Kc Autopilot

KC-Chat Autopilot Mode (Skill)

Read this skill when user explicitly requests "start autopilot/autonomous mode" or when about to call autopilot_* tools. Works with kc-chat.mdc base rules.

1. Mode overview

In autopilot mode, group agents work in automatic cycles under Leader coordination, without per-task user instructions.

• Leader: Dispatch center — manages taskboard, collects results, assigns next round
• Worker: Executes scan/fix/review tasks, reports on completion, waits for next round

2. Startup flow

Autopilot can be started via three methods:

A. User right-clicks in KC Chat group panel — plugin calls AutopilotService.updateState, writes shared memory and delivers prompt to Leader.

B. User tells Leader "start autopilot" — Leader executes initialization commands below.

C. LLM calls MCP tools directly (no extra toggle needed; plugin accepts directly):

autopilot_start({ groupId, sessionId, autoFix:false, focus:"full", maxFilesPerRound:5 })
// Returns OK/FAILED within 5s
// LLM can periodically call autopilot_status({ groupId }) to monitor
// Pause/stop: autopilot_pause / autopilot_stop

Leader receives first-round prompt and executes:

1. memory_write(key:"autopilot-status", content:"active", category:"context", scope:"group:<groupId>")
2. memory_write(key:"autopilot-taskboard", content:JSON.stringify({
     tasks: [], findings: [], completed: [], round: 0
   }), category:"task", scope:"group:<groupId>")
3. publish_event(type:"info", summary:"Autopilot started")
4. Dispatch first-round tasks to Workers (see §4 task templates)
5. Enter auto-loop (see §3)

2.1 Cooldown compatibility (with kc-chat.mdc §6.1)

During autopilot, if group enters cooldown (silentUntil not expired):

• Leader pauses next-round dispatch
• autopilot-status stays active, taskboard unchanged
• Auto-resumes after cooldown expires
• [AUTOPILOT_TICK] suppressed during cooldown (TickScheduler skips cooled groups)
• Worker follows kc-chat.mdc §2.1 silent ack for [COOLDOWN] messages

2.2 KC Flow protocol coordination (see kc-chat.mdc §4 + §10.5)

On autopilot start / each round dispatch / each Worker completion, Leader MUST sync KC Flow so users see project progress in UI instead of scrolling chat:

• On start: flow_step_create({ title:'Start autopilot', kind:'auto', owner:'all', ownerName:'Leader' })
• Each round dispatch: Create a step per Worker task (owner=Worker channelId)
• Worker starts: Immediately flow_step_update_status(seq, 'in_progress')
• Worker completes: flow_step_update_status(seq, 'done', result:{summary:'Fixed 3 issues'})
• Decision points: Create kind:'decision' step, user selects in UI

Time fields are written by mcp-server only — LLM must never pass time values. See kc-chat.mdc §4. Brevity: title ≤ 12 chars, result.summary ≤ 200 chars.

2.3 Autonomy contract (P6+)

autopilot_start now supports five optional new fields for bounded autonomy without user interruption:

autopilot_start({
  groupId, sessionId,
  autoFix:false, focus:'full', maxFilesPerRound:5,
  // ---- P6 new fields (all optional; defaults = legacy behavior) ----
  goal: 'Find and fix all P0 security vulnerabilities',
  successCriteria: ['P0 vulnerability count = 0', 'File coverage ≥ 80%'],
  budget: {
    maxRounds: 10,
    maxWallClockMs: 3600000,           // 60 min
    maxAutoFixesNoConfirm: 3,
    maxFilesPerSession: 50,
  },
  autonomyLevel: 'auto-bounded',       // 'ask'(default) / 'auto-bounded'(recommended) / 'auto'(high risk)
  degradedPolicy: 'pause',             // 'pause'(default) / 'stop' / 'notify'
})

autonomyLevel tiers

Leader per-round self-check loop (mandatory for auto-bounded/auto, recommended for ask)

After each round, before writing autopilot-taskboard:

1. successCriteria check: If all met → autopilot_stop({ reason:'goal-achieved' }) + publish_event + brief report to user
2. budget check:
   • round > maxRounds → autopilot_pause + notify user
   • (now - startedAt) > maxWallClockMs → same + summary:'budget-exhausted: wallClock'
   • autoFix count > maxAutoFixesNoConfirm → switch to autonomyLevel='ask'
3. convergence check: No new findings for 2 consecutive rounds → autopilot_stop({ reason:'converged' })
4. None triggered → proceed to next round dispatch

Watchdog (extension-side AutopilotTickScheduler) auto-fallback

Even if Leader LLM forgets self-check, extension Watchdog catches it: scans every 30s, checks taskboard.round > maxRounds / time limits, triggers action per degradedPolicy.

degradedPolicy options

• pause (default): Write autopilot-status='paused', set degraded=true, push message to Leader
• stop: Write autopilot-status='stopped' + same message
• notify: Push message only, don't change status

Key rule: In auto-bounded/auto mode, Leader must never wait_message for user approval each round. Only reply_message + wait_message on budget exhaustion / goal achieved / degraded.

3. Leader auto-loop (core)

Leader MUST follow this loop in autopilot mode, never self-stop:

LOOP:
  1. Dispatch round tasks:
     For each Worker, choose dispatch method:
     (a) group_broadcast — visible to all, auditable
     (b) send_to_session — direct 1:1 dispatch
     Recommend: group_broadcast for context, then send_to_session for specific tasks

  2. Wait for results:
     wait_message({ sessionId: leaderId })
     → Workers reply via send_to_session(messageType:"result")
     Repeat wait_message until expected members respond or timeout

  3. Analyze results:
     - memory_read("autopilot-taskboard") to get current board
     - Write new findings to findings[]
     - Move completed tasks to completed[]

  4. Decision:
     - New bugs → decide: immediate fix or backlog
     - Fix completed → assign reviewer
     - Review completed → mark closed
     - No new findings → switch scan direction or deeper inspection

  5. Update taskboard:
     memory_write(key:"autopilot-taskboard", content:updated board)
     publish_event(type:"info", summary:"Round N complete, found X issues")

  6. Check interrupt conditions:
     - memory_read("autopilot-status") → if "paused"/"stopped", exit loop
     - User message takes priority: pause auto-loop to handle user request

  GOTO LOOP

4. Task templates

Leader assigns tasks based on Worker roles:

Bug scan / issue detection

Scan project code, focusing on:
- Unhandled exceptions and errors
- Null/undefined risks
- Logic flaws and boundary conditions
- Resource leaks (unclosed connections, streams)

Requirements:
1. List findings (file path + line number + description + severity)
2. Provide fix suggestions for each
3. Use share_context to share files being checked
4. Use publish_event to record important findings
5. Reply to Leader via send_to_session when done

Code review

Review the following code changes/files: [specific files]

Check for: code quality, naming conventions, duplication, performance, security.

Requirements:
1. Give review verdict (pass/needs changes/reject)
2. List specific issues and suggestions
3. Reply to Leader via send_to_session when done

Refactoring

Analyze project code for optimization opportunities:
- Functions > 50 lines
- Duplicate code blocks
- Unreasonable dependencies
- Extractable common modules
- Performance hotspots

Requirements:
1. List suggestions (priority-ordered)
2. Assess benefit and risk of each
3. Reply to Leader via send_to_session when done

5. Worker behavior rules

Worker in autopilot mode:

1. Receive task → execute immediately, use share_context to share working state
2. Find issues → use publish_event(type:"error/info") to record
3. Complete task → send_to_session(messageType:"result", targetSessionId:leaderId) to report
4. Wait for next round → wait_message to keep channel; Leader dispatches via group_broadcast or send_to_session
5. Judgment calls:
   • Critical bugs (crash/security) → report and recommend immediate fix
   • Normal issues → report and wait for Leader decision
   • Optimization suggestions → write to shared memory, low priority

6. Shared memory schema

Taskboard format

{
  "round": 5,
  "mode": "scanning",
  "tasks": [
    { "id": "task-001", "assignee": "kc-mcp-agent-3-xxx", "type": "scan", "target": "mcp-server/src/index.ts", "status": "in_progress" }
  ],
  "findings": [
    { "id": "find-001", "severity": "high", "file": "src/index.ts", "line": 42, "description": "Unhandled exception", "suggestedFix": "...", "reportedBy": "agent-3", "status": "pending_review", "round": 3 }
  ],
  "completed": []
}

7. Stop / Pause

• User says "stop autopilot/pause" → Leader updates autopilot-status to stopped
• Leader notifies all Workers to pause, generates final summary
• Summary: files scanned, issues found, fixed, pending

8. User interruption handling

When user sends a message during autopilot:

1. Leader pauses auto-loop
2. Handles user request
3. After completion, asks whether to resume autopilot
4. If yes, resume loop; otherwise stop

9. Safety constraints

• Auto-fix off by default, requires explicit user authorization
• File deletion never auto-executed
• Core config file changes need Leader confirmation
• Max 5 files modified per round; pause if exceeded

10. [AUTOPILOT_TICK] forced action rules

During autopilot, extension-side AutopilotTickScheduler (scans every 30s) detects stuck Leader/Workers and injects [AUTOPILOT_TICK] messages into Leader's session:

1. Leader stale: Leader's status.json mtime > 90s without update
2. Round stale: autopilot-taskboard.updatedAt > 120s without progress
3. Worker timeout: Worker hasn't replied > 120s

Handling [AUTOPILOT_TICK]

Leader receiving [AUTOPILOT_TICK] is completely different from [POLL_TICK]:

• ❌ [POLL_TICK]: Silently re-call wait_message, no output
• ✅ [AUTOPILOT_TICK]: MUST act immediately, cannot ignore

Message format

[AUTOPILOT_TICK][REASON:<idle-timeout|round-stale|worker-timeout>][TICK:<n>/<max>] <action prompt>

Leader required actions (by REASON)

REASON: idle-timeout: Read autopilot-status + taskboard → dispatch next round or settle current

REASON: round-stale: Collect Worker replies → summarize completed → handle non-responders → update taskboard

REASON: worker-timeout: Retry/reassign/skip timed-out Workers → update taskboard → proceed

Degradation mechanism

• Leader has no action for 2 consecutive ticks → TickScheduler writes autopilot-status=blocked, UI shows red alert banner
• User can one-click "resume" or "stop"; Leader MUST actively restart round dispatch on resume

Leader iron rule

Seeing [AUTOPILOT_TICK] is equivalent to user banging the table saying "move!". Forbidden:

• ❌ Ignoring tick and continuing wait_message
• ❌ Only replying "tick received" without action
• ❌ Asking "what should I do?" (tick message already tells you)

Correct behavior: Immediately execute the actions specified in tick message, reply_message brief progress, then wait_message back to loop.