Spec Kitty Implement Review

spec-kitty-implement-review

Orchestrate the implement-review loop for Spec Kitty work packages. This skill teaches any agent how to dispatch implementation and review to the configured agents, handle rejection loops, enforce cycle limits, and sequence WPs by dependency graph.

When to Use This Skill

• Implement one or more WPs through the full implement-review cycle
• Coordinate cross-agent workflows (different agents for implement vs review)
• Handle rejection feedback loops with cycle tracking
• Run a full mission sprint (WP01 through WP_N)

Core Concepts

Agent Selection

Spec-kitty selects agents from .kittify/config.yaml:

agents:
  available: [claude, codex, opencode]
  auto_commit: true

The orchestrator does NOT hardcode agent names. Instead:

# Check which agents are configured
spec-kitty agent config list

# The workflow commands handle agent selection internally
spec-kitty agent action implement WP01 --agent <tool> --profile <profile>
spec-kitty agent action review WP01 --agent <tool> --profile <profile>

Agent Capabilities

Not all agents can be dispatched the same way. The dispatch method depends on the agent's CLI capabilities:

Tier 1: Full headless CLI. Orchestrator dispatches and agent runs autonomously. Tier 2: CLI exists but needs workarounds (timeout wrappers, retry). Tier 3: GUI-only or no stable CLI. Orchestrator must run move-task after the agent completes, because the agent cannot run shell commands.

Context Boundaries

Keep implement-review sessions narrowly scoped. Compact after task pivots and avoid combining architecture, debugging, and implementation in one long session. If the work changes mode, preserve the current status and start a fresh compacted context before continuing.

Use subagents whenever a task can be done in an isolated context, even if the work is not part of a large parallel sprint. Good candidates include one WP in a separate worktree, a review pass against a fixed diff, a focused debugging investigation, or validation that can run independently from the orchestrator's next scheduling decision.

Sleep Protection (Unattended Runs)

Unattended implement-review runs are long-lived: a dispatched implementation agent can run for many minutes between the short-lived spec-kitty CLI calls that claim and move WPs. None of those CLI calls hold the host machine awake by themselves — only a process that stays alive for the whole session can. If the host sleeps mid-dispatch (default laptop power settings on macOS), the dispatched agent is silently killed or its connection drops, and nothing in the mission state records why. The failure then looks identical to a flaky or stalled agent (see the Stale WP entry under Troubleshooting).

Before dispatching the first agent, hold a keep-awake assertion for the session and release it when the loop ends:

# macOS: idle-sleep assertion tied to this shell's own PID, backgrounded so
# it does not block. Does NOT prevent lid-close sleep -- for a fully
# unattended run, also disable lid-close sleep while plugged in, or keep
# the lid open.
caffeinate -i -w $$ &
CAFFEINATE_PID=$!

Release it after Step 6 (accept/merge) or whenever the loop is explicitly halted:

kill "$CAFFEINATE_PID" 2>/dev/null || true

Linux (systemd distros — the large majority): systemd-inhibit is the equivalent, and its tail --pid form self-releases when the process dies, mirroring caffeinate -w (no polling):

systemd-inhibit --what=idle:sleep --why="spec-kitty implement-review" \
  tail --pid $$ -f /dev/null &
INHIBIT_PID=$!
# release: kill "$INHIBIT_PID" 2>/dev/null || true

Non-systemd Linux (Alpine, minimal containers) has no standard primitive — it no-ops; rely on the host not sleeping.

Windows: there is no shell one-liner. The keep-awake primitive is the Win32 SetThreadExecutionState API (callable from ctypes, no extra dependency), so it must be held by a live process, not a command — run under spec-kitty-orchestrator (below), or accept that an unattended Windows session can sleep.

Canonical cross-platform home. When spec-kitty-orchestrator (the standalone binary, not an agent-harness session following this skill) drives the run, skip all of the above: it holds the idle-sleep assertion for the duration of orchestrate/resume (on by default; --no-caffeinate to opt out — see its README) behind a single no-op-safe prevent_idle_sleep() seam, which is the intended home for this handling across platforms. That coverage is internal to the binary and does not extend to an agent-harness session following this skill directly — which is exactly why the per-platform guidance above exists.

The Mandatory Workflow Pattern

Every WP MUST follow this state flow:

planned --> claimed --> [workflow implement] --> in_progress --> [agent works] --> for_review --> in_review --> approved or planned

After review rejection (WP moves back to planned with review_status: has_feedback):

planned --> claimed --> [workflow implement] --> in_progress --> [agent fixes] --> for_review --> in_review --> approved or planned

After ALL WPs are approved: run spec-kitty accept --mission <slug> first as the mission-readiness nudge. If acceptance passes, run spec-kitty merge --mission <slug> to merge everything and move WPs to done.

approved unblocks dependents immediately. Do NOT wait for done before starting dependent WPs. The done lane is only reached via mission merge.

To determine what to do next, always run:

spec-kitty next --agent <your-name> --mission <mission-slug>

Note: --mission is the canonical selector. The legacy --feature alias has been removed from the internal/agent commands (e.g. agent action implement, agent action review); always use --mission.

This reads the dependency graph and current lane state and returns the exact command. Do NOT reason about lane transitions yourself.

Step 1: Dispatch Implementation

Implementation is a two-step process: claim the workspace, then dispatch an agent to do the work.

Step 1a: Claim the Workspace

OUTPUT=$(spec-kitty agent action implement WP## --mission <slug> --agent <tool> --profile <profile> 2>&1)

Resolved dispatch identity — pass the values selected by the dispatcher so the claim records what actually ran, rather than copying the authored frontmatter recommendation:

--agent <tool> --profile <resolved-profile> [--model <resolved-model> --invocation-id <op-id>]

Examples:

• --agent claude --profile python-pedro
• --agent claude --profile implementer-ivan
• --agent claude --model claude-haiku-4-5 --profile reviewer-renata --invocation-id <op-id>

Pass --model only with a correlated --invocation-id. The durable Op's mission identity, WP, action, profile, and catalog-winning model must match the claim; missing, malformed, or mismatched evidence fails closed. A direct --profile is resolved through the canonical profile registry. A bare --agent claim remains allowed, but records explicit absence for model/profile instead of fabricating values from frontmatter.

This command:

• Moves WP from planned to in_progress
• Creates or re-enters the worktree workspace
• Generates the implementation prompt file

Capture from output:

• Workspace path: line containing Workspace: cd <path>
• Prompt file: line containing cat <path>

WORKSPACE=$(echo "$OUTPUT" | grep 'Workspace: cd ' | sed 's/.*Workspace: cd //')
PROMPT_FILE=$(echo "$OUTPUT" | grep 'cat ' | sed 's/.*cat //')

Step 1b: Dispatch the Implementing Agent

How you dispatch depends on your execution context.

If you are a Claude Code agent orchestrating via subagents (Task tool):

Task(
    subagent_type="general-purpose",
    description="Implement WP##",
    prompt=f"""You are implementing WP## for mission <slug>.

**CRITICAL: Work in the worktree directory:**
cd {WORKSPACE}

**Read the full implementation prompt:**
cat {PROMPT_FILE}

The prompt contains all context, acceptance criteria, and review feedback
(if re-implementing after rejection).

**Your task:**
1. Read the implementation prompt (contains all details)
2. If re-implementing: Read review feedback, update review_status: "acknowledged"
3. Read existing code paths BEFORE implementing
4. Implement all subtasks
5. Write tests that verify the contract (what the spec says)
6. Integration verification (MANDATORY before moving to for_review):
   - Verify new code is ACTUALLY CALLED from live entry points (not just defined)
   - Grep for imports of your new module in the files that should call it
   - If you created a new function/class, grep the codebase for callers — zero callers means the capability is dead code
   - Verify old code paths are removed or redirected
   - Grep for old function/class names to confirm removal
   CRITICAL: A module with passing tests but no callers is NOT implemented.
   The most common review failure is dead code — tests pass but the capability
   is never invoked from the live command path.
7. Run the project's declared validation command before handoff
8. **Diff-scoped lint sweep (MANDATORY before move-task to for_review)**:
   Catches lint regressions before they reach the cycle-1 reviewer — unused-import
   or formatting violations introduced by a WP should be caught here, scoped to the
   diff only so the implementer does not drown in pre-existing warnings owned by
   other WPs. Use the project's declared linter (see charter / project README for
   the configured command and source-file extension).
   ```bash
   # Replace `<ext>` with the project's source-file extension (e.g. py, ts, rs, go)
   # and `<lint-command>` with the project's configured linter invocation.
   CHANGED_SRC=$(git diff --name-only --diff-filter=AMR HEAD | rg '\.<ext>$' || true)
   if [ -n "$CHANGED_SRC" ]; then
     <lint-command> $CHANGED_SRC
   fi

• The command MUST exit 0. If it does not, fix or run the linter's autofix mode and re-run.
• Paste the final command + exit code into the handoff note (e.g. "<lint-command> diff-scoped check: 0 issues, exit 0").
• On cycle-N re-implementation, use the WP's planning base instead of HEAD: git diff --name-only $(git merge-base HEAD main). 8b. Compiler typecheck (MANDATORY when the WP touches typed sources): A test runner does not replace compiler diagnostics. If the diff includes typed sources (including tests), select the project's configured compiler or typecheck command before execution, matching the repository's CI. If no command is configured, select an available compiler appropriate to the project's language. A failed compiler command fails the gate; never try another command to turn that failure into success. The command MUST exit 0. Paste command + exit code into the handoff note. Reviewers reject the WP if typecheck was skipped or is red.

9. Commit only your deliverables: spec-kitty safe-commit <each file you changed> -m "feat(WP##): <description>"
10. Mark subtasks done: spec-kitty agent tasks mark-status T001 T002 ... --status done
11. Move to for_review: spec-kitty agent tasks move-task WP## --to for_review --note "Ready for review" """, run_in_background=True )

**If you are dispatching to an external CLI agent (Tier 1):**

Build a prompt file and pipe to the agent CLI:

```bash
# Read the generated prompt
PROMPT_CONTENT=$(cat "$PROMPT_FILE")

# Dispatch to configured agent (examples for each CLI)

# Claude Code:
claude -p "$PROMPT_CONTENT" --output-format json -C "$WORKSPACE"

# GitHub Codex:
# move-task writes git/status locks and may touch local sync state; workspace-write/full-auto is too narrow.
printf '%s' "$PROMPT_CONTENT" | codex exec --sandbox danger-full-access -C "$WORKSPACE" -

# Google Gemini:
gemini -p "$PROMPT_CONTENT" --yolo --output-format json -C "$WORKSPACE"

# LLxprt Code (no -C flag; cd into the workspace first):
(cd "$WORKSPACE" && llxprt -p "$PROMPT_CONTENT" --approval-mode=yolo)

# OpenCode:
opencode run "$PROMPT_CONTENT" --format json -C "$WORKSPACE"

# Qwen Code:
qwen -p "$PROMPT_CONTENT" --yolo --output-format json -C "$WORKSPACE"

# Kilocode:
kilocode -a --yolo -j "$PROMPT_CONTENT" -C "$WORKSPACE"

# Augment Code:
auggie --acp "$PROMPT_CONTENT" -C "$WORKSPACE"

# Cursor (Tier 2 -- needs timeout wrapper):
timeout 600 cursor agent -p --force --output-format json "$PROMPT_CONTENT" -C "$WORKSPACE"

If the agent is Tier 3 (GUI-only):

The orchestrator cannot dispatch automatically. Instead:

1. Print the workspace path and prompt file for the human operator
2. Wait for the human to run the agent manually in the workspace
3. After the agent finishes, the orchestrator runs move-task on behalf of the agent (since GUI agents cannot execute CLI commands)

echo "Manual dispatch required for agent: <agent-name>"
echo "Workspace: cd $WORKSPACE"
echo "Prompt: cat $PROMPT_FILE"
echo "After agent completes, run:"
echo "  spec-kitty agent tasks move-task WP## --to for_review --note 'Ready'"

Step 2: Monitor Progress

Check WP status at any time:

spec-kitty agent tasks status

This shows:

• Kanban board with WPs in lanes: planned, claimed, in_progress, for_review, in_review, approved, done
• Progress bar showing completion percentage
• Which WPs are ready for review, in progress, and planned

Use this frequently between dispatch and review steps.

Step 3: Dispatch Review

When a WP reaches for_review, dispatch a review agent.

Step 3a: Claim the Review

OUTPUT=$(spec-kitty agent action review WP## --mission <slug> --agent <tool> --profile <profile> 2>&1)
REVIEW_PROMPT=$(echo "$OUTPUT" | grep -o '/var/folders[^ ]*/spec-kitty-review-WP[0-9]*.md' || echo "$OUTPUT" | grep 'cat ' | sed 's/.*cat //')
WORKTREE=$(echo "$OUTPUT" | grep 'Workspace: cd ' | sed 's/.*Workspace: cd //')

Step 3b: Dispatch the Review Agent

If you are a Claude Code agent (Task tool):

Task(
    subagent_type="general-purpose",
    description="Review WP##",
    prompt=f"""You are reviewing WP## for mission <slug>.

**CRITICAL: Work in the worktree directory:**
cd {WORKTREE}

**Read the full review prompt:**
cat {REVIEW_PROMPT}

The review prompt contains:
- Acceptance criteria for this WP
- Git diff commands with the correct base branch
- Dependency warnings for downstream WPs
- Completion instructions (approve/reject commands)

**Your task:**
1. Read the review prompt (it is the source of truth)
2. Run the git diff commands listed in the prompt
3. Check each acceptance criterion against the diff
4. Check for unrelated changes outside WP scope
5. Issue exactly one verdict:

**If ALL acceptance criteria met:**
spec-kitty agent tasks move-task WP## --to approved --note "Review passed: <summary>"

**If criteria NOT met:**
Write structured feedback to a temp file, then:
spec-kitty agent tasks move-task WP## --to planned --review-feedback-file <feedback-path> --agent <reviewer-agent-id>
""",
    run_in_background=True
)

If dispatching to an external CLI agent:

Build a combined prompt and pipe to the agent. The mandatory instruction ensures the agent runs the move-task command after reviewing.

# Build combined prompt with mandatory instruction.
# Scope temp paths by <mission> so concurrent missions sharing a WP id (e.g.
# two missions both with WP01) do not collide on the same /tmp file (#1831).
printf 'IMPORTANT: After reviewing, you MUST execute the appropriate spec-kitty agent tasks move-task command shown at the bottom of this prompt.\n---\n' > /tmp/review-prompt-<mission>-WP##.md
cat "$REVIEW_PROMPT" >> /tmp/review-prompt-<mission>-WP##.md

# Dispatch to configured reviewer (same CLI patterns as Step 1b)
# Example for codex:
cat /tmp/review-prompt-<mission>-WP##.md | codex exec --sandbox danger-full-access \
  -C "$WORKTREE" --add-dir "$(pwd)" \
  -o "/tmp/review-result-<mission>-WP##.md" -

# Example for claude:
claude -p "$(cat /tmp/review-prompt-<mission>-WP##.md)" --output-format json -C "$WORKTREE"

# Example for gemini:
gemini -p "$(cat /tmp/review-prompt-<mission>-WP##.md)" --yolo --output-format json -C "$WORKTREE"

# Example for llxprt (no -C flag; cd into the worktree first):
(cd "$WORKTREE" && llxprt -p "$(cat /tmp/review-prompt-<mission>-WP##.md)" --approval-mode=yolo)

Capture the reviewer command exit status. If the configured/chosen reviewer fails, do not silently approve and do not fall back to the implementing agent. Print a structured error and halt in unattended/noninteractive mode:

ERROR: Configured reviewer '<agent>' failed (<exit code or reason>).
Options:
  a) Retry with '<agent>' after fixing the error
  b) Switch reviewer: spec-kitty agent action review WP## --mission <slug> --agent <other-reviewer>
  c) Proceed with self-review -- WARNING: no independent review

Self-review fallback is only allowed after an explicit operator decision. The approval command must preserve the failure metadata:

spec-kitty agent tasks move-task WP## --to approved --force \
  --self-review-fallback \
  --intended-reviewer <failed-agent> \
  --reviewer-failure-reason "<exit code or reason>" \
  --note "Self-review fallback after reviewer failure: <summary>"

If the reviewer is Tier 3 (GUI-only) or cannot run move-task:

After the reviewer completes, the orchestrator must:

1. Read the reviewer's output to determine pass/fail
2. Run move-task on behalf of the reviewer

# Read reviewer output, then:
# If approved:
spec-kitty agent tasks move-task WP## --to approved --note "Review passed (by <agent>): <summary>"

# If rejected:
spec-kitty agent tasks move-task WP## --to planned \
  --review-feedback-file /tmp/feedback-<mission>-WP##.md \
  --agent <reviewer-agent-id>

When relaying a verdict, act as the reviewer: pass the reviewer's own --agent identity, not the orchestrator's and not --force.

Step 3c: Verify the Outcome

After review completes:

# Check the WP lane
spec-kitty agent tasks status

# If reviewer output was captured to a file:
cat /tmp/review-result-<mission>-WP##.md

Before final approval, if spec.md references GitHub issues, ensure issue-matrix.json (JSON-first canonical artifact; legacy issue-matrix.md missions are read via failover, never re-authored) exists and every referenced issue has a verdict: fixed, verified-already-fixed, a documented follow-up (deferred-with-followup), or in-mission (being closed by a later WP in this same mission). unknown/empty verdicts block approval. The CLI enforces this guard on move-task --to approved/done. An in-mission verdict passes per-WP approved (so a dependency chain is not blocked on its own downstream WPs) but is rejected on done — resolve every in-mission row to a terminal verdict before the mission merges.

Step 4: Handle Review Rejection

When a reviewer moves a WP back to planned with feedback, the orchestrator must re-dispatch implementation.

What Happens on Rejection

1. The reviewer runs move-task WP## --to planned --review-feedback-file <path> --agent <reviewer-agent-id> (a rejection is an ordinary move and needs no force flag)
2. The event log records the lane transition and review-feedback reference; the WP planning file remains byte-stable.

The implementer resumes the rework with its own --agent identity. The resubmission (--to for_review) and the re-review need no --force, so they are not recorded as an override. The CLI compares the agent tool of the requester with the latest implementer's; it does not prove that the reviewer is an independent agent, so choose reviewers deliberately.

Re-Implementation Steps

1. Re-dispatch implementation using the same two-step pattern from Step 1:

   OUTPUT=$(spec-kitty agent action implement WP## --mission <slug> --agent <tool> --profile <profile> 2>&1)
   WORKSPACE=$(echo "$OUTPUT" | grep 'Workspace: cd ' | sed 's/.*Workspace: cd //')
   PROMPT_FILE=$(echo "$OUTPUT" | grep 'cat ' | sed 's/.*cat //')

   Then dispatch the implementing agent (Step 1b). The prompt file now includes the review feedback.

2. Wait for re-implementation to complete (WP reaches for_review)

3. Re-dispatch review (Step 3)

4. Track the cycle count (max 3):

   • Cycle 1: First rejection
   • Cycle 2: Second rejection
   • Cycle 3: Third rejection triggers ARBITER MODE

Example Rejection Loop

# 1. Reviewer rejected WP03 -- verify status
spec-kitty agent tasks status
# Shows: WP03 in planned (review_status: has_feedback)

# 2. Re-dispatch implementation (two-step pattern)
OUTPUT=$(spec-kitty agent action implement WP03 --mission <slug> --agent <tool> --profile <profile> 2>&1)
WORKSPACE=$(echo "$OUTPUT" | grep 'Workspace: cd ' | sed 's/.*Workspace: cd //')
PROMPT_FILE=$(echo "$OUTPUT" | grep 'cat ' | sed 's/.*cat //')

# 3. Dispatch fixing agent (Task tool or CLI -- see Step 1b)
# Include cycle info: "This is cycle 2/3"

# 4. Wait for WP to reach for_review, then re-dispatch review (Step 3)

Step 5: Arbiter Mode (After 3 Rejections)

If a WP is rejected 3 times, the orchestrator steps in as arbiter. Issue the arbiter decision while the WP is still in planned after the third rejection: an override is recorded only for a forced planned to approved/done, so after an in_review to in_progress rejection the decision is not recorded as an override.

Arbiter Decision Options

Option A -- Approve with notes (implementation is correct, reviewer is too strict):

spec-kitty agent tasks move-task WP## --to approved --force \
  --note "Arbiter decision: Approved after 3 cycles. Meets acceptance criteria. Rationale: <explain>"

Option B -- Escalate to human (disagreement is substantial):

spec-kitty agent tasks move-task WP## --to blocked --force \
  --note "Escalated to human after 3 cycles. Conflict: <summarize>"

Option C -- Accept and move on (feedback is contradictory):

spec-kitty agent tasks move-task WP## --to approved --force \
  --note "Arbiter decision: Proceeding after 3 cycles with inconsistent feedback. <explain>"

Arbiter Guidelines

• Favor functional correctness over style preferences
• If all acceptance criteria are met, lean toward approval
• Look for genuine blocker bugs vs cosmetic complaints
• Consider diminishing returns of another full cycle
• Document your decision clearly in the --note

Parallel Sprint Pattern

When multiple independent WPs can execute simultaneously, dispatch them all at once instead of processing sequentially. This is the fastest path for missions with mixed dependency graphs.

Identifying Parallel Opportunities

# lanes.json shows which WPs are independent (different lanes)
cat kitty-specs/<mission>/lanes.json

Independent WPs are in separate lanes. WPs in the same lane have a dependency chain and must execute sequentially within that lane.

Parallel Dispatch

Claim all workspaces first (sequential — each modifies git state), then dispatch all agents in parallel:

# 1. Claim workspaces (must be sequential — git state mutations)
for wp in WP01 WP03 WP04 WP05 WP06; do
  spec-kitty agent action implement $wp --mission <slug> --agent <tool> --profile <profile>
done

# 2. Dispatch agents in parallel (method depends on orchestrator)
#    - CLI orchestrator: launch background processes
#    - Claude Code: use Agent tool with run_in_background=True
#    - CI/CD: parallel matrix jobs
#    - Human operator: open multiple terminals

The dispatch mechanism is orchestrator-dependent. The key constraint is that workspace claiming must be sequential, but agent execution can be fully parallel.

Completion-Driven Review Scheduling

As each implementation agent completes (notified asynchronously):

1. Check which WP reached for_review
2. Dispatch a review agent for that WP immediately
3. If the reviewed WP unblocks a dependent (e.g., WP01 approval unblocks WP02), dispatch the dependent's implementation immediately
4. Continue until all WPs are approved

This pattern sustains maximum parallelism throughout the sprint. Do NOT wait for all implementations to complete before starting reviews.

Tracking Parallel State

Maintain a status table of the current state. Update after each agent completion:

| WP   | Stage                | Agent         |
|------|----------------------|---------------|
| WP01 | Review in progress   | <reviewer>    |
| WP03 | Approved             | --            |
| WP04 | Implementation       | <implementer> |
| WP05 | Review in progress   | <reviewer>    |
| WP06 | Fix cycle 1/3        | <implementer> |
| WP02 | Blocked on WP01      | --            |

Use spec-kitty agent tasks status --mission <slug> as the authoritative source. The table above is the orchestrator's working copy for scheduling decisions between status checks.

Dependency-Aware Sequencing