run-once

name	run-once
description	Manually run an ad-hoc debugging task through an aweek agent — simulates a heartbeat tick with the agent's .env, per-agent lock, and full execution log
trigger	aweek run, aweek debug run, aweek test agent, aweek manual run, aweek run task, aweek one-off, aweek exec agent

aweek:run-once

Manually dispatch an ad-hoc, one-shot task to an aweek agent. The task goes through the exact same execution path the hourly heartbeat uses — per-agent heartbeat lock, per-agent .env, executeSessionWithTracking with dangerouslySkipPermissions, and a full activity-log / execution-log entry — but the task itself is ephemeral: it is never written to the weekly plan and does not consume a weekly-plan slot.

Use this to debug a single agent without waiting for the next tick, or to probe an agent that is currently budget-paused (the skill deliberately force-runs through config.paused).

What this skill does NOT do

• Does not add the task to the weekly plan.
• Does not pause, resume, or top-up the agent (use /aweek:manage for lifecycle ops).
• Does not bypass the per-agent heartbeat lock — a manual run will wait for, or skip, any in-flight tick.
• Does not skip the destructive-operation confirmation gate.

Destructive operation policy

Running an ad-hoc task costs tokens (counted against the agent's weekly budget exactly like a scheduled tick) and can force-run through a pause. Per project policy, this requires an explicit AskUserQuestion confirmation before confirmed: true is set. Never pass confirmed: true without collecting the user's "yes".

Instructions

You MUST follow this exact workflow when this skill is invoked. All persistence goes through src/skills/run-once.js via the dispatcher — never spawn Claude Code CLI sessions directly.

Step 1: Pick the target agent

List all agents so the user can pick one:

echo '{"dataDir":".aweek/agents"}' \
  | aweek exec agent-helpers listAllAgents --input-json -

Project each entry to { id, name: identity?.name, paused: budget?.paused } for display. If the roster is empty, tell the user to run /aweek:hire first and stop.

Ask the user via AskUserQuestion: "Which agent should run the ad-hoc task?" — pass the agent choices (including any [paused] marker so the user knows they're forcing through).

Step 2: Collect the prompt (and optional title)

Combine the prompt and an optional short title into a single AskUserQuestion with input fields:

• Prompt (required) — the task prompt fed to the CLI session. Multiline allowed.
• Title (optional) — short label shown in the activity dashboard. Defaults to Ad-hoc debug run.

Step 3: Confirm the destructive intent

Show a preview summary:

--- Ad-hoc Run ---
Agent:     <NAME> (<ID>)[ paused → will be force-run]
Title:     <TITLE or "Ad-hoc debug run">
Prompt:    <first 200 chars of PROMPT…>
Budget:    counts toward this week's budget
Lock:      waits for/shares the per-agent heartbeat lock

Ask via AskUserQuestion: "Dispatch this ad-hoc task now? (yes/no)"

If the user says no, reply "Ad-hoc run cancelled. No session was dispatched." and stop.

Step 4: Dispatch

Only after an explicit yes, invoke the dispatcher:

echo '{
  "agentId": "<AGENT_ID>",
  "prompt": "<PROMPT>",
  "title": "<TITLE_OR_OMIT>",
  "confirmed": true
}' | aweek exec run-once execute --input-json -

Replace placeholders with JSON-escaped values. Omit title if the user didn't supply one so the default applies.

The call blocks until the Claude Code CLI session finishes. On a fresh aweek agent this may take seconds to several minutes.

Step 5: Present the result

The response JSON has this shape:

{
  "agentId": "…",
  "task": { "id": "adhoc-XXXXXXXX", "title": "…", "prompt": "…", "status": "in-progress" },
  "execResult": { /* ExecutionResult from session-executor */ },
  "activityEntry": { /* appended activity log entry */ },
  "executionLogBasename": "adhoc-XXXXXXXX_session-YYY",
  "durationMs": 12345,
  "finalStatus": "completed" | "failed",
  "error": "…"   // only present when finalStatus === 'failed'
}

Show the user, as direct assistant output (NOT just bash output):

• finalStatus with a clear indicator (completed / failed).
• durationMs formatted as seconds/minutes.
• When executionLogBasename is non-null: a clickable dashboard link of the form /executions/<agentId>/<executionLogBasename> so the user can jump straight to the rendered summary page.
• When finalStatus === 'failed': the error string verbatim, plus any stderr excerpt from execResult.sessionResult.stderr.
• Token usage from execResult.tokenUsage when present.

Error handling

• ERUN_NOT_CONFIRMED — the confirmation gate rejected the call. Restart at Step 3.
• ERUN_UNKNOWN_AGENT — the provided agentId does not exist on disk. Restart at Step 1.
• ERUN_LOCKED — another heartbeat tick is holding the per-agent lock. Tell the user to wait for the tick to complete (typically seconds to a few minutes) and retry.

Data directory

Agent state is read from .aweek/agents/<agentId>.json. The ad-hoc session reads the agent's .env from .aweek/agents/<agentId>/.env and appends an entry to .aweek/agents/<agentId>/logs/<week-monday>.json. The full NDJSON execution log is written to .aweek/agents/<agentId>/executions/<taskId>_<sessionId>.jsonl — the dashboard renders it at /executions/<agentId>/<basename>.