Agent Mux

agent-mux

QUICKSTART (95% of use cases)

1. Discover workers:

agent-mux config prompts        # live roster with effort defaults
agent-mux config prompts --json # structured
agent-mux config engines --json # engine capabilities + model allowlists

Selection: scout for reads, lifter for writes, architect for plans, grunt for bulk edits.

2. Profile dispatch (the 80% case):

agent-mux -P=lifter --cwd /repo "Fix retry logic in src/client.go" 2>/dev/null

Blocks until done. Result is stdout JSON with status, response, activity.files_changed.

3. Prompt-file dispatch (for heredoc quoting issues):

cat > /tmp/prompt.md << 'EOF'
Fix the `parseConfig()` function. Handle edge cases:
- Empty input returns `{}`
- Nested backticks in strings
EOF
agent-mux -P=lifter -C=/repo --prompt-file=/tmp/prompt.md

Or use --stdin JSON mode for full quoting control. --stdin implies --yes, so machine JSON dispatches do not stop for interactive confirmation by default.

4. Effort note: Don't override -e unless upgrading. Check config prompts for profile defaults -- passing -e=high to an xhigh profile (scout, auditor, grunt, architect) is a DOWNGRADE.

For opt-in GPT-5.6 depth, inspect config engines --json first. Sol and Terra may expose max/ultra; Luna stops at max; GPT-5.5/5.4-family models stop at xhigh on Codex 0.144.1. Max/Ultra require an explicit model and verified live/bundled catalog support:

agent-mux -E=codex -m=gpt-5.6-sol -e=max --cwd /repo "hard bounded task"

Ultra may trigger native Codex collaboration, but delegation is not guaranteed. --reasoning is a deprecated one-release alias; use --effort, and never pass differing explicit values.

Key flags: -P profile, -E engine, -m model, -e effort, --cwd / -C cwd, --async, --timeout. Prefer spelling --cwd in plans for clarity. Note: -E (engine) and -e (effort) are different — case matters.

5. Sync wait + result:

agent-mux -P=lifter --cwd /repo "task"  # blocks until done
# stdout is JSON: {"status":"completed","response":"...","activity":{"files_changed":2},...}

Async-first: Prefer --async for dispatches over 2 minutes — lets you monitor, steer, or bail early. Sync is fine for quick probes.

6. Async pattern:

agent-mux -P=scout --async --cwd /repo "probe" > /tmp/dispatch.json 2>/dev/null &
# ... do other work ...
wait  # shell wait: waits for the async launcher process to finish writing the ack
ID=$(jq -r .dispatch_id < /tmp/dispatch.json)
agent-mux wait --poll 30s "$ID" 2>/dev/null
RESULT=$(agent-mux result "$ID" --json 2>/dev/null)
printf '%s\n' "$RESULT" | jq -e '.status == "completed"' >/dev/null

Note: --async does NOT daemonize -- the CLI writes an ack and keeps a host process alive. Shell & is required for true parallel startup. wait --poll <duration> <id> is the lifecycle wait primitive; result <id> --json is the structured collection primitive.

DETAILS (remaining 5%)

1. Engine override:

agent-mux -P=researcher -E=gemini -m gemini-3.1-pro-preview --cwd /repo "Analyze auth module"

• Codex: implementation, debugging, precise edits. Model discovery uses the PATH-resolved CLI once per process: live catalog, bundled fallback, then conservative legacy fallback. DispatchSpec.Effort is forwarded on initial and resume invocations.
• Claude: planning, synthesis, review
• Gemini: analysis, second opinion (models: gemini-3-flash-preview, gemini-3.1-pro-preview)
• agy: experimental CLI-first model access. Plain stdout; agent-mux internally invokes the local agy CLI with its fixed --sandbox, discovers Antigravity conversation IDs from agy.log, and uses inbox + --conversation for nudge/redirect resume. Operator-supplied sandbox/permission/reasoning/max-turn/full-access options are rejected. This does not imply plugins, MCP, browser automation, Google services, or provider service actions.
• Grok: Grok Build CLI support. Default dispatch is headless prompt-file plus streaming JSON with exact UUID resume, adapter-owned prompt/session/debug files, model grok-4.5 by default, omitted native sandbox flags by default, and public artifact scanning with private Grok sidecars excluded. Opt-in v1.5 structured reports use private prompt-file transport plus grok-worker-report-schema=agent-mux.worker-report.v1; this is text/code-only, fake-covered, and live-smoke-proven on Grok CLI 0.2.93 for bounded worker reports. worker_report.json is the sanitized public report sidecar; grok_prompt.md sidecars are private. Web/X search, image input/output, structured activity, token/cost usage, custom sandbox aliases, native allow/deny, max-turns, plugins/MCP/hooks/memory/subagents, and GAAL-native indexing are not claimed. Use agent-mux config engines --json for the active Grok model allowlist; refresh explicitly with agent-mux config engines --refresh-models --engine grok --json.

Gemini ignores -e -- use model selection for depth control.

2. Steering:

agent-mux status <dispatch_id> --json 2>/dev/null
agent-mux steer <dispatch_id> redirect "Narrow scope to src/parser only; skip the broad audit." 2>/dev/null
agent-mux wait --poll 30s <dispatch_id> 2>/dev/null
agent-mux steer <dispatch_id> abort 2>/dev/null

• status <id> --json is the one-off live check. Do not poll status in a loop.
• steer <id> redirect "new direction" narrows or redirects a running worker; then wait and collect result.
• steer <id> nudge "message" asks for a wrap-up/status nudge.
• steer <id> abort SIGTERMs async host when live; otherwise writes control.json for the watchdog.

3. Data model (3 paths):

• ~/.agent-mux/dispatches/<ULID>/ -- durable (meta.json, result.json)
• $XDG_RUNTIME_DIR/agent-mux/<ULID>/ or /tmp/agent-mux-$UID/<ULID>/ -- runtime (events.jsonl, inbox.md)
• ~/.agent-mux/prompts/ -- profiles; <cwd>/.agent-mux/ -- skills/, hooks/

4. Gaal tracing: (session_id available after session start; absent for startup failures; Grok session IDs are exact UUIDs for agent-mux resume and are not GAAL-native trace IDs)

jq -r .session_id ~/.agent-mux/dispatches/<id>/meta.json
gaal inspect <session_id>

5. Bash timeout: Claude Code defaults to 120s. Long dispatches need explicit timeout: 300000 or run_in_background: true on the Bash tool call.

6. Recovery:

• timed_out + activity.files_changed non-empty -> --recover=<id> with a continuation prompt
• timed_out + empty activity.files_changed -> reframe into a narrower fresh dispatch
• failed -> check cause, retry once or escalate

agent-mux -P=lifter --recover=<id> --cwd /repo "Continue from the existing edits. Finish remaining parser tests."

7. Landmines:

• --async does NOT daemonize -- use shell &
• status: completed does not mean task success -- check response
• -e is effort, -E is engine (case matters)
• Codex Max/Ultra require -m; legacy discovery cannot prove them. Luna rejects Ultra and GPT-5.5/5.4-family models reject Max/Ultra on the Codex 0.144.1 matrix
• --reasoning is deprecated; it aliases --effort, and a differing explicit pair fails preflight
• agy rejects explicit portable sandbox, permission, reasoning, max-turn, and full-access options; use model/add-dir/prompt only
• Grok rejects explicit add-dir, permission, reasoning, full-access, max-turns, raw prompt-json/json-schema, web/X, image, native allow/deny, plugin/MCP/hook, trace/export/debug, subagent, and memory controls; use model, optional supported Grok sandbox (off, workspace, read-only, strict, devbox), prompt, timeout, and the AgentMUX-owned v1.5 report schema only
• --recover resumes with NEW prompt, does not fetch result (use result <id>)
• -e=high on xhigh profiles is a downgrade, not a floor
• wait has no --timeout — only --poll for cadence (e.g., wait --poll 30s <id>)
• There is no agent-mux config pipelines command in the current CLI. Use agent-mux config prompts and compose multi-step flows with profiles unless a named pipeline profile appears in the prompt roster.

8. References: Deep dives in references/:

• first-run.md -- first-launch/onboarding contract for cold machines
• pitfalls.md -- audit-grounded landmines with session evidence
• async-and-steering.md -- lifecycle, fan-out, steering mechanics
• recovery-guide.md -- timeout/failure handling
• gemini-specifics.md -- Gemini approval mode, resume quirks