Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/bassimeledath/dispatch/llms.txt

Use this file to discover all available pages before exploring further.

Dispatch provides robust error handling and recovery mechanisms. Workers can ask questions, mark blockers, and recover from failures — all without losing context.

​
Understanding Checklist Markers

The plan file uses four markers to track progress:
Meaning: Item completed successfullyExample:
- [x] Scan for hardcoded secrets — found 2 instances in config.ts
- [x] Review authentication logic — JWT implementation looks good
Workers can add notes after the marker to provide context.

​
Blocked Items and Question Flow

When a worker needs clarification, it uses the IPC (Inter-Process Communication) system to ask questions without losing context.

​
Primary Flow: IPC (Monitor-Triggered)

This is the preferred method — the worker stays alive while waiting for an answer.

​
1. Worker Asks a Question

Worker encounters a blocker and writes a question: 
# Worker writes atomically
echo "Found both JWT and session-based auth. Which should I focus on?" > \
  .dispatch/tasks/security-review/ipc/001.question.tmp
mv .dispatch/tasks/security-review/ipc/001.question.tmp \
  .dispatch/tasks/security-review/ipc/001.question
The worker then polls for .dispatch/tasks/security-review/ipc/001.answer.

​
2. Monitor Detects the Question

A lightweight monitor script polls the IPC directory: 
# Monitor checks for unanswered questions
for q in .dispatch/tasks/security-review/ipc/*.question; do
  seq=$(basename "$q" .question)
  if [ ! -f ".dispatch/tasks/security-review/ipc/${seq}.answer" ]; then
    exit 0  # Triggers <task-notification>
  fi
done
When it finds an unanswered question, it exits — triggering a notification.

​
3. Dispatcher Surfaces the Question

The dispatcher reads the question and shows it to you: 
Worker is asking: "Found both JWT and session-based auth. Which should I focus on?"

​
4. You Answer

You provide an answer: 
Focus on JWT — session-based auth is deprecated.

​
5. Dispatcher Writes the Answer

The answer is written atomically: 
echo "Focus on JWT — session-based auth is deprecated." > \
  .dispatch/tasks/security-review/ipc/001.answer.tmp
mv .dispatch/tasks/security-review/ipc/001.answer.tmp \
  .dispatch/tasks/security-review/ipc/001.answer

​
6. Worker Continues

Worker detects the answer, acknowledges it, and continues: 
# Worker acknowledges
touch .dispatch/tasks/security-review/ipc/001.done

# Continues with the task using the answer
Critical: The worker never lost context. It stayed alive the entire time, preserving all state in memory.

​
Fallback Flow: Context Dump ([?] Marker)

If no answer arrives within ~3 minutes, the worker falls back to the legacy behavior:
  1. Dumps context to .dispatch/tasks/<task-id>/context.md
  2. Marks the item [?] with the question
  3. Exits cleanly
Example plan file after timeout: 
# Security Review

- [x] Scan for hardcoded secrets — found 2 instances
- [?] Review authentication logic

> **Blocker:** Found both JWT and session-based auth. Which should I focus on?
You can answer and re-dispatch: 
/dispatch continue security-review from the blocked item. 
Answer: Focus on JWT — session-based auth is deprecated.
The new worker:
  1. Reads the plan file
  2. Reads context.md (preserves the previous worker’s state)
  3. Uses your answer to continue from the blocked item

​
Proactive Recovery When Workers Fail

When a worker fails to start or errors immediately, Dispatch proactively offers recovery.

​
Scenario: CLI Unavailable

/dispatch use gpt-5.3-codex to review the auth module
Worker fails: 
Error: codex: command not found
Dispatcher’s proactive recovery: 
The Codex CLI is no longer available. Alternatives from your config:

1. opus (Claude)
2. sonnet (Claude)  
3. gemini-3.1-pro (Cursor)

Want me to switch your default and retry with one of these?
You choose: 
Use opus
Dispatcher updates config and retries: 
# ~/.dispatch/config.yaml
default: opus  # Updated from gpt-5.3-codex

models:
  opus:   { backend: claude }
  sonnet: { backend: claude }
  # Removed: gpt-5.3-codex (CLI unavailable)

Updated default to opus. Retrying security-review...
Dispatched security-review using opus.

​
Scenario: Auth Error

/dispatch use claude-opus-4.5 to do a security review
Worker fails: 
Error: Authentication failed. Claude Opus 4.5 requires upgraded subscription.
Dispatcher’s proactive recovery: 
Claude Opus 4.5 returned an auth error. Alternatives:

1. sonnet (Claude) — works on your plan
2. gemini-3.1-pro (Cursor)

Switch to one of these and retry?
You choose: 
Sonnet
Dispatcher updates config and retries: 
# Removes the broken model to prevent future failures
models:
  sonnet: { backend: claude }
  haiku:  { backend: claude }
  # Removed: claude-opus-4.5 (auth failed)

​
Timeout Behavior

Workers and monitors have timeouts to prevent hanging indefinitely.

​
Monitor Timeout

The monitor script times out after 30 minutes: 
TIMEOUT=1800  # 30 minutes
START=$(date +%s)

while true; do
  # Check for .done or questions...
  
  ELAPSED=$(( $(date +%s) - START ))
  [ "$ELAPSED" -ge "$TIMEOUT" ] && exit 1
  
  sleep 3
done
If a worker is still running after 30 minutes without progress, the monitor exits with error.

​
IPC Answer Timeout

Workers wait ~3 minutes for an answer to their question. If no answer arrives:
  1. Worker dumps context to context.md
  2. Marks the item [?] with the question
  3. Exits cleanly
This prevents workers from hanging indefinitely when you’re away.

​
Handling Multiple Failures

When multiple parallel workers fail, Dispatch handles each independently: 
Error report:

● security-review (opus) — FAILED
  Worker exited with error on item 2:
  [!] Review authentication logic
  > Error: auth/ directory not found

● api-tests (gpt-5.3-codex) — FAILED
  Worker failed to start: codex: command not found

● update-readme (gemini) — COMPLETE
  ✓ All items checked
Dispatcher’s recovery: 
1. security-review failed due to missing directory. Fix it and I can retry.
2. api-tests failed because Codex CLI is unavailable. Switch to opus or sonnet?
3. update-readme completed successfully.
You respond: 
Retry api-tests with sonnet
Dispatcher updates and retries: 
Updated api-tests to use sonnet. Dispatching...
Dispatched api-tests using sonnet.

​
Error Recovery Best Practices

Answer Questions Quickly

Workers stay alive for ~3 minutes waiting for answers. Respond within this window to avoid context loss.

Review Context Dumps

If a worker times out, check context.md before re-dispatching. It contains the worker’s state before exit.

Fix Root Causes

If a worker fails due to missing files or broken CLI, fix the underlying issue before retrying.

Use Aliases for Fallbacks

Create aliases with fallback models so you can quickly switch when a primary model fails.
See the Configuration guide for managing model fallbacks and the Parallel Tasks guide for handling errors in concurrent workers.

Build docs developers (and LLMs) love

Get started for freeTalk to us