prompt-executor

name	prompt-executor
description	Execute prompts from ./prompts/ directory with various AI models. Use when user asks to run a prompt, execute a task, delegate work to an AI model, run prompts in worktrees/tmux, or run prompts with verification loops.
allowed-tools	["Bash(git:*)","Bash(mkdir:*)","Bash(cp:*)","Bash(rm:*)","Bash(python3 ~/.claude/plugins/cache/daplug/daplug/*/skills/prompt-executor/scripts/executor.py:*)","Bash(codex:*)","Bash(gemini:*)","Bash(agy:*)","Bash(tmux:*)","Bash(cat:*)","Bash(pgrep:*)","Bash(REPO_ROOT=:*)","Bash(REPO_NAME=:*)","Bash(WORKTREE_PATH=:*)","Bash(WORKTREES_DIR=:*)","Bash(BRANCH_NAME=:*)","Bash(TIMESTAMP=:*)","Bash(\\:*)","Read","Edit","Write","Task","Glob","Grep"]

Prompt Executor

Auto-Approval Setup

If the user has to manually confirm the executor bash command, suggest they add this rule to ~/.claude/settings.json under permissions.allow:

"Bash(PLUGIN_ROOT=$(jq -r '.plugins.\"daplug@cruzanstx\"[0].installPath' ~/.claude/plugins/installed_plugins.json):*)"

Quick command to add it:

# Add auto-approval rule for prompt executor
jq '.permissions.allow += ["Bash(PLUGIN_ROOT=$(jq -r '"'"'.plugins.\"daplug@cruzanstx\"[0].installPath'"'"' ~/.claude/plugins/installed_plugins.json):*)"]' ~/.claude/settings.json > /tmp/settings.json && mv /tmp/settings.json ~/.claude/settings.json

---

Execute prompts from ./prompts/ (including subfolders) using various AI models (Claude, Codex, Antigravity/Gemini, ZAI, etc).

When to Use This Skill

• User says "run prompt 123" or "execute prompt 123"
• User says "run that prompt with codex/agy/gemini/zai"
• User wants to "run a prompt in a worktree"
• User wants to "run prompts in parallel"
• User asks to "delegate this to codex/agy/gemini"
• User wants to "run with verification loop" or "keep retrying until complete"
• User asks to "check loop status" for a running prompt

Executor Script

IMPORTANT: Get the executor path from Claude's installed plugins manifest:

PLUGIN_ROOT=$(jq -r '.plugins."daplug@cruzanstx"[0].installPath' ~/.claude/plugins/installed_plugins.json)
EXECUTOR="$PLUGIN_ROOT/skills/prompt-executor/scripts/executor.py"
python3 "$EXECUTOR" [prompts...] [options]

Options:

• --model, -m: claude, cc-sonnet, cc-opus, fable, codex, codex-spark, codex-high, codex-xhigh, sol, terra, luna, gpt54, gpt54-high, gpt54-xhigh, gpt55, gpt55-high, gpt55-xhigh, gpt52, gpt52-high, gpt52-xhigh, gemini, gemini-high, gemini-xhigh, gemini25pro, gemini25flash, gemini25lite, gemini3flash, gemini3pro, gemini31pro, zai, glm5, glm52, kimi, synthetic, syn-flash, syn-kimi, syn-qwen, opencode, local, qwen, devstral, glm-local, qwen-small, qwen36, qwen36-27b
  • glm52: GLM-5.2 via Z.AI / OpenCode (1M context)
  • synthetic: GLM-5.2 via Synthetic / OpenCode (syn:large:text, requires SYNTHETIC_API_KEY)

• --moa: Mixture-of-agents — comma-separated list of 2+ models (e.g. codex,synthetic,qwen36). Entries may carry a per-model CLI override as model:cli (e.g. codex:opencode). Each run gets its own worktree (implies --worktree); a manifest for the judge/consolidation phase is written to ~/.claude/loop-state/moa/. Mutually exclusive with --model and the global --cli; the bare claude Task-subagent shorthand is not allowed (use cc-sonnet, cc-opus, or claude:claude).
• --cli: Override CLI wrapper (codex, opencode, claude, agy, or gemini; aliases: claudecode, cc, antigravity). Unsupported explicit combinations fail with a clear error (no silent fallback). Not allowed with --moa.
• --variant: Reasoning variant override (none|low|medium|high|xhigh). Explicit --variant overrides alias defaults (codex-high, gpt55-high, gpt54-high, gpt52-high, etc.).
• --cwd, -c: Working directory for execution
• --run, -r: Actually run the CLI (default: just return info)
• --info-only, -i: Only return prompt info, no CLI details
• --worktree, -w: Create isolated git worktree for execution
• --sandbox: Enable sandboxing (Linux default backend: bubblewrap)
• --sandbox-type: Sandbox backend override (bubblewrap)
• --no-sandbox: Explicitly disable sandboxing
• --sandbox-profile: Isolation profile (strict|balanced|dev, default balanced)
• --sandbox-workspace: Override sandbox workspace path (default: execution cwd)
• --sandbox-net: Network override (on|off; default comes from profile)
• --base-branch, -b: Base branch for worktree (default: main)
• --on-conflict: How to handle existing worktree (error|remove|reuse|increment)
• --loop, -l: Enable iterative verification loop until completion
• --max-iterations: Max loop iterations before giving up (default: 3)
• --completion-marker: Text pattern signaling completion (default: VERIFICATION_COMPLETE)
• --require-diff: Reject completion marker when no file changes detected (created, modified, or committed) in the execution directory. Excludes TASK.md and .sisyphus/.
• --loop-status: Check status of an existing verification loop

Output: JSON with prompt content, CLI command, log path, worktree info, and loop state if enabled

Execution Flows

Direct Execution (default)

# Get executor path from installed plugins manifest
PLUGIN_ROOT=$(jq -r '.plugins."daplug@cruzanstx"[0].installPath' ~/.claude/plugins/installed_plugins.json)
EXECUTOR="$PLUGIN_ROOT/skills/prompt-executor/scripts/executor.py"

# Get prompt info
python3 "$EXECUTOR" 123 --model codex

# Force OpenCode path with reasoning variant
python3 "$EXECUTOR" 123 --model codex --cli opencode --variant high

# Folder-qualified prompt (resolves prompts/providers/011-*.md)
python3 "$EXECUTOR" providers/011 --model codex

# Run in current directory
python3 "$EXECUTOR" 123 --model codex --run

# Run in bubblewrap sandbox (Linux)
python3 "$EXECUTOR" 123 --model codex --run --sandbox

# Strict profile (no network by default)
python3 "$EXECUTOR" 123 --model codex --run --sandbox --sandbox-profile strict

# Explicit opt-out
python3 "$EXECUTOR" 123 --model codex --run --no-sandbox

With Worktree (built-in)

Single command creates worktree, copies TASK.md, and optionally runs:

# Create worktree and get info
python3 "$EXECUTOR" 123 --worktree --model codex

# Create worktree and run immediately
python3 "$EXECUTOR" 123 --worktree --model codex --run

# Use different base branch
python3 "$EXECUTOR" 123 --worktree --base-branch develop --model codex

The worktree directory is read from worktree_dir in <daplug_config> within CLAUDE.md (via config-reader), or defaults to ../worktrees/.

Mixture of Agents (--moa)

Run the same prompt with multiple models in parallel, one worktree per model, then judge and consolidate the results in the main session:

# 3 models, 3 worktrees, launched in parallel
python3 "$EXECUTOR" 123 --moa codex,synthetic,qwen36 --run

# With per-runner verification loops (state keyed as 123-moa-<label>)
python3 "$EXECUTOR" 123 --moa codex,glm5 --run --loop

# Per-entry CLI override: run the codex model through OpenCode
python3 "$EXECUTOR" 123 --moa codex:opencode,qwen36 --run

# Same model on two CLIs (distinct labels: codex, codex-opencode)
python3 "$EXECUTOR" 123 --moa codex,codex:opencode --run

Per-run info lands in prompts[].moa.runs[] (worktree, branch, log, state file, launch status); the same data is persisted as a manifest at ~/.claude/loop-state/moa/{N}-{timestamp}.json. One model failing to launch does not abort the other runs. --variant applies per model where supported and is dropped (with variant_dropped: true) where not. After all runs finish, compare diffs and test results across the worktrees, pick or synthesize a winner, and merge — see the run-prompt command's "Judge & Consolidation Phase".

With tmux (use tmux-manager skill)

1. Get CLI command from executor:

python3 "$EXECUTOR" 123 --model codex
# Returns: {"cli_command": ["codex", "exec", "--full-auto"], "content": "...", "log": "..."}

2. Create tmux session using tmux-manager patterns:

SESSION_NAME="prompt-123-$(date +%Y%m%d-%H%M%S)"
tmux new-session -d -s "$SESSION_NAME" -c "$WORKTREE_PATH"

3. Send command to session:

tmux send-keys -t "$SESSION_NAME" "codex exec --full-auto '...' 2>&1 | tee $LOG_FILE" C-m

With Verification Loop

Run prompts with automatic retries until the task is verified complete:

# Run with verification loop (background, default 3 iterations)
python3 "$EXECUTOR" 123 --model codex --run --loop

# With custom max iterations
python3 "$EXECUTOR" 123 --model codex --run --loop --max-iterations 5

# With custom completion marker
python3 "$EXECUTOR" 123 --model codex --run --loop --completion-marker "TASK_DONE"

# Worktree + loop combo
python3 "$EXECUTOR" 123 --model codex --worktree --run --loop

# Require file changes before accepting completion (--require-diff)
python3 "$EXECUTOR" 123 --model codex --run --loop --require-diff

Output includes:

{
  "execution": {
    "status": "loop_running",
    "pid": 12345,
    "loop_log": "~/.claude/cli-logs/codex-123-loop-20251229-120000.log",
    "state_file": "~/.claude/loop-state/123.json",
    "max_iterations": 3,
    "completion_marker": "VERIFICATION_COMPLETE"
  }
}

Log paths follow cli_logs_dir from <daplug_config> if configured (default ~/.claude/cli-logs/).

Completion markers (required):

• To end the loop, the model must output a final-line verification tag: <verification>VERIFICATION_COMPLETE</verification>.
• To request another iteration, output: <verification>NEEDS_RETRY: [reason]</verification>.
• The executor ignores any markers that appear inside echoed prompt instructions (some CLIs print the full prompt into logs).

--require-diff (optional):

• When enabled, the completion marker is rejected if the execution directory has no file changes (created, modified, or committed since the loop started).
• Executor-injected artifacts (TASK.md, .sisyphus/) are excluded from the diff.
• A rejected completion injects a synthetic retry reason so the model sees why it was rejected and gets another chance.
• Terminal status on the final iteration with no diff: completed_unverified (not completed).

Dead-loop detection (always on):

• Stalled: Two consecutive iterations with the same retry reason (case/whitespace-insensitive) → loop aborts with status stalled.
• Blocked: A retry reason matching impossible-gate patterns (references to paths outside the worktree, isolation-boundary refusals) → loop aborts immediately with status blocked and a suggested next step.

Check Loop Status

# Check specific prompt's loop
python3 "$EXECUTOR" 123 --loop-status

# List all active loops
python3 "$EXECUTOR" --loop-status

Model Reference

Model	CLI	Description
claude	(Task subagent)	Complex reasoning in current Claude Code context
cc-sonnet	claude --print --no-session-persistence --output-format text --input-format text --permission-mode dontAsk --model sonnet	Claude Code CLI Sonnet runs
cc-opus	claude --print --no-session-persistence --output-format text --input-format text --permission-mode dontAsk --model opus	Claude Code CLI Opus runs
fable	claude --print --no-session-persistence --output-format text --input-format text --permission-mode dontAsk --model fable	Claude Code CLI Fable 5 runs
codex	codex exec --full-auto -m gpt-5.6-terra	Fast coding execution (default Codex shorthand)
codex-spark	codex exec --full-auto -m gpt-5.3-codex-spark	Lowest-latency quick edits
codex-high	codex exec --full-auto -m gpt-5.6-terra -c model_reasoning_effort="high"	Complex coding
codex-xhigh	codex exec --full-auto -m gpt-5.6-terra -c model_reasoning_effort="xhigh"	Large refactors
sol	codex exec --full-auto -m gpt-5.6-sol	Hardest agentic coding tasks
terra	codex exec --full-auto -m gpt-5.6-terra	Balanced everyday coding (codex default)
luna	codex exec --full-auto -m gpt-5.6-luna	Fast, cost-efficient coding
gpt54	codex exec --full-auto -m gpt-5.4	Explicit GPT-5.4 shorthand
gpt54-high	codex exec --full-auto -m gpt-5.4 -c model_reasoning_effort="high"	Deep reasoning with GPT-5.4
gpt54-xhigh	codex exec --full-auto -m gpt-5.4 -c model_reasoning_effort="xhigh"	Maximum reasoning with GPT-5.4
gpt55	codex exec --full-auto	Explicit GPT-5.5 shorthand
gpt55-high	codex exec --full-auto -c model_reasoning_effort="high"	Deep reasoning with GPT-5.5
gpt55-xhigh	codex exec --full-auto -c model_reasoning_effort="xhigh"	Maximum reasoning with GPT-5.5
gpt52	codex exec --full-auto -m gpt-5.2	Planning, research, analysis
gpt52-high	codex exec --full-auto -m gpt-5.2 -c model_reasoning_effort="high"	Deep reasoning
gpt52-xhigh	codex exec --full-auto -m gpt-5.2 -c model_reasoning_effort="xhigh"	Maximum reasoning (30+ min)
gemini	gemini -y -m gemini-3-flash-preview	Fast daily driver (default)
gemini-high	gemini -y -m gemini-2.5-pro	Stable, more capable
gemini-xhigh	gemini -y -m gemini-3-pro-preview	Most capable Gemini fallback
gemini25pro	gemini -y -m gemini-2.5-pro	Explicit stable Pro selection
gemini25flash	gemini -y -m gemini-2.5-flash	Fast, cost-effective
gemini25lite	gemini -y -m gemini-2.5-flash-lite	Fastest Gemini option
gemini3flash	gemini -y -m gemini-3-flash-preview	Explicit 3 Flash selection
gemini3pro	gemini -y -m gemini-3-pro-preview	Explicit 3 Pro selection
gemini31pro	gemini -y -m gemini-3.1-pro-preview	Latest Pro model (if available)
zai	codex exec --full-auto --profile zai	General coding fallback
glm5	opencode run --format json -m zai/glm-5.2 --pure --agent build	Latest GLM 5.x tasks via OpenCode
glm52	opencode run --format json -m zai/glm-5.2 --pure --agent build	Explicit GLM-5.2 pin via OpenCode
kimi	opencode run --format json -m opencode/kimi-k2.5 --pure --agent build	Kimi K2.5 via OpenCode
synthetic	opencode run --format json -m synthetic/syn:large:text --pure --agent build	GLM-5.2 default, 512k context
syn-flash	opencode run --format json -m synthetic/syn:small:text --pure --agent build	Fast GLM-4.7-Flash fallback
syn-kimi	opencode run --format json -m synthetic/syn:large:vision --pure --agent build	Kimi-K2.6 vision tasks
syn-qwen	opencode run --format json -m synthetic/syn:small:vision --pure --agent build	Qwen3.6-27B vision tasks
opencode	opencode run --format json -m zai/glm-4.7 --pure --agent build	Recommended OpenCode JSON runner
local	opencode run --format json -m lmstudio/qwen3.6-35b-a3b --pure --agent build	Local qwen-coder model with no quota
qwen	opencode run --format json -m lmstudio/qwen3.6-35b-a3b --pure --agent build	Local qwen-coder model with no quota
devstral	opencode run --format json -m lmstudio/devstral-small-2-2512 --pure --agent build	Local Devstral model with no quota
glm-local	opencode run --format json -m lmstudio/glm-4.7-flash --pure --agent build	Local GLM-4.7 Flash model with no quota
qwen-small	opencode run --format json -m lmstudio/qwen3-4b-2507 --pure --agent build	Small/fast local Qwen model
qwen36	opencode run --format json -m lmstudio/qwen3.6-35b-a3b --pure --agent build	Best local coding model, MoE 35B, no quota
qwen36-27b	opencode run --format json -m lmstudio/qwen3.6-27b --pure --agent build	Local Qwen 3.6 dense 27B, no quota

OpenCode runs include --variant <value> when a variant is set.

GLM-5.2 uses the Z.AI Coding Plan endpoint (https://api.z.ai/api/coding/paas/v4) with raw model ID glm-5.2. OpenCode receives zai/glm-5.2; Claude Code env examples use glm-5.2[1m] with ANTHROPIC_DEFAULT_SONNET_MODEL, ANTHROPIC_DEFAULT_OPUS_MODEL, and CLAUDE_CODE_AUTO_COMPACT_WINDOW=1000000. daplug passes only the model ID; the 1M context window is provided by the Coding Plan endpoint.

Synthetic shorthands use OpenCode provider refs such as synthetic/syn:large:text; set SYNTHETIC_API_KEY and configure OpenCode's synthetic provider with https://api.synthetic.new/openai/v1. Raw hf: IDs are intentionally not daplug shorthands; add them to OpenCode as synthetic/hf:owner/model pass-through refs when needed.

OpenCode permissions (headless runs): configure ~/.config/opencode/opencode.json to avoid interactive permission prompts, e.g.:

{
  "permission": {
    "*": "allow",
    "external_directory": "allow",
    "doom_loop": "allow"
  }
}

Output Display

After executing the prompt, display a clear summary that includes the prompt title from the JSON output:

## Execution Started

**Prompt 295**: Add transcript success monitoring with retry logic

| Field | Value |
|-------|-------|
| Model | codex (gpt-5.5) |
| Status | 🟢 Running (PID 12345) |
| Loop | Max 3 iterations |

Worktree: `.worktrees/repo-prompt-295-20251229-181852/`
Branch: `prompt/295-transcript-success-monitoring`

Important: Always include the title field from the executor JSON output. This tells the user what the prompt actually does, not just its number.

Monitoring Pattern

After launching, spawn a haiku monitor subagent:

Task(
  subagent_type: "general-purpose",
  model: "haiku",
  run_in_background: true,
  prompt: """
    Monitor prompt execution:
    - Log file: {log_path}
    - PID: {pid}
    - {If tmux: Session: {session}}
    - {If worktree: Worktree: {worktree_path}}

    IMPORTANT: Use Bash tool for all file operations (not Read tool):

    Every 30 seconds, check status using Bash:
    ```bash
    # Check if process is running
    ps -p {pid} > /dev/null 2>&1 && echo "RUNNING" || echo "STOPPED"

    # Tail last 20 lines of log
    tail -20 "{log_path}"
    ```

    On completion (process ended):
    ```bash
    # Get summary from log
    tail -50 "{log_path}"

    # If worktree, show git status
    cd "{worktree_path}" && git log --oneline -5 && git diff --stat
    ```
    - Summarize what was done
    - Report final status
  """
)

Cleanup

For worktree executions, after completion:

# Remove TASK.md before merge
rm "$WORKTREE_PATH/TASK.md"

# Merge if requested
git checkout main
git merge --no-ff "$BRANCH_NAME" -m "Merge prompt: $BRANCH_NAME"

# Cleanup
git worktree remove "$WORKTREE_PATH"
git branch -D "$BRANCH_NAME"
git worktree prune