crispy-agent

name	crispy-agent
description	Unified wrapper around crispy-dispatch for multi-vendor IPC dispatch through the running Crispy host. Supports claude, codex, and other vendors. Outputs session_id for conversation continuation.
allowed-tools	Bash

Crispy Agent (unified IPC wrapper)

Execution

Works in both foreground and background:

# Foreground (blocks until turn completes)
Bash(command: "$CRISPY_AGENT Your prompt here")

# Background (non-blocking — check output later with TaskOutput)
Bash(command: "$CRISPY_AGENT Your prompt here", run_in_background: true)

Background is preferred when launching multiple agents in parallel.

---

Unified wrapper around crispy-dispatch (Crispy IPC) for multi-vendor dispatch:

• Provide prompt via arguments, PROMPT_FILE, or stdin
• Vendor selection via --vendor (default: claude)
• Model selection via --model / -m
• Session resume via --resume / -r
• Default: no timeout, bypass approvals, sessions persisted
• Sessions auto-close by default. This is almost always what you want — auto-closed sessions are persisted and fully resumable via --resume <uuid>. Pass --no-auto-close only when you need the in-memory channel kept attached (e.g. live observation, postMessage, waitForIdle). Do not pass it just to enable a follow-up turn.
• All output goes through Crispy host — sessions stream live in the UI

Prerequisites

Running Crispy host required. The script checks for ~/.crispy/ipc/servers.json or $CRISPY_SOCK.

Usage

# Arguments (default vendor: claude)
$CRISPY_AGENT Your prompt here

# Vendor selection
$CRISPY_AGENT --vendor codex "Your prompt"

# File via env var
PROMPT_FILE=task.md $CRISPY_AGENT

# Stdin
cat task.md | $CRISPY_AGENT

Session Resume

# Resume by session ID
$CRISPY_AGENT --resume <UUID> "Follow-up question"

# Resume with vendor
$CRISPY_AGENT --vendor codex --resume <UUID> "Continue"

Options

Flag	Description
--vendor <v>	Vendor to dispatch through (default: claude)
-m, --model <model>	Model override
-r, --resume <UUID>	Resume session by ID
--timeout <ms>	Override timeout (default: no timeout)
--no-auto-close	Keep the in-memory channel attached after the turn settles. Use only for live observation (postMessage, waitForIdle, watching events). Not required for --resume — auto-closed sessions are persisted and resumable.
--auto-close	Close the in-memory channel when the turn settles (this is the default — flag exists to make intent explicit). Session is still persisted to disk and fully resumable via --resume <uuid>.
-f, --fork	Fork from session (requires --resume)
--resume-at <msg-id>	Fork at specific message (requires --fork)
--no-persist	Don't save session to disk (default: persist)
--approval <mode>	Approval mode: fail, bypass (default), manual
--debug	Print diagnostics to stderr

Environment Variables

Variable	Description
PROMPT_FILE	Read prompt from this file
CRISPY_SOCK	Override socket path (skip discovery)
BYPASS_PERMISSIONS	Set to 1 for --approval bypass

Output

Returns plain text response followed by the session ID:

<response text>

[session_id: <uuid>]

Capture the session ID to continue the conversation with --resume <uuid>.

For full transcript access (turn-by-turn user/assistant content), use the readDialogue RPC:

$CRISPY_DISPATCH rpc readDialogue '{"sessionId": "<uuid>"}'

See also

crispy:live-sessions — for sessions kept alive with --no-auto-close, you can postMessage (deliver a turn), waitForIdle (block until the target settles), and readDialogue (read user↔assistant pairs) without spawning a new crispy-agent subprocess. Use the live-sessions verbs when you already have the session ID and the session is intentionally alive; use --resume when the session may have closed and needs rehydration.

Exit Codes

Code	Meaning
0	Completed successfully
10	Approval required (session paused, can resume)
11	Timeout (only if explicit --timeout set)
12	Transport error / no Crispy host
13	Invalid usage