acp-job-stuck

Use when an ACP job is stuck at a phase. Phase table (who acts next, with which command, when to wait, nudge or close), the read-only checks that tell you where it really is, and the links to the cata

SKILL.md

GitHub

ACP job stuck — find the phase, find who owes the next move

When to use

  • A job has not moved for longer than its SLA, or the owner asks "what happened to job <job-id>?".
  • A command answered but the job status did not change.

The state machine itself is in acp-jobs. This skill is the triage: where is the job, who must act, what you may do.

Prerequisites

  • The <job-id> and the chain (Base, 8453). Commands run through the app door ~/.iclone-os/bin/acp --inft <iNFT>.
  • Which side you are on: client (you hired) or provider (you sell). job history says it in roles.

Quick reference (acp 1.0.38 has no job status, job steps or job deliver subcommand)

~/.iclone-os/bin/acp --inft <iNFT> job history --job-id <job-id> --chain-id 8453 --json
~/.iclone-os/bin/acp --inft <iNFT> job watch --job-id <job-id> --timeout 600 --json
~/.iclone-os/bin/acp --inft <iNFT> message send --job-id <job-id> --chain-id 8453 --content "<text>" --json

job history is the ground truth (status, budget, deliverable, every message). job watch blocks until the job needs YOUR action, prints the event and exits — run it as a background task, never in the foreground of a conversation.

Procedure

  1. Read job history. Take the top-level status and the last entry. Ignore job list for state (acp-troubleshooting E11).
  2. Find the row and its owner:
Last event Status Next move by Their command Stuck means
job.created open provider provider set-budget provider offline or no handler
budget.set budget_set client client fund --amount <exactly the budget> waiting for the owner, or fund failed
job.funded funded provider provider submit provider still working or crashed
job.submitted submitted evaluator (client by default) client complete / client reject nobody evaluated
job.completed / job.rejected / job.expired final nobody — not stuck: report the outcome
  1. If the next move is YOURS, do it now with the commands in acp-jobs (provider: price, deliver; client: fund, evaluate). Funding, completing and rejecting move value: ask through the consent door (key spendTrade), state amount, asset, chain and destination, and wait for the owner's yes — you operate, the owner spends (economyos law 1):

    ~/.iclone-os/bin/acp --inft <iNFT> consent request spendTrade --reason "Fund job <job-id>: <usdc> USDC on Base to escrow"
    
  2. If the next move is THEIRS, nudge once with message send. It speaks for the owner: ask first through the consent door (key sendMessages) and wait for the owner's yes. Say the job id, the phase and what you need:

    ~/.iclone-os/bin/acp --inft <iNFT> consent request sendMessages --reason "Nudge the counterparty on job <job-id>"
    

    Then start job watch --timeout 600 in the background and stop polling by hand.

  3. Past the SLA with no answer: tell the owner the phase, the time waited and the options — keep waiting, or client reject with a reason (the CLI help reads "Reject a job or deliverable (as evaluator)"). Never reject on your own.

  4. Your fund failed with transfer amount exceeds balance: acp-troubleshooting E20, then acp-wallet-ops for the top-up.

  5. You are the provider and never saw job.funded: check that exactly one listener runs for this agent (acp-troubleshooting E17) and that it did not crash-loop (E6).

Pitfalls

  • message send defaults to chain 84532 (Base Sepolia). Always pass --chain-id 8453, or the message goes nowhere.
  • The events listen help text points to a job status command that does not exist in 1.0.38. Use job history.
  • A deliverable that is empty or generic is a provider fault, not a stuck job (acp-troubleshooting E12, E19).
  • Signer approval pending forever: the URL expired (acp-troubleshooting E16) — that blocks every write, not one job.
  • The door refuses a command it does not allow-list ('<root> <sub>' is not a command the agent may run). Report the refusal verbatim to the owner; never run the CLI another way.

Never

  • Never fund, complete or reject without the owner's yes through the consent door (spendTrade).
  • Never send a message to the counterparty without sendMessages allowed.
  • Never re-create a job to "unstick" it: the first one may still be funded.

Verification

Run job history again: the status moved to the phase you expected, or the owner has the report (phase, owner of the next move, time waited, the choice you need from them).