At Prompt Runner

at-prompt-runner

Run existing prompts using agent-team orchestration.

Use this skill when the user invokes /​run-at-prompt.

Group Syntax

• -> separates sequential phases.
• , separates prompts inside the same phase (parallel).

Examples:

• 220,221 -> 222,223 -> 224
• 220 -> 221,222,223 -> 224
• 220

Reference parse output:

[
  {"phase": 1, "prompts": [220, 221], "strategy": "parallel"},
  {"phase": 2, "prompts": [222, 223], "strategy": "parallel"},
  {"phase": 3, "prompts": [224], "strategy": "parallel"}
]

Resolve Helpers

PLUGIN_ROOT=$(jq -r '.plugins."daplug@cruzanstx"[0].installPath' ~/​.claude/​plugins/​installed_plugins.json)
AT_RUNNER="$PLUGIN_ROOT/​skills/​at-prompt-runner/​scripts/​at_runner.py"
PROMPT_MANAGER="$PLUGIN_ROOT/​skills/​prompt-manager/​scripts/​manager.py"

CLI Contract

python3 "$AT_RUNNER" parse "220,221 -> 222,223 -> 224"
python3 "$AT_RUNNER" parse "220 221 222" --auto-deps
python3 "$AT_RUNNER" validate "220,221 -> 222"
python3 "$AT_RUNNER" plan "220,221 -> 222" --model codex --worktree --json

Required Execution Flow

1. Parse Input

• If user passes explicit groups, parse directly.
• If user passes space-separated prompts with --auto-deps, infer group order.

2. Validate Prompt References

Run:

python3 "$AT_RUNNER" validate "$GROUPS" [--auto-deps]

If validation fails, stop and report missing prompt IDs.

3. Auto-Deps Mode (--auto-deps)

When --auto-deps is set:

• Read prompt contents via prompt-manager.
• Spawn at-planner (model: sonnet) to review inferred dependencies.
• Ask user to confirm proposed group syntax before execution.
• Do not execute any /​run-prompt command until confirmation is explicit.

Planner output should include:

• Proposed group syntax
• Dependency rationale per phase
• Any uncertainty/​assumptions

4. Build Plan

Run:

python3 "$AT_RUNNER" plan "$GROUPS" --model "$MODEL" [--worktree] [--loop] [--validate] [--json]

If --dry-run is set:

• Show plan and generated /​run-prompt commands only.
• Do not execute anything.

5. Execute Phase-by-Phase

For each phase:

• Launch one at-monitor task per prompt (model: haiku, run_in_background: true).
• Each monitor launches /​run-prompt ... through tmux and produces an Execution Report.
• Collect reports via TaskOutput() and triage in orchestrator.
• Monitors only report and flag; they do not choose retries, merges, or architecture changes.

Decision rules:

• All reports OK -> continue to next phase.
• Any ESCALATE -> inspect evidence and resolve before proceeding.
• Complex unresolved failure -> spawn at-fixer (model: opus).
• File overlap/​conflicts -> spawn at-merger (model: sonnet).

6. Optional Final Validation (--validate)

Spawn at-validator (model: sonnet) after all phases complete. Validator decides:

• PASS -> done
• RETRY -> rerun specific prompts
• ESCALATE -> call at-fixer with explicit log line references

Model Tiering (Mandatory)

• at-monitor: haiku
• at-planner: sonnet
• at-validator: sonnet
• at-merger: sonnet
• at-fixer: opus (only on escalation)

Do not assign sonnet/​opus to monitor role.

Monitor Contract (haiku)

Monitor agents are mechanical watchers. They do not debug or choose strategy.

Each report must include:

• log path
• exit code
• last 20 lines
• triage flags with line references

Report format requirement: ESCALATE|OK: {description} (see log lines N-N if applicable)

Flags

• --model <model>: default model for delegated /​run-prompt commands
• --auto-deps: infer phase ordering from prompt contents
• --validate: append final validator phase
• --worktree: run each prompt in isolated worktree
• --loop: enable verification loop per prompt
• --dry-run: show plan only

Orchestrator Prompt Pattern

Use explicit role-based Task calls:

Task(subagent_type: "at-monitor", model: "haiku", run_in_background: true, prompt: "...phase command...")
Task(subagent_type: "at-validator", model: "sonnet", prompt: "...review reports...")
Task(subagent_type: "at-merger", model: "sonnet", prompt: "...resolve overlap...")
Task(subagent_type: "at-fixer", model: "opus", prompt: "...complex failure details...")

Always pass log path + line ranges when escalating.