| name | autonomous-issue-dispatch |
| description | Orchestration overview for the autonomous issue dispatch system. Handles bugs AND features from ANY entry point. References the component skills (dispatcher, handler) and per-project skills (bug-intake, qa-submission). Use when user says "dispatch", "fix these issues", "batch implement", "implement this", "build this", "develop this", "here's a plan", or any implementation/feature/fix request โ including direct conversational requests to Claude Code. |
| user-invocable | false |
Autonomous Issue Dispatch System
Goal
Claude Code receives issues (bugs OR features), diagnoses them, implements fixes/features, tests, deploys, and submits for QA - all without human intervention (except final QA approval).
This SOP covers ALL implementation work, not just dispatched issues. Whether the trigger is a Slack bug report, a GitHub issue, or a direct request in the Claude Code conversation โ the same pipeline applies. The CTO Agent is always the implementation authority for non-trivial work.
System Architecture
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ CHANNEL (Slack โ unified bug + QA per workspace) โ
โ Bug report: "Caption generation is broken" โ
โ QA response: thread reply "Approved" / "Rejected" โ
โโโโโโโโโโโโโโโโโโโโโโโฌโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ
โผ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ BUG INTAKE (per-project) โ
โ Scans channel threads for actionable items: โ
โ โ
โ New top-level messages (bug reports): โ
โ - Create GitHub issue with source reference โ
โ - Reply in thread: "Documented as #123" โ
โ - Route to โ Dispatcher โ
โ โ
โ Thread replies (QA responses): โ
โ - Rejection: Update issue โ Route directly to CTO (skip Handler)โ
โ - Approval: Close issue โ Add checkmark to thread โ
โ โ
โ Skill: .claude/skills/bug-intake-override/SKILL.md โ
โโโโโโโโโโโโโโโโโโโโโโโฌโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ
โโโโโโโโโโโโโโโดโโโโโโโโโโโโโโ
โผ โผ
NEW ISSUES QA REJECTION
โ โ
โผ โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ
โ ISSUE DISPATCHER (global)โ โ
โ - Triage queue, hygiene โ โ
โ - Prioritize actionable โ โ
โ - Invoke Handler โ โ
โโโโโโโโโโโโโฌโโโโโโโโโโโโโโโโ โ
โ โ
โผ โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ
โ ISSUE HANDLER (global) โ โ
โ - Query logs, diagnose โ โ
โ - Reproduce via test โ โ
โ - Package for CTO โ โ
โโโโโโโโโโโโโฌโโโโโโโโโโโโโโโโ โ
โ โ
โโโโโโโโโโโโโฌโโโโโโโโโโโโ
โผ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ CTO AGENT (global) โ
โ - Developer: implements fix โ
โ - QA Engineer: code review + tests โ
โ - Integration Tester: E2E validation โ
โ - Completion Gate: verify all criteria โ
โ Agent: strategic-cto-planner โ
โโโโโโโโโโโโโโโโโโโโโโโฌโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ
โผ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ QA SUBMISSION (per-project) โ
โ - Reply in bug thread: "Fix ready for QA" + test steps โ
โ - Include source reference + issue link โ
โ - Human responds in same thread (approve/reject) โ
โ Skill: .claude/skills/qa-submission/SKILL.md โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ
โผ
โโโโโโโโโโโโโโโโโโโ
โ Human reviews โ
โ in thread โ
โโโโโโโโโโฌโโโโโโโโโ
โ
โโโโโโโโโโโโโโโโโดโโโโโโโโโโโโโโโโ
โผ โผ
APPROVED REJECTED
โ โ
โผ โผ
Bug Intake Bug Intake
sees approval sees rejection
โ โ
โผ โผ
Close issue Route to CTO
Add checkmark (skip Handler)
to thread Iterate on fix
Component Responsibilities
Global Components (Uniform Across All Repos)
| Component | Purpose | Invokes |
|---|
| Issue Dispatcher | Queue triage, hygiene, prioritization | Issue Handler |
| Issue Handler | Diagnosis, reproduction, evidence | CTO Agent |
| CTO Agent | Fix implementation, testing, deployment | QA Submission |
Per-Project Components (Customized Per Repo)
| Component | Purpose | Project Defines |
|---|
| Bug Intake | Receive reports, create issues, notify | Channel ID, QA mode, repo routing |
| QA Submission | Submit for review, notify reporter | QA process, notification channel. Currently: MyProjectA only. Other workspaces use deploy-gate (YourCompany: approval-required) or auto-stage (ProjectB). |
Generalized flows:
~/.claude/skills/bug-intake/SKILL.md โ common scan/triage/fix/communicate procedure. Per-repo configs extend with ## Override: sections.~/.claude/skills/dev-loop/SKILL.md โ continuous loop (5-min intervals). Runs the full pipeline (intake โ dispatcher โ handler โ CTO โ QA) per iteration.
Workspaces
| Workspace | Slack Domain | Bug Channel | Repos |
|---|
| YourBrand | yourbrand | #cf-bugs-qa (unified) | MyProjectA |
| CompanyHQ | yourcompanyhq | #bugs | CompanyProject_NextJS, yourcompany-site, CompanyProject_AI |
| ProjectB | projectb-hq | #bugs-general | ProjectB_Bot, ProjectB_Website, projectb-schema |
CompanyProject_AI special case: When invoked locally, scans Google Sheet (manager feedback) instead of Slack. When dispatched externally, handles Slack-routed issue + optionally scans Sheet with user approval. See ~/repos/CompanyProject_AI\.claude\skills\bug-intake\SKILL.md.
Unified Pipeline
All work โ regardless of source โ flows through the same pipeline:
ANY entry point โ Triage โ [Handler if bug] โ CTO Agent (implement) โ Deploy โ QA
| Entry Point | Intake | Dispatcher | Handler | CTO Agent | Deploy | QA |
|---|
| Slack bug | Yes | Yes | Yes | Yes | Yes | Yes |
| GitHub issue | Skip | Yes | Yes | Yes | Yes | Yes |
| Sheet (CS AI) | Yes | Yes | Yes | Yes | Yes | Yes |
| Direct request | Skip | Triage only | Skip (feature) / Yes (bug) | Yes | Yes | Yes |
| QA rejection | Skip | Skip | Skip | Yes | Yes | Yes |
- Slack-sourced: Bug Intake creates GitHub issue โ enters unified queue
- GitHub-native: Already in queue โ Dispatcher picks up directly
- Sheet-sourced (CS AI): Bug Intake creates GitHub issue โ enters unified queue
- Direct request: User asks Claude Code directly ("implement this", "build X", "here's a plan"). Triage for complexity โ CTO Agent for non-trivial โ Deploy โ QA. No intake or handler needed.
- QA rejection: Routes back to CTO Agent directly (skip Handler โ diagnosis exists)
CTO Agent is always the implementation authority. Never dispatch directly to developer agent.
Trigger Mechanisms
Batch Issue Dispatch (Features + Bugs from GitHub)
When user says "dispatch these issues", "implement these", "batch fix", or similar:
- Triage the issues (easy vs big, per Issue Dispatcher skill)
- Diagnose each via Issue Handler (query logs, reproduce, evidence)
- Dispatch each to CTO Agent (
strategic-cto-planner) โ NEVER raw developer agents
- CTO Agent handles the full DoD per issue (see Orchestrator-Owns-Deploy section)
- Orchestrator tracks progress across all dispatched issues
This path skips Bug Intake (issues already exist in GitHub) but still uses Dispatcher and Handler for triage and diagnosis.
Direct Request (Conversational)
When user directly asks Claude Code to implement something ("implement this", "build X", "here's a plan", "add feature Y"):
- Triage โ Is this trivial (typo, 1-line fix, single obvious change) or non-trivial?
- Trivial (touches โค2 files, no architectural decisions): Implement directly โ Deploy โ QA
- Non-trivial (3+ files, new migration, new system/pattern, prompt changes): Continue to step 2
- Dispatch to CTO Agent (
strategic-cto-planner) โ CTO orchestrates Developer, QA Engineer, Integration Tester
- Deploy โ Run migrations, commit, push. This is NOT a TODO โ do it as part of the work.
- QA Submission โ Per project's qa-submission skill. No exceptions.
The direct request path skips Bug Intake (no channel to scan) and Handler (no bug to diagnose for features). But CTO Agent and QA are mandatory for non-trivial work.
Self-Discovered Work (Auto-Dispatch)
๐จ When Claude Code identifies a needed feature, bug, or capability gap during work โ dispatch to CTO Agent immediately in background. NEVER ask "Want me to build it?" or "Should I implement this?"
Rationale: CTO Agent runs in its own context window (minimal token cost to orchestrator). User may be AFK. Asking blocks progress on obvious engineering decisions.
Flow:
- Identify the gap (missing feature, broken behavior, missing capability)
- Create GitHub issue (via
github-issue-manager) if one doesn't exist
- Dispatch to CTO Agent (
strategic-cto-planner) via background Task โ immediately
- Continue with other work while CTO runs
- When CTO completes: audit output, deploy, submit QA
Exception: Destructive/irreversible changes (data deletion, breaking APIs, major architectural pivots) still need user confirmation.
What counts as non-trivial?
- Touches 3+ files
- New database migration
- Architectural changes (new systems, new patterns)
- Prompt/pipeline logic changes
- New UI sections or components
Bug Intake (Slack Channel Scanning)
Bug Intake can be triggered in multiple ways:
| Trigger | When | Command |
|---|
| Manual | On-demand by operator | "Scan for bug reports" / "Check QA responses" |
| Daily scheduled | Morning or night cron | Future: automated daily run |
| Webhook | Real-time on new message | Future: Slack webhook triggers scan |
Current implementation: Manual trigger only. Design is compatible with future automation.
Manual trigger commands:
"Scan for new bug reports" โ Check bug channel for new unaddressed messages
"Check QA responses" โ Check threads for QA approve/reject responses
"Run intake scan" โ Full scan (new bugs + thread responses)
Loop-Back Notifications
The reporter receives updates at key milestones:
| Milestone | Who Notifies | Where | Message |
|---|
| Bug documented | Bug Intake | Thread reply | "Created issue #123" |
| Fix in QA | QA Submission | Thread reply | "Fix ready for QA" + test steps |
| QA approved | Bug Intake | Thread reply + checkmark | "Resolved and deployed" |
| QA rejected | Bug Intake | Thread reply | Acknowledgment, then routes to CTO |
Unified thread flow (all in same bug thread):
- Reporter posts bug โ Bug Intake replies in thread with issue link
- Fix ready โ QA Submission replies in same thread with test steps
- QA reviewer responds in same thread (approve/reject)
- Bug Intake sees response:
- Approved โ close issue, add checkmark to thread
- Rejected โ route to CTO for iteration
Source reference storage: Bug Intake stores original report location (Slack thread URL) in GitHub issue metadata. This enables loop-back to the correct thread.
Invocation Patterns
Full Intake Scan (Recommended Entry Point)
"Run intake scan" / "Scan for bug reports and QA responses"
- Bug Intake scans bug channel for new unaddressed top-level messages
- Bug Intake scans threads for QA approve/reject responses
- New reports โ create issues โ Dispatcher โ Handler โ CTO
- Rejections โ route directly to CTO (skip Handler)
- Approvals โ close issues, add checkmark to thread
Manual Trigger Points
| Command | Starts At | Use When |
|---|
| "Run intake scan" | Bug Intake | Full channel scan (new bugs + QA responses) |
| "Scan for new bug reports" | Bug Intake | Only check for new top-level messages |
| "Check QA responses" | Bug Intake | Only check threads for QA responses |
| "Check pending issues" / "Find low-hanging fruit" | Dispatcher | Triage GitHub queue |
| "Diagnose issue #71" | Handler | Investigate specific issue |
| "Fix issue #71" | CTO Agent | Known diagnosis, ready to fix |
| "Implement this" / "Build X" / "Here's a plan" | Direct Request โ CTO Agent | User requests feature/fix directly |
| "Submit #71 for QA" | QA Submission | Ready for human review |
Skip Layers When Appropriate
- Direct GitHub issue: Skip Bug Intake (no Slack notification needed)
- QA rejection: Skip Handler โ straight to CTO (diagnosis exists)
- Known diagnosis: Skip Handler โ straight to CTO
- Simple fix: Handler notes "trivial" โ CTO may fast-track
Future Automation Triggers
| Trigger Type | Implementation | Status |
|---|
| Daily scheduled | Cron job runs "Run intake scan" | ๐ฎ Future |
| Webhook on Slack message | Slack app triggers scan | ๐ฎ Future |
| GitHub issue created | Webhook triggers Dispatcher | ๐ฎ Future |
Repository Readiness Checklist
For autonomous issue dispatch to work, each repo needs:
Infrastructure (Technical)
[ ] Error logging captures full context (inputs, outputs, errors)
[ ] Logs are queryable (database or structured files)
[ ] E2E tests exist for critical paths
[ ] E2E tests runnable via CLI
[ ] Database read access for log queries
[ ] CI/CD runs tests before deploy
Skills (Context Engineering)
[ ] .claude/skills/bug-intake-override/SKILL.md - workspace config (channel ID, QA mode, deploy mode, repo routing)
Global base: ~/.claude/skills/bug-intake/SKILL.md
Override example: MyProjectA/.claude/skills/bug-intake-override/SKILL.md
[ ] .claude/skills/dev-loop-override/SKILL.md - continuous scanning override (optional)
Global base: ~/.claude/skills/dev-loop/SKILL.md
Override example: MyProjectA/.claude/skills/dev-loop-override/SKILL.md
[ ] .claude/skills/qa-submission/SKILL.md - QA process (required for repos with dedicated QA channels; other repos rely on deploy-gate or approval-required mode)
GitHub Setup
[ ] Project board with status columns (Todo, Doing, QA, Done, Backlog)
[ ] Labels for categorization (bug, enhancement, security, etc.)
[ ] needs-qa label for QA tracking
Failure Modes & Recovery
Handler Blocked (Can't Diagnose)
Symptom: Missing logs, can't reproduce, no test coverage.
Recovery:
- Handler reports blocker type
- Dispatcher may ask Bug Intake to request more info from reporter
- Create infrastructure gap issue if systemic
CTO Blocked (Can't Fix)
Symptom: Requires architectural change, external dependency, unclear requirements.
Recovery:
- CTO reports blocker
- Issue moves to "blocked" status
- Dispatcher skips on future runs until unblocked
Needs Owner Decision (Dilemma Escalation)
Symptom: Epic needs breakdown, requirements ambiguous, architectural fork with no clear answer, conflicting acceptance criteria.
Recovery:
- Post in the issue's Slack thread (or open new one in the project's bug/QA channel). Tag owner with: the specific dilemma, recommended option(s) with tradeoffs, default action if no response.
- Also print the dilemma in the terminal (owner may be in session).
- Label issue
blocked, move on to other work.
- Owner responds in Slack โ next scan iteration picks it up and proceeds.
- Owner responds in terminal โ proceed immediately.
This is NOT a cop-out for complexity. Big scope with clear spec = dispatchable. Only escalate when the WHAT is genuinely unclear or requires a product decision you can't make.
QA Rejected
Symptom: Fix doesn't work, introduces regression, incomplete.
Recovery:
- Human posts rejection as thread reply in bug channel
- Bug Intake sees rejection during next scan (or webhook trigger)
- Bug Intake routes directly to CTO (skips Handler - diagnosis still valid)
- CTO iterates on fix
- QA Submission re-posts in same thread with updated test steps
- Cycle repeats until approved
Reporter Unresponsive
Symptom: Asked for more info, no response.
Recovery:
- Wait threshold (e.g., 7 days)
- Dispatcher marks as stale
- Eventually close with "unable to reproduce, please reopen with details"
Metrics (Future)
Track system health:
| Metric | Target |
|---|
| Intake โ Closed (median) | < 48 hours |
| Handler diagnosis rate | > 80% (can diagnose) |
| QA first-pass rate | > 90% (pass on first try) |
| Reporter satisfaction | Notified at all milestones |
Parallel Claude Code Instances
The user may run multiple Claude Code instances on the same repo simultaneously. This is normal and sanctioned.
Rules for the orchestrator:
- NEVER assume all unstaged changes are yours. Another instance may be working on different issues in the same working directory.
- Before reverting/discarding ANY unstaged changes: ASK the user if another instance is working. If changes look unrelated to your dispatched work, they probably belong to another instance.
- Stage only YOUR files. When committing, explicitly stage only the files your agents modified. Never
git add -A or git add . on a shared working directory.
- If you see unexpected changes (files you didn't touch, migrations you didn't create, schema changes outside your scope): treat them as belonging to another instance until the user confirms otherwise.
- Journal/config files may have entries from both instances. Don't remove entries you didn't add โ you could nuke another instance's migrations.
The safe pattern:
git add file1.ts file2.ts file3.ts # Explicit files only
git add -A # NEVER on shared workdir
git restore file.ts # Could nuke another instance's work
Orchestrator-Owns-Deploy Model
The Problem
The CTO Agent (strategic-cto-planner) handles: Developer โ QA Engineer โ Integration Tester โ QA Submission. But its Completion Gate does NOT explicitly include "commit + push to deploy." When multiple issues are dispatched in parallel, having each CTO sub-agent deploy independently causes:
- Merge conflicts from parallel commits
- Rogue scope going undetected (no diff audit)
- No batch deployment opportunity
- Risk of nuking parallel instance work
The Solution: Orchestrator = CTO Agent with Deploy Authority
When dispatching multiple issues, the orchestrator itself is a CTO Agent with a special temporary role:
- It coordinates and orchestrates all sub-CTO agents
- It is exclusively responsible for deployment (commit, push, QA submission)
- Sub-CTO agents are explicitly told: do NOT deploy โ your DoD is "ready for deployment"
Two-Tier CTO Architecture
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ ORCHESTRATOR CTO (you, temporary role) โ
โ Responsibilities: โ
โ - Triage issues (easy vs big) โ
โ - Dispatch to sub-CTO agents โ
โ - Audit agent output for rogue scope โ
โ - Batch commit + push (owns deployment) โ
โ - Submit QA for all issues โ
โ - Label issues pending-qa โ
โโโโโโโโโโโโโโโโโโโโโโโโฌโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ
โโโโโโโโโโโโโโโผโโโโโโโโโโโโโโ
โผ โผ โผ
โโโโโโโโโโโโโโโโ โโโโโโโโโโโโโโโโ โโโโโโโโโโโโโโโโ
โ Sub-CTO #119 โ โ Sub-CTO #108 โ โ Sub-CTO #126 โ
โ โ โ โ โ โ
โ Developer โ โ Developer โ โ Developer โ
โ QA Engineer โ โ QA Engineer โ โ QA Engineer โ
โ Int. Tester โ โ Int. Tester โ โ Int. Tester โ
โ โ โ โ โ โ
โ DoD: "Ready โ โ DoD: "Ready โ โ DoD: "Ready โ
โ for deploy" โ โ for deploy" โ โ for deploy" โ
โโโโโโโโโโโโโโโโ โโโโโโโโโโโโโโโโ โโโโโโโโโโโโโโโโ
Sub-CTO Agent DoD (reports back to orchestrator)
- Implement fix/feature (Developer agent)
- Write/update tests (Developer agent)
- Run test suite โ all pass
- Code review (QA Engineer)
- Integration test (Integration Tester, if UI changes)
- Report: "Ready for deployment" โ list files modified, tests passed
Orchestrator DoD (after all sub-CTOs report ready)
- Audit all agent output for rogue scope (files outside issue scope)
- Run full test suite across all changes
- Stage only in-scope files (explicit
git add, never -A)
- Commit with all issue references
- Push to deploy (trigger CI/CD)
- Submit QA to Slack for each issue (per qa-submission skill)
- Label all issues
pending-qa
- Wait for QA pass โ close issues
Standard Sub-CTO Dispatch Prompt Suffix
Include this in every sub-CTO dispatch:
DEPLOYMENT RULES:
- Do NOT commit or push. Your DoD is "ready for deployment."
- Report back: files modified, tests written, tests passed count.
- Only modify files directly related to this issue.
- Use @playwright/test for tests (NOT vitest, NOT jest).
QA RULES:
- After tests pass, invoke the project's qa-submission skill
(.claude/skills/qa-submission/SKILL.md). Do NOT post to Slack manually.
- If no qa-submission skill exists, report "QA: no qa-submission skill
found โ orchestrator must handle QA submission."
- NEVER skip QA submission. Every code change gets QA. No exceptions.
Never stop at "code is written." That's step 1 of the orchestrator's 8-step pipeline.
References
references/re-engineering-guide.md โ Change impact matrix, component contracts, extension procedures. Read before modifying any skill in this framework.
Related Skills
| Skill | Location | Purpose |
|---|
| issue-dispatcher | ~/.claude/skills/issue-dispatcher/ | Queue triage |
| issue-handler | ~/.claude/skills/issue-handler/ | Diagnosis |
| bug-intake | ~/.claude/skills/bug-intake/ (global) + .claude/skills/bug-intake/ (per-project config) | Intake + notification |
| dev-loop | ~/.claude/skills/dev-loop/ (global) + .claude/skills/dev-loop/ (per-project override) | Continuous scanning loop |
| qa-submission | .claude/skills/qa-submission/ (per-project) | QA + notification |
| strategic-cto-planner | ~/.claude/agents/ | Fix orchestration |
