kiro

name	kiro
description	Orchestrate Kiro CLI coding sessions via ACP. Spawns Kiro for coding tasks, monitors progress, creates branches and PRs when done. Handles both chat-initiated and Linear-spec tasks. Use when asked to code, build, or implement features.

Kiro Agent — Automated Coding via Kiro CLI

You (OpenCode) are the orchestrator. Kiro is the coder. You spawn Kiro, give it tasks, monitor its work, and handle the git/PR workflow when it's done.

Everything is remote. Kiro reads and writes code via the GitHub API — no local file changes. Multiple sessions can run simultaneously on different branches without conflicts.

When to Use This Skill

• User asks to "build", "implement", "code", "create", or "fix" something or /kiro
• A Linear task is ready for development
• A previous deploy failed and the fix needs to be coded

Architecture

OpenCode (you) ──ACP──> Kiro CLI (codes via GitHub API)
                              │
                              ├── reads files via github-dev scripts
                              ├── writes files via github-dev scripts
                              ├── commits via github-dev scripts
                              │
OpenCode (you) ──gh CLI──> GitHub (branch, PR, labels)

Paths & Constants

KIRO_ACP_CLIENT = "/Users/kokayi/Dev/claude-agent-workflow/.opencode/skills/kiro-acp/client.py"
GITHUB_DEV      = "/Users/kokayi/Dev/claude-agent-workflow/.opencode/skills/github-dev/dev"
GITHUB_DEV_DIR  = "/Users/kokayi/Dev/claude-agent-workflow/.opencode/skills/github-dev"
LINEAR_HELPER   = "/Users/kokayi/Dev/claude-agent-workflow/.opencode/skills/kiro/linear_helper.py"
KIRO_AGENT      = "/Users/kokayi/Dev/claude-agent-workflow/.opencode/skills/kiro/kiro_agent.py"

DEFAULT_CWD     = "/Users/kokayi/Dev/consuelo_on_call_coaching"
DEFAULT_REPO    = "kokayicobb/consuelo_on_call_coaching"

Full Workflow

Step 1: Receive Task

Two entry points:

Chat entry — user says "build X" or "implement feature Y":

• Extract the coding task from the conversation
• Construct a detailed prompt for Kiro (more detail = better results)

Linear spec entry — task ID provided (e.g., CON-456):

• Fetch the spec: python3 {GITHUB_DEV_DIR}/fetch_task.py CON-456
• Read the saved spec file: {GITHUB_DEV_DIR}/tasks/CON-456.md
• The spec contains requirements, acceptance criteria, and context

Step 2: Create Branch

Start a github-dev session to create the working branch:

# For Linear tasks:
python3 {GITHUB_DEV} task CON-456

# For chat tasks (no Linear):
python3 {GITHUB_DEV} create-branch "add user authentication flow"

This creates a remote branch on GitHub and saves session state. Note the branch name from the output.

Step 3: Spawn Kiro ACP Session

Use the orchestrator to spawn and manage Kiro:

import sys
sys.path.insert(0, "/Users/kokayi/Dev/claude-agent-workflow/.opencode/skills/kiro")
from kiro_agent import KiroSession

# Start a session
session = KiroSession(
    cwd="/Users/kokayi/Dev/consuelo_on_call_coaching",
    repo="kokayicobb/consuelo_on_call_coaching",
)
session.start()

# Send the task
result = session.run_task(
    task_description="Implement user authentication with JWT tokens",
    branch_name="remote/consuelo_on_call_coaching-a3f2--add-user-auth",
    linear_spec=None,  # or the spec text if from Linear
    linear_task_id=None,  # or "CON-456" if from Linear
)

# result contains: what Kiro did, whether it created specs or coded, files changed

Step 4: What Kiro Does (You Don't Control This)

Kiro receives a prompt that instructs it to:

1. Assess complexity — is this a straightforward coding task or does it need specs?
2. If complex — create Linear issues tagged kiro for each piece of work, then STOP
3. If straightforward — code the solution using github-dev scripts:
   • Read files: python3 {GITHUB_DEV} read src/file.tsx
   • Explore: python3 {GITHUB_DEV} ls src/ and python3 {GITHUB_DEV} tree src/
   • Write files: echo 'content' | python3 {GITHUB_DEV} write src/new-file.tsx
   • Or: python3 {GITHUB_DEV} write src/file.tsx --content-file /tmp/edit.tsx
   • Commit: python3 {GITHUB_DEV} commit "feat: add user auth"
   • Can do multiple write+commit cycles

Step 5: Monitor Progress

While Kiro works, you monitor via streaming. The orchestrator (kiro_agent.py) handles this, but key things to watch for:

• ToolCall events with "linear" mentions — Kiro is creating specs → session will end without coding
• ToolCall events with file operations — Kiro is coding, relay significant updates to chat
• Kiro saying "created specs" or "created issues" — end session, report specs
• TurnEnd — Kiro's turn is complete, evaluate

Relay significant progress to Ko's chat:

• "Kiro is reading the codebase structure..."
• "Kiro is writing src/components/Auth.tsx..."
• "Kiro committed 3 files: feat: add auth component"
• "Kiro created Linear specs: CON-457, CON-458. Session ending."

Step 6: After Kiro Finishes

If Kiro created Linear specs (complex task):

• Report the created specs to chat
• Tag the Linear issues with kiro label
• END — no PR, no coding. The specs will be picked up later.

If Kiro coded (straightforward task):

1. Check what was committed (the github-dev session tracks this)
2. Review the changes:

   # Check what's on the branch
   python3 {GITHUB_DEV} status
   

3. If satisfied, create PR:

   python3 {GITHUB_DEV} create-pr \
     --title "feat: add user authentication" \
     --label kiro \
     --label kiro-small
   

4. Determine PR size for labeling:
   • kiro-small: <10 files changed AND <500 total lines changed AND single concern
   • kiro-big: everything else
5. If Linear task exists, the create_pr() method auto-updates Linear status to "In Review"

Step 7: Report Results

Tell Ko:

• PR URL and number
• What was changed (brief summary)
• Whether it was labeled kiro-small (auto-review) or kiro-big (manual review)
• Linear task status update (if applicable)

Kiro Prompt Template

When sending the coding task to Kiro, use this template:

You are coding a feature for the Consuelo project (on-call coaching platform).

## Task
{task_description}

{linear_spec_section if applicable}

## Working Branch
You are working on branch: `{branch_name}`

## How to Read/Write Code

You MUST use the github-dev CLI for ALL file operations. Do NOT use your built-in
file tools — all code changes must go through the GitHub API.

### Read files
python3 {GITHUB_DEV} read <path>

### List directories
python3 {GITHUB_DEV} ls <path>

### View file tree
python3 {GITHUB_DEV} tree <optional_prefix>

### Write/edit a file
Write content to a temp file first, then stage it:
cat > /tmp/kiro-edit.tsx << 'KIROEOF'
... your code here ...
KIROEOF
python3 {GITHUB_DEV} write src/path/to/file.tsx --content-file /tmp/kiro-edit.tsx

### Commit your changes
python3 {GITHUB_DEV} commit "type: description of changes"

### Check status
python3 {GITHUB_DEV} status
python3 {GITHUB_DEV} staged

## Complexity Decision

If this task is complex (needs multiple PRs, architectural decisions, or a spec):
1. Create Linear issues for each piece of work using:
   python3 {LINEAR_HELPER} create --title "Issue title" --description "Full description" --label kiro
2. Report what specs you created
3. STOP — do not code anything

If this task is straightforward:
1. Read the relevant code to understand the codebase
2. Plan your changes
3. Write the code using github-dev commands
4. Commit with descriptive messages
5. Report what you changed and why

## Rules
- Use conventional commit messages (feat:, fix:, refactor:, etc.)
- Write clean, production-ready code
- Add comments for complex logic
- Do NOT delete files unless the task specifically requires it
- If you're unsure about something, implement the most reasonable approach
- When done, list all files you changed/created

Error Handling

• Kiro ACP process dies: Report error, clean up session state
• github-dev operations fail: Report the error — Kiro can retry or work around it
• Branch creation fails: Report error, do NOT spawn Kiro
• Timeout (4 hours): Graceful shutdown, report partial progress, save state
• Deploy failure (retry from pr-watcher): Same flow but prompt includes error context:

  The previous deploy failed with this error:
  {error_details}
  
  Fix the issue on branch `{branch_name}` and commit the fix.
  

Branch Naming

• From Linear task: Uses RemoteDev convention: remote/{repo_name}-{hash}--{description}
• From chat: kiro/{kebab-description} (we use create-branch which generates remote/ prefix)

PR Labeling

Always apply these labels:

• kiro — marks this PR as part of the automated Kiro pipeline
• kiro-small OR kiro-big — determines review automation behavior

If a Linear task is involved:

• Link the task in the PR body
• Update task status to "In Review"
• Add kiro label to the Linear task

The Specs Watcher — Autonomous Kiro Pipeline

Kiro can run autonomously without OpenCode. The specs-watcher polls Linear and spawns fresh Kiro sessions.

Files

.opencode/skills/kiro/
  specs-watcher.sh     # Main polling script
  config.sh            # Configuration (API keys, state IDs, paths)
  linear-api-kiro.sh   # Linear API wrapper
  kiro_agent.py        # Kiro session orchestrator

How It Works

specs-watcher.sh (cron/launchd)
    │
    ├── polls linear every 30 min for issues with "kiro" label in "Open" state
    │
    ├── for each issue:
    │   ├── creates remote branch via github-dev
    │   ├── spawns FRESH kiro ACP session (no token bloat)
    │   ├── kiro codes via GitHub API
    │   ├── kiro commits
    │   ├── creates PR with "kiro" label
    │   ├── moves Linear issue to "In Review"
    │   └── sends Slack notification
    │
    └── pr-watcher.ts plugin (separate) picks up the PR for review/deploy

One task = One Kiro session = Fresh context

Running the Specs Watcher

# Manual (run once)
.opencode/skills/kiro/specs-watcher.sh

# Setup Linear API first
.opencode/skills/kiro/specs-watcher.sh --setup

# Continuous mode (polls every 30 min)
.opencode/skills/kiro/specs-watcher.sh --daemon

# Via cron (every 30 minutes)
*/30 * * * * /Users/kokayi/Dev/claude-agent-workflow/.opencode/skills/kiro/specs-watcher.sh

# Via launchd (macOS)
# Create ~/Library/LaunchAgents/com.kiro.specs-watcher.plist

Configuration

Set environment variables in ~/.zshrc:

export LINEAR_API_KEY="..."
export LINEAR_TEAM_ID="..."
export SLACK_WEBHOOK_URL="..."

Or edit config.sh directly for:

• LINEAR_LABEL_NAME — default "kiro"
• POLL_INTERVAL — default 1800 seconds (30 min)
• MAX_TASKS_PER_RUN — default 0 (unlimited)

Post-Session: Update OBSERVATIONS.md (REQUIRED)

After every kiro session (chat-initiated or autonomous), the orchestrator MUST update:

.kiro/skills/kiro/OBSERVATIONS.md

Write observations about:

• Bugs — anything that broke or failed silently
• Token waste — where tokens were burned unnecessarily (large file reads, redundant searches, etc.)
• Speed — what slowed the session down
• Suggestions — concrete fixes to improve the next run

This is non-optional. The pipeline is new and needs constant refinement. Only skip if genuinely nothing went wrong (rare early on).

Important Notes

• Remote-only: Kiro codes via GitHub API. No local file changes. Period.
• Parallel sessions: Multiple Kiro sessions can run on different branches simultaneously.
• Fresh sessions: Each Linear task spawns a new Kiro ACP session — no context bloat.
• The pr-watcher plugin checks every 30 min for PRs tagged kiro and triggers automated review.
• kiro-small PRs get auto-reviewed and deployed. kiro-big PRs get a Slack notification for manual review.
• Linear specs created by Kiro will be picked up by the next specs-watcher poll — don't try to code them in the same session.