Oa Dispatch

Dispatch Tasks to Other Agents

Use oa to delegate scoped coding tasks to a different AI model or backend. This is useful when you want to parallelize work, use a specific model's strengths, or offload a well-defined edit while you continue working.

When to dispatch

Dispatch works best for targeted, self-contained tasks with clear instructions:

• Edit a specific file or function
• Generate a test suite for a module
• Refactor a block of code to match a pattern
• Review a file for a specific concern (security, performance, etc.)

Do not dispatch vague or open-ended tasks. The receiving agent has no access to your conversation context — the prompt must be self-contained.

Basic command

Always pipe the prompt into oa instead of passing it as an argument. This keeps prompt content out of process listings (ps aux):

echo "<prompt>" | oa -b <backend> -m <model> --thinking <level>

All flags are optional. --thinking controls reasoning effort — supported levels vary by backend (e.g. low, medium, high). Omit it to use the backend's default.

For background execution with streaming visibility:

echo "Edit path/to/file.go: <specific instructions>" | oa -b claude --jsonl

For multi-line prompts, use a heredoc:

cat <<'EOF' | oa -b codex --jsonl
Edit internal/auth/handler.go: add rate limiting to the Login function.
Use x/time/rate with a limit of 5 requests per second per IP.
Do not modify any other functions. Do not ask for confirmation, just make the edits.
Run go test ./internal/auth/... when done.
EOF

Constructing effective prompts

The prompt must include everything the agent needs. Always include:

1. The file path to edit or read
2. What to do — be specific about the change
3. Constraints — what not to change, patterns to follow, tests to pass

Good prompt:

Edit internal/auth/handler.go: Add rate limiting to the Login function.
Use x/time/rate with a limit of 5 requests per second per IP.
Do not modify any other functions. Do not ask for confirmation, just make the edits.
Run go test ./internal/auth/... when done.

Bad prompt:

Add rate limiting to the auth system

Always include "Do not ask for confirmation, just make the edits" or similar. Without this, some backends may plan or brainstorm instead of making changes, especially if they have skills or settings that encourage a "plan first" workflow.

Available backends and models

Built-in backends:

To use a specific model or thinking level:

echo "prompt" | oa -b pi -m openai-codex/gpt-5.4
echo "prompt" | oa -b claude -m sonnet --thinking high
echo "prompt" | oa -b codex --thinking high

Run oa list to see all configured backends.

Running in the background

Run the dispatch as a background task so you can continue working:

# Background with JSONL streaming (can inspect progress)
echo "Edit file.go: add error handling to ProcessOrder" | oa -b codex --jsonl &

# Or with a specific model via Pi
echo "Edit file.go: add error handling to ProcessOrder" | oa -b pi -m openai-codex/gpt-5.4 --jsonl &

After the task completes, verify the changes:

git diff path/to/file.go

Timeouts: Dispatched tasks can take anywhere from 30 seconds to 30 minutes depending on complexity. When running oa from a shell tool, set a timeout of at least 10 minutes (600000ms). The default 2-minute timeout will kill most non-trivial tasks prematurely.

Parallelism rule: You can dispatch multiple tasks in parallel if they edit different files. If multiple tasks touch the same file, run them sequentially — otherwise the later task may overwrite or conflict with the earlier one.

Using threads for multi-step tasks

When a task needs follow-up work, use threads to carry context:

# First pass with Claude
echo "Refactor internal/auth/handler.go to extract middleware" | oa -b claude -t refactor-auth

# Follow-up on the same thread, different backend
echo "Now add tests for the extracted middleware" | oa -b codex -t refactor-auth

Threads are portable — switch backends mid-thread and context carries over automatically.

Output modes

Verifying results

Always verify dispatched work before proceeding:

1. Check the diff: git diff to see what changed
2. Run tests: go test ./... or equivalent
3. Read the output: Check for errors or unexpected changes

If the result is wrong, either fix it yourself or dispatch a follow-up with more specific instructions.

Multi-agent scripts

For complex dispatch patterns, use the bundled scripts:

• scripts/converse.sh — Two agents discuss a problem on a shared thread, alternating turns, then synthesize findings into an action plan.

  bash scripts/converse.sh -a claude -b codex -t 3 "the service crashes after 5 minutes under load"
  

• scripts/multi-review.sh — Dispatch parallel code reviews to different agents (security + performance), then merge into one summary.

  bash scripts/multi-review.sh
  

Both scripts use threads, pipes, and hooks — the same primitives available for any custom workflow.

Gotchas

• The dispatched agent runs in your working directory but has no access to your conversation history. Every prompt must be self-contained.
• If a backend CLI is not installed or not signed in, oa will return an error. Check with oa list first.
• --thread and --session are mutually exclusive. Use threads for cross-backend continuity, sessions for single-backend resume.
• Large prompts with full file contents may hit token limits on some models. Point the agent at the file path instead of pasting contents.