/work-milestone

Execute a groomed milestone. Reads the dispatch plan produced by /groom-milestone, creates an integration branch, spawns parallel kage-bunshin sessions (independent claude -p processes) per wave — each in its own worktree with full orchestrator capabilities (Agent tool, TeamCreate, all MCP servers). Sequentially merges their branches and lands the integration branch to main when all waves complete.

Entry Conditions

• Milestone number provided as argument
• Dispatch plan exists: plan/milestone-{N}-dispatch.yaml
• All items in dispatch plan are groomed (groomed: true)
• Backlog MCP and SAM MCP responding
• Clean git state on main

Run /groom-milestone {N} first if the dispatch plan is missing or stale.

Main Workflow

flowchart TD
    Start(["Input: milestone number N"]) --> LoadPlan["Step 1: Load Dispatch Plan<br>Read plan/milestone-{N}-dispatch.yaml<br>Output: waves, conflict groups, quality gates"]

    LoadPlan --> Validate{"Step 2: Validate Plan<br>File exists?<br>All items still open?<br>Observable: file + backlog_list_issues(milestone=N)"}

    Validate -->|"Plan missing"| Block1(["BLOCKED — run /groom-milestone {N}"])
    Validate -->|"Items changed since groom"| Regroom["Re-run /groom-milestone {N}"]
    Validate -->|"Plan valid"| CreateBranch
    Regroom --> LoadPlan

    CreateBranch["Step 3: Create Integration Branch<br>github_branches create<br>milestone/{N}-{slug} from main.<br>Switch to integration branch locally."]

    CreateBranch --> FetchItems["Step 3b: Fetch All Items Once<br>Call backlog_view once per issue in the dispatch plan.<br>Store results in context keyed by issue number.<br>DO NOT call backlog_view again for any issue<br>already fetched this session."]

    FetchItems --> WaveLoop["Step 4: Wave Dispatch Loop<br>Read next wave from dispatch plan"]

    WaveLoop --> WaveItems{"Wave has items?"}
    WaveItems -->|"No waves remain"| Land

    WaveItems -->|"Yes"| CreateWorktrees["Step 5a: Create Worktrees<br>For each item in wave:<br>git worktree add worktrees/{slug}<br>from integration branch"]

    CreateWorktrees --> WriteLocks["Step 5b: Write Lock Files<br>For each worktree:<br>Write .claude/kage-bunshin.lock<br>(prevents recursive spawning)"]

    WriteLocks --> SpawnSessions["Step 5c: Spawn Kage-Bunshin Sessions<br>For each worktree: cd into it, then:<br>claude -p --model sonnet<br>--permission-mode auto --output-format json<br>'Load /dh:work-backlog-item #{issue}'<br>Each session is a FULL orchestrator —<br>has Agent tool, TeamCreate, all MCP.<br>All sessions launch as background processes."]

    SpawnSessions --> WaitReturn["Step 6: Wait for All Sessions<br>Poll PIDs for exit.<br>Read result JSON from each<br>/tmp/kb-work-{issue}.json"]

    WaitReturn --> CheckResults{"Step 6a: Any sessions failed?"}
    CheckResults -->|"All succeeded"| ParseResults
    CheckResults -->|"Some failed"| InvestigateFail{"Fixable?"}
    InvestigateFail -->|"Yes"| FixRetry["Fix and re-spawn<br>failed item only"]
    FixRetry --> WaitReturn
    InvestigateFail -->|"No"| SkipFailed["Skip failed item<br>Log to output"]
    SkipFailed --> ParseResults

    ParseResults["Step 6b: Parse Completion Reports<br>Extract STATUS, BRANCH, FILES_CHANGED,<br>COMMITS, NOTES from each result JSON."]

    ParseResults --> MergeLoop["Step 6b: Merge Worktree Branches<br>Sequential merge into integration branch.<br>One at a time, in return order."]

    MergeLoop --> MergeResult{"Any merge conflicts?"}
    MergeResult -->|"All clean"| WaveComplete
    MergeResult -->|"Trivial/Medium conflict"| Resolve["Auto-resolve or spawn<br>conflict-resolution agent"]
    Resolve --> ResolveGates{"Resolution gates pass?"}
    ResolveGates -->|"Yes"| WaveComplete
    ResolveGates -->|"No"| EscalateConflict["Create backlog item<br>for conflict resolution.<br>Add to milestone."]

    MergeResult -->|"Heavy conflict (3+ files)"| EscalateConflict

    WaveComplete["Step 6c: Wave Complete<br>All branches merged.<br>Delete worktree branches."]

    WaveComplete --> DiscoveryRelay["Step 6d: Discovery Relay<br>Build relay document from agent outputs.<br>Include FILES_CHANGED, COMMITS, NOTES<br>in next wave agent prompts."]

    DiscoveryRelay --> PartialCheck{"Any agents returned PARTIAL?"}
    PartialCheck -->|"Yes"| HandlePartial["Create backlog items for<br>blocked tasks. Add to milestone."]
    PartialCheck -->|"No"| NextWaveCheck

    HandlePartial --> NextWaveCheck{"More waves?"}
    NextWaveCheck -->|"Yes"| WaveLoop
    NextWaveCheck -->|"No"| Land

    Land["Step 8: Land Integration Branch<br>Run full quality gate suite<br>pre_merge + post_merge"]

    Land --> FinalResult{"Gates pass?"}
    FinalResult -->|"Fail"| FinalFix["Spawn fix agent in worktree<br>on integration branch."]
    FinalFix --> Land

    FinalResult -->|"Pass"| MergeMain["Step 9: Merge to Main<br>git switch main<br>git merge --no-ff milestone/{N}-{slug}<br>git push origin main"]

    MergeMain --> Complete["Step 10: Complete Milestone<br>/complete-milestone {N}.<br>Delete integration branch."]

    Complete --> Done(["Exit: milestone complete"])

Step 3b: Fetch All Items Once (Before Any Wave)

Before entering the wave dispatch loop, call backlog_view once per issue listed across all waves in the dispatch plan. Store each result in context keyed by issue number.

Fetch-once rule: Do NOT call backlog_view for the same issue more than once per session. Use the already-fetched data for all subsequent references — wave loop iterations, discovery relay construction, result reporting. If an item's state genuinely changes (e.g., after a backlog_update call), replace the cached value with a single new backlog_view call for that issue only.

Pass the fetched data (title, AC, description) into spawned session prompts directly rather than having each spawned session re-fetch — see Step 5c prompt construction.

Dispatch Step (Step 5 Detail)

All items in a wave are independent by construction (guaranteed non-overlapping by the conflict group analysis in the dispatch plan). Each item gets its own worktree and its own kage-bunshin session — an independent claude -p process with full orchestrator capabilities.

Why Kage-Bunshin Instead of TeamCreate

Teammates and subagents do NOT have the Agent tool. The /work-backlog-item flow is an orchestration skill that needs to spawn sub-agents (feature-researcher, codebase-analyzer, python-cli-architect, etc.). A teammate running /work-backlog-item is BLOCKED at the first agent delegation step.

A kage-bunshin session is an independent claude CLI process — a full orchestrator that inherits all MCP servers, skills, plugins, and agents from the project directory. It CAN use Agent tool and TeamCreate internally.

Worktree + Session Setup Per Item

Use the kage-bunshin spawn script — it handles worktree creation, .venv/node_modules symlinking, lock file writing, and process launch:

SPAWN="plugins/development-harness/skills/kage-bunshin/scripts/spawn.py"
PIDS=()
SPAWN_INFO=()

for ISSUE in "${WAVE_ISSUES[@]}"; do
  OUTPUT=$($SPAWN --worktree \
    --branch "${INTEGRATION_BRANCH}" \
    --name "work-item-${ISSUE}" \
    --model sonnet \
    "Load /dh:work-backlog-item #${ISSUE}. Execute the full work flow. \
     You are in a worktree on integration branch ${INTEGRATION_BRANCH}. \
     Use MCP tools for plan artifact discovery — plan/ files are in the root worktree. \
     Prior wave context: ${DISCOVERY_RELAY}")
  PIDS+=($(echo "$OUTPUT" | python3 -c "import sys,json; print(json.load(sys.stdin)['pid'])"))
  SPAWN_INFO+=("$OUTPUT")
done

What the Spawned Session Gets

Via the prompt:

• Issue number (from dispatch plan wave item)
• Integration branch name (from dispatch plan)
• Discovery relay content from prior waves (orchestrator-accumulated, empty for wave 1)
• Instruction to use MCP for artifact discovery

Via capability inheritance (automatic):

• All MCP servers (backlog, SAM, artifact registry)
• All skills (including /dh:work-backlog-item which it loads and executes)
• All agent types (it CAN and WILL spawn sub-agents as needed)
• Full tool access (Agent, TeamCreate, Read, Write, Edit, Bash, etc.)

Via self-discovery (the session does this itself):

• Groomed description and acceptance criteria — via backlog_view(selector="#{issue}")
• Plan artifacts — via artifact_list(issue_number={issue}) then artifact_read(...)
• SAM task plan — via sam_read(plan="P{N}") if a plan exists
• Skills to load — from skills field in SAM task metadata

Monitoring and Result Collection

# Wait for all sessions in the wave to exit
for i in "${!PIDS[@]}"; do
  wait "${PIDS[$i]}"
  EXIT_CODE=$?
  INFO="${SPAWN_INFO[$i]}"
  RESULT_FILE=$(echo "$INFO" | python3 -c "import sys,json; print(json.load(sys.stdin)['result_file'])")

  if [ $EXIT_CODE -eq 0 ] && [ -s "$RESULT_FILE" ]; then
    echo "$(python3 -c "import sys,json; d=json.load(sys.stdin); print(d.get('result','')[:200])" < "$RESULT_FILE")"
  else
    ERROR_FILE=$(echo "$INFO" | python3 -c "import sys,json; print(json.load(sys.stdin)['error_file'])")
    echo "FAILED (exit ${EXIT_CODE}): $(tail -5 "$ERROR_FILE")"
  fi
done

Model Selection

The --model flag on the kage-bunshin controls the spawned session's orchestrator model only. Sub-agents spawned inside that session use their own model per their agent frontmatter definition.

Recommended: --model sonnet for spawned sessions. Haiku viability as orchestrator is an open experiment.

Agent Result Handling

flowchart TD
    Result(["Agent returned"]) --> Status{"Agent output indicates success?"}
    Status -->|"STATUS: COMPLETE — tasks done, changes committed"| Merge["Proceed to merge"]
    Status -->|"STATUS: PARTIAL — some tasks done, some blocked"| Partial["Merge completed work.<br>Create backlog item for remaining tasks.<br>Add to current milestone."]
    Status -->|"Failure — no useful work done"| Failure["Log failure context.<br>Escalate to user:<br>item title, error, agent output summary."]

Discovery Relay Between Waves

After all wave agents return, the orchestrator builds a relay document from their completion reports. This is injected as discovery_relay_content in the next wave's agent prompts.

## Prior Wave Results

### Wave 1 Results

#### Item: #{issue1} — {title1}
- Status: COMPLETE
- Files changed: {file_list}
- Key commits:
  - {hash}: {message}
- Design notes: {notes_if_any}

#### Item: #{issue2} — {title2}
- Status: COMPLETE
- Files changed: {file_list}
- Key commits:
  - {hash}: {message}

Items in the same wave are guaranteed non-overlapping by the dispatch plan's conflict group analysis. The relay provides cross-wave awareness for items with depends_on relationships or shared conflict groups.

For milestones with 5+ waves, cap the relay at the most recent 3 waves.

Merge Conflict Classification

Conflict scope	Classification	Action
0 files	Clean	Merge immediately
1-2 files — whitespace or adjacent additions	Trivial	Auto-resolve, run gates
1-2 files — same function edited differently	Medium	Spawn conflict-resolution agent
1-2 files — file restructured by both worktrees	Heavy	Create backlog item for conflict resolution
3+ files	Heavy	Abort merge, create backlog item

Conflict resolution agent receives both branches' diffs and resolves in-place on the integration branch. No PRs are created for worktree branches — they are local-only, never pushed to origin.

Tools Used

Tool	Purpose
read_dispatch_plan	Read plan/milestone-{N}-dispatch.yaml
dispatch_wave_start	Register wave + items in dispatch state DB before spawning (Step 5)
dispatch_spawn	Background MCP task that launches parallel kage-bunshin sessions for a wave (Step 5c)
dispatch_wave_status	Poll wave progress and detect stale PIDs (Step 6)
dispatch_item_status	Called by spawned sessions to record completion or failure
git worktree add	Create isolated worktree per wave item
claude -p (kage-bunshin)	Spawn independent orchestrator session per item — has Agent tool, TeamCreate, all MCP
backlog_view	Read item description, AC, design decisions (used by spawned sessions)
artifact_list	Discover plan artifacts registered for an issue (used by spawned sessions)
artifact_read	Read plan artifact content from root worktree via MCP (used by spawned sessions)
sam_read	Read SAM task plan for an item (used by spawned sessions)
sam_status / sam_list	Check whether item has a SAM plan (used by spawned sessions)
github_branches create	Create integration branch
github_branches merge	Merge worktree branch into integration branch
github_branches delete	Delete integration branch after landing
run_quality_gates	Execute gate commands from dispatch plan
backlog_list_issues(milestone=N)	Validate plan against current item state

Error Conditions

• Dispatch plan missing: BLOCKED — direct to /groom-milestone {N}
• Items changed since groom: re-run /groom-milestone {N} to regenerate plan
• Backlog MCP unavailable: PROCESS ERROR — report with exact error text
• SAM MCP unavailable: PROCESS ERROR — report with exact error text
• Integration branch already exists: check for stale branch (no commits in 7+ days) — offer to delete and recreate, or resume
• Kage-bunshin session exited non-zero: read error log, investigate if fixable, re-spawn if yes, skip item if no
• Kage-bunshin result contains PARTIAL status: create backlog items for blocked tasks, add to milestone, continue with other items
• All quality gates fail on integration branch: escalate to user before landing
• Main diverged during milestone work: rebase integration branch onto main before landing

References

• Worktree Worker Protocol — full worker lifecycle: setup, direct task execution, quality gates, completion report format, blocker handling, skill loading
• Merge Queue Protocol — merge slot lifecycle, conflict classification, conflict-resolution agent, quality gate commands