
-      │   → Main Claude calls: [specifier+planner] → builder → [verifier+committer]
-      │
-      └─ full  (multi-domain / complex DAG / new module / 5+ tasks)
-          → Main Claude calls: [specifier+planner] → scheduler → [builder → verifier+committer] × N
+                    ▼ spawn once, mode=gated|auto
+              ┌──────────────┐
+              │ orchestrator │  ← TASK DAG scheduling + gate/decision mediation
+              └──────┬───────┘
+                     │ nested spawn (depth 2)
+                     ▼
+              specifier → Requirement.md          ← [GATE-1]
+                     │
+              planner   → PLAN.md + TASK DAG      ← [GATE-2]
+                     │
+              STEP C: DAG-ordered [builder → verifier → committer] × N (parallel where READY)
 ```
 
-**Sub-agent Spawn Count by Mode:**
+- `mode=gated`: pauses with `<gate type="stage">` after specifier (**[GATE-1]**) and after planner (**[GATE-2]**); also pauses anytime with `<gate type="decision">` (background + options + recommendation) when orchestrator or a nested child needs a user call. Main Claude relays the gate, waits for approval/choice, then resumes the parked orchestrator with `SendMessage` (falls back to a log-based re-spawn if the handle is gone).
+- `mode=auto`: **one spawn, zero gates** — every judgment point is resolved with the recommended option and logged to `works/{WORK}/DECISIONS.md` plus the final report's `## 자동 결정 사항` section.
+- Execution (STEP C: builder → verifier → committer) never gates on the user — only WORK creation (specifier) and planning (planner) do.
+
+**Sub-agent Spawn Count:**
 
-| Mode | Spawns | Agents |
-|------|--------|--------|
-| direct | 3 | specifier + builder + verifier+committer |
-| pipeline | 3 | specifier+planner + builder + verifier+committer |
-| full (N TASKs) | 2+2N | specifier+planner + scheduler + [builder + verifier+committer] × N |
+| Main → Orchestrator | → Specifier | → Planner | → Builder | → Verifier | → Committer | Total |
+|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
+| 1 | 1 | 1 | N | N | N | **3 + 3N** |
 
-Compared to the previous approach (separate spawns): **6 TASKs: 20 → 14 spawns (-30%)**.
+`gated` vs `auto` doesn't change the spawn count — only whether execution pauses for approval (see [Approval Gates, Nested Autonomy, and DECISIONS.md](#approval-gates-nested-autonomy-and-decisionsmd) below, `agent-flow.md` § 4).
 
-All three modes output to `works/WORK-NN/` and guarantee `result.md` + `COMMITTER DONE` callback.
+Every WORK outputs to `works/WORK-NN/` and guarantees `result.md` + `DECISIONS.md`.
 
-### WORK (Multi-Task, full mode)
+### WORK and TASK
 
-A two-level hierarchy for complex features:
+A two-level hierarchy:
 
 ```
 WORK (unit of work)       A single goal. The unit requested by the user.
@@ -406,94 +396,83 @@ WORK (unit of work)       A single goal. The unit requested by the user.
     └── result            Completion proof. Auto-generated after verification.
 ```
 
-### pipeline mode (Single Task, Delegated)
-
-Subagent-delegated path for moderate single tasks. specifier+planner run in a single spawn, then builder, then verifier+committer.
-
-```
-Main Claude → [specifier+planner](opus) → builder(sonnet) → [verifier+committer](haiku)
-              (each group called as a single spawn by Main Claude)
-```
-
-### direct mode (Trivial)
-
-Main Claude calls specifier, which determines direct mode and returns a dispatch XML. Main Claude then calls builder (implements the change) and verifier+committer (verifies and commits) as a single spawn.
+Orchestrator nests planner once, then loops builder → verifier → committer per TASK in DAG order.
 
 ```
-Main Claude → specifier: Analyze → return dispatch XML → [back to Main Claude]
-Main Claude → builder: Implement → Self-check → [back to Main Claude]
-Main Claude → [verifier+committer]: Verify → Commit → result.md
+orchestrator → planner(opus, nested)                 → PLAN.md + TASK DAG
+             → [builder(sonnet) → verifier(haiku) → committer(haiku)] × N   ← STEP C, no gates
+              (each nested spawn made by orchestrator, not Main Claude)
 ```
 
 ---
 
 ## Pipeline
 
-### WORK Pipeline (Complex)
+### WORK Pipeline (Nested)
 
-> Subagents cannot nest — Main Claude (CLI terminal) orchestrates every call.
+> Main Claude spawns `orchestrator` **once**. Every other call below is a *nested* sub-agent spawn made by orchestrator itself — Main Claude is not in this loop.
 
 ```
-                               Main Claude (orchestrator)
-                    ┌─────────────┼──────────────────────┐
-                    │             │                       │
-  specifier        planner          scheduler         builder          verifier         committer
- ┌──────────┐    ┌─────────┐     ┌──────────┐     ┌──────────┐     ┌──────────┐     ┌──────────┐
- │Request   │────▶│Create   │────▶│Dependency│────▶│Code      │────▶│Build/Test│────▶│Result    │
- │Analysis  │     │WORK/TASK│     │DAG+Order │     │Implement │     │Verify    │     │→ git     │
- └──────────┘    └─────────┘     └──────────┘     └────┬─────┘     └────┬─────┘     └────┬─────┘
-                                                        │                │                │
-                                                        └── Retry on fail┘                │
-                                                           (max 3 times)                  │
-                                                                          Next TASK loop ◀┘
+Main Claude ── spawn once (mode=gated|auto) ──▶ orchestrator
+                                                      │  nested spawn (depth 2)
+                    ┌─────────────┬──────────────────┴──────────────────┐
+                    │             │                                     │
+  specifier        planner                    builder          verifier         committer
+ ┌──────────┐    ┌─────────┐               ┌──────────┐     ┌──────────┐     ┌──────────┐
+ │Request   │────▶│Create   │──────────────▶│Code      │────▶│Build/Test│────▶│Result    │
+ │Analysis  │     │WORK/TASK│               │Implement │     │Verify    │     │→ git     │
+ └────┬─────┘     └────┬────┘               └────┬─────┘     └────┬─────┘     └────┬─────┘
+      │                │                         │                │                │
+ [GATE-1]          [GATE-2]                      └── Retry on fail┘                │
+ (gated mode        (gated mode                     (max 3 times, then             │
+  only; yield        only; yield                    <needs-decision> escalation)   │
+  + SendMessage       + SendMessage)                                Next READY TASK loop ◀┘
+  to resume)                              (orchestrator schedules the DAG itself — STEP C, no gate)
 ```
 
-### pipeline mode (Simple → Delegated)
+- `mode=gated` (default): pause + yield at **[GATE-1]** (after specifier) and **[GATE-2]** (after planner); resume via `SendMessage(agentId, decision)`, fallback to log-based re-spawn.
+- `mode=auto`: no gates — orchestrator completes the entire diagram in one spawn and records any judgment calls to `DECISIONS.md`.
+- STEP C (the builder → verifier → committer loop) never gates on the user, in either mode.
 
-```
-  specifier+planner    builder       verifier+committer
- ┌──────────────┐     ┌──────────┐     ┌────────────────┐
- │PLAN          │────▶│Code      │────▶│Build/Test      │
- │+TASK         │     │Implement │     │Verify→ git     │
- └──────────────┘     └──────────┘     └────────────────┘
-   (opus, 1 spawn)     (sonnet)         (haiku, 1 spawn)
-              ← each group called by Main Claude →
-```
-
-### direct mode (Trivial)
+### Stage Detail
 
 ```
-  specifier        builder                    verifier+committer
- ┌──────────┐     ┌──────────────────────┐     ┌──────────────┐
- │ Analyze  │────▶│ Implement → Self-chk │────▶│Verify→Commit │
- │ dispatch │     └──────────────────────┘     │→ result      │
- └──────────┘      (no build/test required)    └──────────────┘
+  specifier   →  planner  →  [builder → verifier → committer] × N
+ ┌──────────┐   ┌─────────┐   ┌──────────┐  ┌──────────┐  ┌────────────┐
+ │Request   │──▶│PLAN     │──▶│Code      │─▶│Build/Test│─▶│Result→ git │
+ │Analysis  │   │+TASK DAG│   │Implement │  │Verify    │  │            │
+ └──────────┘   └─────────┘   └──────────┘  └──────────┘  └────────────┘
+   (opus)          (opus)       (sonnet)       (haiku)        (haiku)
+              ← all nested spawns made by orchestrator →
 ```
 
 ### Agents
 
-Six agents work together in a clean, isolated pipeline:
+Six agents work together in a clean, isolated pipeline — Main Claude spawns only `orchestrator`; orchestrator nests the rest:
 
 | Agent | Role | Model | Permission | MCP | Spawn |
 |-------|------|-------|------------|-----|-------|
-| **specifier** | `[]` tag detection, execution-mode selection (direct/pipeline/full), PLAN creation, WORK-LIST management, returns dispatch XML for all modes | **opus** | read + dispatch | Serena (codebase exploration), sequential-thinking (complexity check) | combined with planner (pipeline/full) |
-| **planner** | Create WORK + decompose TASKs + generate PLAN.md (full mode) + pre-create progress templates | **opus** | read-only | Serena (codebase exploration), sequential-thinking (task decomposition) | combined with specifier |
-| **scheduler** | Manage DAG for a specific WORK + run pipeline with sliding window context | **haiku** | read + dispatch | — | standalone (full mode only) |
-| **builder** | Code implementation + progress.md checkpoint recording | **sonnet** | full access | Serena (symbol-level explore/edit) | standalone |
-| **verifier** | Progress gate (Status=COMPLETED) → build/lint/test verification (read-only) | **haiku** | read + execute | — | combined with committer |
-| **committer** | Gate check (progress.md) → write result.md → git commit → COMMITTER DONE callback | **haiku** | read + write + git | — | combined with verifier |
+| **orchestrator** | Nests specifier→(planner)→builder→verifier→committer; schedules the TASK DAG (STEP C); mediates fixed/dynamic gates; batch-records the activity log | **opus** | read + nested spawn | Serena (optional) | spawned **once** by Main Claude per WORK |
+| **specifier** | `[]` tag detection, requirement analysis, complexity assessment, WORK-LIST management, returns dispatch XML | **opus** | read + dispatch | Serena (codebase exploration), sequential-thinking (complexity check) | nested by orchestrator |
+| **planner** | Create WORK + decompose TASKs + generate PLAN.md + pre-create progress templates | **opus** | read-only | Serena (codebase exploration), sequential-thinking (task decomposition) | nested by orchestrator |
+| **builder** | Code implementation + progress.md checkpoint recording | **sonnet** | full access | Serena (symbol-level explore/edit) | nested by orchestrator, per TASK |
+| **verifier** | Progress gate (Status=COMPLETED) → build/lint/test verification (read-only) | **haiku** | read + execute | — | nested by orchestrator, per TASK |
+| **committer** | Gate check (progress.md) → write result.md → git commit | **haiku** | read + write + git | — | nested by orchestrator, per TASK |
+
+> Activity-log recording is done **once, by orchestrator**, on behalf of the agent it just nested — child agents do not write logs themselves.
 
 ### Support Files (included in Plugin)
 
-In addition to the 6 pipeline agents, the plugin includes 6 support files that agents reference at startup.
-These are located in `plugin/skills/sdd-pipeline/references/` (synced from `develop/references/`):
+In addition to the 6 pipeline agents, the plugin includes 7 support files that agents reference at startup.
+These are located in `plugin/references/` (synced from `develop/references/`); when installed via npm they land in `.claude/references/`:
 
 | File | Purpose |
 |------|---------|
-| `agent-flow.md` | Pipeline orchestration rules — how Main Claude calls each agent in sequence |
+| `agent-flow.md` | Pipeline orchestration rules — Main Claude's trigger/gate boundary + orchestrator's internal nested-spawn flow |
+| `context-policy.md` | Sliding window context transfer rules between agents |
 | `file-content-schema.md` | Single source of truth for all file formats (PLAN.md, TASK.md, progress.md, result.md) |
+| `ref-cache-protocol.md` | 4-step ref-cache protocol — checks `<ref-cache>` in the dispatch XML and skips disk reads when cached references are present |
 | `shared-prompt-sections.md` | Shared prompt sections with cache_control — reduces repeated token cost up to 90% |
-| `context-policy.md` | Sliding window context transfer rules between agents |
 | `work-activity-log.md` | Activity log format for builder stage tracking |
 | `xml-schema.md` | XML communication format for dispatch and task-result messages |
 
@@ -507,6 +486,8 @@ works/
 ├── WORK-01/                          ← "User Authentication"
 │   ├── PLAN.md                       ← Plan + dependency graph
 │   ├── PROGRESS.md                   ← Progress tracking (auto-updated)
+│   ├── work_WORK-01.log               ← Orchestrator activity log (STAGE_*/GATE_WAIT/DECISION_WAIT/DECISION)
+│   ├── DECISIONS.md                  ← Auto-decisions + resolved gate decisions (PENDING to RESOLVED)
 │   ├── TASK-00.md                    ← Task specification
 │   ├── TASK-00_progress.md           ← Real-time checkpoint (builder writes)
 │   ├── TASK-00_result.md             ← Completion report (committer writes)
@@ -525,6 +506,8 @@ works/
 | Completion report | `TASK-NN_result.md` |
 | Plan | `PLAN.md` |
 | Work progress | `PROGRESS.md` |
+| Activity log | `work_{WORK_ID}.log` (orchestrator writes; drives gate resume) |
+| Decision log | `DECISIONS.md` (auto-decisions + user-approved decisions, `PENDING`/`RESOLVED`) |
 
 ### WORK-LIST.md
 
@@ -572,7 +555,7 @@ Requests without `[]` tags are handled directly by Claude without routing. If yo
 
 ### Parallel TASKs
 
-The planner creates dependency-aware TASK graphs. Independent TASKs (same `blockedBy` set) can be dispatched concurrently by the scheduler — mention it when approving:
+The planner creates dependency-aware TASK graphs. Independent TASKs (same `blockedBy` set) are dispatched concurrently by orchestrator's DAG scheduling (STEP C) — mention it when approving:
 
 ```
 > Approve. Run independent tasks in parallel.
@@ -586,7 +569,7 @@ If Claude loses context mid-pipeline, you can always resume:
 > Resume WORK-02 from where it stopped
 ```
 
-The scheduler reads `PROGRESS.md` to determine the last completed TASK and continues.
+Orchestrator reads `work_{WORK}.log` (and `PROGRESS.md`) to determine the last completed stage/TASK and continues — reattaching via `SendMessage` if its parked handle is still alive, or a fresh nested re-spawn reconstructed from the log otherwise. An unresolved `GATE_WAIT`/`DECISION_WAIT` is always re-presented rather than skipped.
 
 ---
 
@@ -595,11 +578,15 @@ The scheduler reads `PROGRESS.md` to determine the last completed TASK and conti
 ```
 User: [new-feature] Build a comment feature for the blog system.
 
-Claude: [specifier → WORK path]
-  Complexity: 4+ files, DB schema change, multiple modules
+Claude: [Main Claude spawns orchestrator once, mode=gated]
+
+Claude: [orchestrator nests specifier → WORK path]
+  Complexity: 4+ files, DB schema change, multiple modules → Large
   → Creating new WORK
 
-Claude: [planner]
+  [GATE-1] Requirement approved — continuing to planner
+
+Claude: [orchestrator nests planner]
   Project analysis
      Tech Stack: Next.js + Prisma + PostgreSQL
      Existing code: Post CRUD done, Auth done
@@ -612,11 +599,11 @@ Claude: [planner]
   WORK-03: TASK-03: Frontend comment component         ← TASK-00 (parallelizable)
   WORK-03: TASK-04: Integration + notifications        ← TASK-02, TASK-03
 
-  Do you approve?
+  [GATE-2] Do you approve this plan?
 
-User: Approve. Run automatically.
+User: Approve.
 
-Claude: [scheduler → auto mode]
+Claude: [orchestrator resumed via SendMessage → STEP C: DAG execution, no further gates]
   WORK-03: TASK-00 → builder → verifier ✅ → committer [a1b2c3d]
   WORK-03: TASK-01 → builder → verifier ✅ → committer [d4e5f6g]
   WORK-03: TASK-02 → builder → verifier ✅ → committer [h7i8j9k]
@@ -624,6 +611,7 @@ Claude: [scheduler → auto mode]
   WORK-03: TASK-04 → builder → verifier ✅ → committer [p3q4r5s]
 
   🎉 WORK-03 completed! 5 tasks, 5 commits
+  ## 자동 결정 사항: 없음 (both gates required explicit approval)
 ```
 
 ---
@@ -664,13 +652,13 @@ Each agent file follows a consistent four-section structure:
 
 Each agent reads 4-5 shared reference files (shared-prompt-sections.md, file-content-schema.md, xml-schema.md, etc.) on startup — totaling ~26 file reads across a full pipeline. **ref-cache** eliminates this redundancy:
 
-1. **First agent** reads reference files normally and returns `<ref-cache>` in its result XML
-2. **Main Claude** copies ref-cache into the next agent's dispatch
-3. **Subsequent agents** use cached content instead of reading files from disk
+1. **First agent** (orchestrator itself, per its STARTUP step) reads reference files normally and can return `<ref-cache>` when nesting the next agent's dispatch
+2. **Orchestrator** copies ref-cache into each nested child's dispatch instead of Main Claude relaying it
+3. **Subsequent nested agents** use cached content instead of reading files from disk
 
-Phase 2 (selective delivery) further reduces token usage by passing only the sections each agent needs — not the full file contents. The section mapping per agent is defined in `agent-flow.md`.
+The protocol itself is defined in `references/ref-cache-protocol.md` (4 steps). Phase 2 (selective delivery) further reduces token usage by passing only the sections each agent needs — not the full file contents. The section mapping per agent is defined in `agent-flow.md`.
 
-**Measured impact** (direct mode, 3 agents):
+**Measured impact** (3 agents):
 - File reads: 14 → 5 (**-64%**)
 - Token usage: ~85K → ~72K (**-15%**)
 
@@ -689,10 +677,10 @@ This ensures:
 
 ### Context Isolation
 
-Each subagent runs in an independent context. Even if the builder creates 50 files using 20,000 tokens, the scheduler only receives a 3-line summary.
+Each subagent runs in an independent context. Even if the builder creates 50 files using 20,000 tokens, orchestrator (which nested it) only receives a 3-line summary via the sliding-window context-handoff.
 
 ```
-scheduler's context after 5 TASKs:
+orchestrator's context after 5 TASKs:
 
   PLAN.md (loaded once)                              ~500 tokens
   WORK-01: TASK-00 result: "20 files, PASS"           ~200 tokens
@@ -714,113 +702,19 @@ scheduler's context after 5 TASKs:
 | Tracking | Scroll chat history | File-based (PLAN.md, result.md) |
 | Verification | Manual | Automated (build/lint/test) |
 
-### Router Rule Config (`.agent/router_rule_config.json`)
+### Approval Gates, Nested Autonomy, and DECISIONS.md
 
-The specifier reads `.agent/router_rule_config.json` from the project root to determine routing criteria. If the file is absent, the specifier uses its built-in defaults.
+Main Claude spawns `orchestrator` exactly once per WORK; orchestrator alone decides when to nest specifier/planner/builder/verifier/committer, and when to pause and ask a human. Since nested sub-agents cannot prompt the user directly, **every approval or decision is surfaced at the Main Claude boundary**:
 
-**File location:**
-```
-{project-root}/.agent/router_rule_config.json
-```
+- **Fixed gates** (`<gate type="stage">`) — exactly two, and only in `mode=gated` (the default): **[GATE-1]** right after specifier (Requirement.md ready), and **[GATE-2]** right after planner (PLAN.md + TASK DAG ready).
+- **Dynamic gates** (`<gate type="decision">`) — raised by orchestrator or any nested child, at *any* point (design trade-off, scope creep, destructive change, 3 failed retries, ambiguous requirement…), carrying `<context>` + `<options>` + `<recommended>`.
+- At a gate, orchestrator **yields (parks)** rather than exiting. Main Claude presents the gate, waits for the human, then resumes the parked orchestrator with `SendMessage(agentId, decision)` — context is preserved, no re-reading of files. If the handle is lost (new session, crashed terminal), Main Claude re-spawns orchestrator with the `WORK_ID`; orchestrator replays `work_{WORK}.log` and re-presents the exact same unresolved gate — **an unapproved gate is never silently skipped**, because `STAGE_DONE` is only ever written *after* the gate resolves.
+- Once the final report lands, Main Claude calls `TaskStop(agentId)` to release the parked handle.
+- `mode=auto` ("auto"/"자동으로" in the request) skips every gate: orchestrator resolves each judgment point with its own recommended option and completes in a single spawn.
 
-**JSON structure:**
-```json
-{
-  "$schema": "http://uc-taskmanager.local/schemas/specifier-rules/v1.0.json",
-  "version": "1.1.0",
-  "description": "Specifier execution-mode decision criteria. Customize per project.",
-  "decision_flow": [
-    "1. build_test_required? → false → direct",
-    "2. single_domain + sequential DAG → pipeline",
-    "3. any full_conditions met → full"
-  ],
-  "rules": {
-    "direct": {
-      "criteria": {
-        "build_test_required": false,
-        "note": "File/line count irrelevant. If no build/test needed → direct (text edits, config changes, simple substitutions)"
-      }
-    },
-    "pipeline": {
-      "criteria": {
-        "build_test_required": true,
-        "single_domain_only": true,
-        "max_tasks": 5,
-        "dag_complexity": "sequential"
-      }
-    },
-    "full": {
-      "criteria": {
-        "any_of": [
-          "task_count > 5",
-          "dag_complexity == complex (2+ dependency levels)",
-          "multi_domain == true (BE + FE simultaneously)",
-          "new_module == true (design → implement → verify multi-phase)",
-          "partial_rollback_needed == true"
-        ]
-      }
-    }
-  },
-  "customization_guide": {
-    "doc-heavy projects (md edits)": "Widen direct scope. Most build_test_required=false cases → direct",
-    "code-heavy projects": "Center on pipeline/full. Simple bug fixes → pipeline, multi-domain → full",
-    "max_tasks tuning": "Adjust pipeline max_tasks between 3–7 based on team size or context limits"
-  }
-}
-```
+Every decision — whether a human approved it or orchestrator auto-resolved it — is written to `works/{WORK_ID}/DECISIONS.md` (status `PENDING` while parked, `RESOLVED` once settled) and summarized in the final report's `## 자동 결정 사항` section. STEP C itself (the builder → verifier → committer TASK loop) never gates, in either mode — it's the part of the pipeline nobody needs to approve.
 
-**Key fields:**
-| Field | Description |
-|-------|-------------|
-| `rules.direct.criteria.build_test_required` | `false` → specifier handles implementation, then committer commits |
-| `rules.pipeline.criteria.max_tasks` | Max task count before escalating to full (default: 5) |
-| `rules.pipeline.criteria.dag_complexity` | `sequential` only; complex DAG → escalates to full |
-| `rules.full.criteria.any_of` | List of conditions — any match triggers full mode |
-
-**Fallback behavior:** If `.agent/router_rule_config.json` is absent or malformed, the specifier falls back to its built-in defaults (equivalent to the structure above).
-
-**Per-project customization example:**
-
-For a documentation-heavy project where most changes are text edits:
-```json
-{
-  "rules": {
-    "direct": {
-      "criteria": { "build_test_required": false }
-    },
-    "pipeline": {
-      "criteria": { "max_tasks": 3, "single_domain_only": true, "dag_complexity": "sequential" }
-    }
-  }
-}
-```
-
-For a monorepo with strict build requirements:
-```json
-{
-  "rules": {
-    "pipeline": {
-      "criteria": { "max_tasks": 7 }
-    },
-    "full": {
-      "criteria": {
-        "any_of": ["task_count > 7", "multi_domain == true"]
-      }
-    }
-  }
-}
-```
-
-### Three Execution Modes
-
-The specifier matches effort to complexity via `execution-mode`:
-- **direct**: 1-line typo fix — Main Claude calls specifier, which returns a dispatch XML. Main Claude then calls builder (implements) + [verifier+committer] (verifies and commits). 3 spawns total.
-- **pipeline**: Moderate fix — Main Claude calls [specifier+planner] → builder → [verifier+committer] in sequence. 3 spawns total.
-- **full**: Complex features — [specifier+planner] → scheduler → [builder + verifier+committer] × N. 2+2N spawns total.
-
-Agent pairs are combined into a single spawn to eliminate redundant context loading. This reduces total spawns by **~30%** across all modes.
-
-All three modes output to `works/WORK-NN/` with identical artifact structure (PLAN.md + result.md + COMMITTER DONE callback), ensuring Runner integration works regardless of mode.
+Both modes (gated/auto) output to `works/WORK-NN/` with identical artifact structure (PLAN.md + result.md + `DECISIONS.md`), ensuring downstream integration works regardless of mode.
 
 ### Structured Agent Communication
 
@@ -828,7 +722,7 @@ Instead of ambiguous natural language prompts, agents communicate using structur
 
 **Dispatch Format** (Caller → Receiver):
 ```xml
-<dispatch to="builder" work="WORK-03" task="TASK-00" execution-mode="pipeline">
+<dispatch to="builder" work="WORK-03" task="TASK-00">
   <context>
     <project>uc-taskmanager</project>
     <language>ko</language>
@@ -848,7 +742,7 @@ Instead of ambiguous natural language prompts, agents communicate using structur
 <task-result work="WORK-03" task="TASK-00" agent="builder" status="PASS">
   <summary>Created shared-prompt-sections.md and xml-schema.md</summary>
   <files-changed>
-    <file action="created" path="agents/shared-prompt-sections.md">Common sections with cache_control</file>
+    <file action="created" path="references/shared-prompt-sections.md">Common sections with cache_control</file>
   </files-changed>
   <verification>
     <check name="file_existence" status="PASS">Both files created</check>
@@ -862,7 +756,7 @@ Instead of ambiguous natural language prompts, agents communicate using structur
 - **Prompt Caching**: Common sections (Output Language Rule, Build Commands) are marked with Anthropic API `cache_control`, saving up to **90% on repeated tokens**
 - **Scalability**: Cache hit rates improve with WORK count (5 TASKs at ~0.03 tokens/token vs 2K+ tokens without cache)
 
-See `agents/xml-schema.md` for complete format, and `agents/shared-prompt-sections.md` for cacheable sections.
+See `references/xml-schema.md` for complete format, and `references/shared-prompt-sections.md` for cacheable sections.
 
 ### Sliding Window Context Transfer
 
@@ -893,24 +787,6 @@ Each agent outputs a **context-handoff** — a structured reasoning document, no
 
 See `docs/spec_sliding-window-context.md` for full design details.
 
-### External System Callback (Optional)
-
-uc-taskmanager is generic by default. To integrate with an external system (e.g., a CI/CD backend), add callback URLs to `CLAUDE.md`:
-
-```markdown
-## Task Callbacks
-TaskCallback: http://localhost:3000/api/v1/runner/{{executionId}}/task-result
-ProgressCallback: http://localhost:3000/api/v1/runner/{{executionId}}/task-progress
-CallbackToken: <your-token>
-```
-
-- **No config** → works as-is, no external calls made
-- **TaskCallback** → committer POSTs result after each TASK commit
-- **ProgressCallback** → builder POSTs checkpoint after each progress.md update
-- Callback failures are non-fatal — a warning is printed and the pipeline continues
-
-See `docs/spec_callback-integration.md` for payload schema and implementation guide.
-
 ---
 
 ## Output Language
@@ -963,8 +839,9 @@ Place a file with the same name in `.claude/agents/` to override.
 
 | What | File | Section |
 |------|------|---------|
-| Routing criteria | `specifier.md` | 3-2. Execution-Mode 결정 |
-| Approval policy | `scheduler.md` | Phase 1: User Approval |
+| Complexity criteria | `specifier.md` | 4. 역할 결정 |
+| Approval gates / mode handling | `orchestrator.md` | 3-3. 게이트 및 동적 의사결정 |
+| TASK DAG scheduling / retries | `orchestrator.md` | STEP C: TASK DAG 실행 |
 | Commit message format | `committer.md` | Step 3: Stage + Commit |
 | Verification steps | `verifier.md` | Verification Pipeline |
 | Task granularity | `planner.md` | Task Decomposition Rules |
@@ -995,27 +872,32 @@ Auto-detected from project files. No configuration needed.
 uc-taskmanager/
 ├── develop/                 ← Source of truth (edit here)
 │   ├── agents/              ← 6 agent prompts (language-agnostic)
-│   │   ├── specifier.md     ← [] tag detection + execution-mode routing
+│   │   ├── orchestrator.md  ← Nested spawn coordinator: specifier→(planner)→builder→verifier→committer, TASK DAG scheduling, gates/decisions, batch log
+│   │   ├── specifier.md     ← [] tag detection + requirement analysis
 │   │   ├── planner.md       ← WORK creation + TASK decomposition
-│   │   ├── scheduler.md     ← DAG management + pipeline orchestration
 │   │   ├── builder.md       ← Code implementation
 │   │   ├── verifier.md      ← Build/lint/test verification
 │   │   └── committer.md     ← git commit + result.md
-│   ├── references/          ← 6 support files (shared across agents)
-│   │   ├── agent-flow.md    ← Pipeline orchestration rules
-│   │   ├── file-content-schema.md  ← File format definitions
-│   │   ├── shared-prompt-sections.md  ← Cacheable shared sections
-│   │   ├── context-policy.md    ← Sliding window context rules
-│   │   ├── work-activity-log.md ← Activity log format
-│   │   └── xml-schema.md    ← XML communication format
-│   ├── skills/              ← Skill definitions
-│   └── hooks/               ← Hook scripts
+│   ├── references/          ← 7 support files (shared across agents)
+│   │   ├── agent-flow.md          ← Pipeline orchestration rules
+│   │   ├── context-policy.md      ← Sliding window context rules
+│   │   ├── file-content-schema.md ← File format definitions
+│   │   ├── ref-cache-protocol.md  ← ref-cache protocol (4 steps)
+│   │   ├── shared-prompt-sections.md ← Cacheable shared sections
+│   │   ├── work-activity-log.md   ← Activity log format
+│   │   └── xml-schema.md          ← XML communication format
+│   ├── skills/              ← Skill definitions (sdd-pipeline, uctm-init, work-pipeline, work-status)
+│   └── .claude-plugin/
+│       └── plugin.json      ← Plugin manifest source (name, version, agents array)
 ├── npm/                     ← npm package (published as `uctm`)
-│   ├── agents/              ← Synced from develop/agents/ + develop/references/
+│   ├── agents/              ← Synced from develop/agents/
+│   ├── references/          ← Synced from develop/references/ (7 files)
+│   ├── skills/              ← Synced from develop/skills/ (4 SKILL.md)
 │   ├── bin/cli.mjs          ← CLI entry point (uctm init/update)
 │   ├── lib/                 ← CLI implementation (constants.mjs, init.mjs, update.mjs)
-│   ├── .agent/              ← Default router config bundled with npm
-│   │   └── router_rule_config.json
+│   ├── .claude-plugin/
+│   │   └── plugin.json      ← Synced from develop/.claude-plugin/plugin.json
+│   ├── README.md            ← Synced from README.md (exposed on npmjs.com)
 │   ├── package.json         ← npm package config
 │   ├── .npmignore
 │   └── LICENSE
@@ -1024,17 +906,15 @@ uc-taskmanager/
 │   ├── references/          ← Synced from develop/references/
 │   ├── skills/              ← Plugin skills
 │   │   ├── sdd-pipeline/
-│   │   │   ├── SKILL.md     ← Skill manifest
-│   │   │   └── references/  ← Synced from develop/references/
+│   │   │   └── SKILL.md     ← Skill manifest
+│   │   ├── uctm-init/
+│   │   │   └── SKILL.md     ← /uctm-init (setup works/, CLAUDE.md, permissions)
 │   │   ├── work-pipeline/
 │   │   │   └── SKILL.md
-│   │   ├── work-status/
-│   │   │   └── SKILL.md
-│   │   └── init/
-│   │       └── SKILL.md     ← /uctm-init (setup works/, CLAUDE.md, permissions)
-│   ├── .claude-plugin/
-│   │   └── plugin.json      ← Plugin manifest (name, version, agents array)
-│   └── README.md
+│   │   └── work-status/
+│   │       └── SKILL.md
+│   └── .claude-plugin/
+│       └── plugin.json      ← Plugin manifest (name, version, agents array)
 ├── .claude/                 ← Local Claude settings (not committed)
 │   └── settings.local.json
 ├── README.md                ← English (default, this file)
@@ -1044,16 +924,14 @@ uc-taskmanager/
 ├── docs/                    ← Design specifications
 │   ├── spec_pipeline-architecture_v1.3.md  ← Pipeline architecture v1.3 (ref-cache, Specifier-based)
 │   ├── spec_sliding-window-context.md      ← Sliding window context design
-│   ├── spec_callback-integration.md        ← External system callback integration
 │   ├── spec_SDD_with_ucagent_requirement.md ← SDD v1.5 requirement management system design
 │   ├── pipeline-architecture-v1.3-visual.html ← Interactive pipeline visualization (with ref-cache tab)
 │   ├── SDD-requirement-visual.html         ← Interactive SDD visualization (with ref-cache tab)
-│   ├── callback-integration-visual.html    ← Interactive callback visualization
 │   ├── sliding-window-context-visual.html  ← Interactive sliding window visualization
 │   └── _archive/                           ← Legacy docs (Router-based)
 └── works/                   ← WORK directories (auto-generated)
     ├── WORK-LIST.md          ← Master index
-    ├── WORK-01/              ← all modes output here (direct/pipeline/full)
+    ├── WORK-01/              ← every WORK outputs here
     └── ...
 ```