--exec: a sub-agent that reaches max_iterations ends the whole run instead of handing back to its parent

Open
#4,947 0 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 1 day

Nobody has claimed this yet.

Assessment

Difficulty
4/5
Estimated time
3-5 days
Newbie friendliness
55/100
Issue type
Bug
Clarity
Clearly specified
Activity status
Active
Tech stack
go
Domain
ai, cli

Research direction

Start with pkg/runtime/agent_delegation.go to trace how child sessions inherit iteration and non-interactive settings, then read enforceMaxIterations in pkg/runtime/loop_steps.go and the --exec handling in pkg/cli/runner.go. Reproduce with the supplied team configuration and non-terminal stdin; confirm that a helper capped at one iteration returns control to the root, which then completes its run.

Written by the indexing model from the issue text.

Description

area/agent area/cli area/runtime
Description

In docker agent run --exec, a sub-agent that reaches its max_iterations ends the whole run. Its lead never gets control back, so the lead cannot finish, summarise or report what the sub-agent did. The run exits 0.

Versions: v1.152.0 and v1.161.0, Linux, stdin not a terminal (CI or an unattended job).

Steps to reproduce

The provider is an OpenAI-compatible stub that answers with tool calls.

agents:
  root:
    model: m
    instruction: lead
    max_iterations: 20
    sub_agents: [helper]
  helper:
    model: m
    instruction: helper
    max_iterations: 1
    toolsets:
      - type: todo
  1. The root's first answer is transfer_task to helper.
  2. The helper answers with several create_todos calls, then text.
  3. Run docker agent run team.yaml --exec go </dev/null.

Result: one helper call, then

⚠️  Maximum iterations (1) reached. The agent may be stuck in a loop.
Do you want to continue for 10 more iterations? (y/n):
Failed to read input, exiting...

and exit 0. The root's next model call never happens.

With the helper at max_iterations: 0 (no cap), the helper finishes and the root gets transfer_task response → "…" as expected.

Why (from the source, v1.161.0)
  • Each transfer_task starts a child session with the sub-agent's own cap (pkg/runtime/agent_delegation.go, session.WithMaxIterations(childAgent.MaxIterations())).
  • In enforceMaxIterations (pkg/runtime/loop_steps.go), a NonInteractive session stops itself and appends "Execution stopped after reaching the configured max_iterations limit". For a sub-agent, that would return its turn to the lead.
  • The --exec CLI path never marks the session NonInteractive, and transfer_task copies the parent's value. So the MaxIterationsReachedEvent reaches the CLI runner (pkg/cli/runner.go). The runner prompts, the read fails without a terminal, and it calls rt.Resume(ResumeReject("")) and return nil, which ends the whole run.
Expected

In --exec with no terminal, a capped sub-agent stops and hands its turn back to its parent, as a NonInteractive session already does, instead of the CLI ending the run. Possible approaches:

  • mark --exec sessions NonInteractive when stdin is not a terminal;
  • or have the CLI, on a failed read, reject only the session that reached its cap.

--yolo avoids the prompt, but it also auto-approves every tool call, which an unattended run with a restricted permission set cannot accept.

Related: #4560 (the continuation increment).

Dominant language
Go
Stars
4.7k
Forks
565
Avg merge
18h 52m
Merged PRs (30d)
221

Getting set up

Open in Codespaces

Starts the project's dev container in your browser, under your own GitHub account.

  • Ships a Dockerfile or Docker Compose file
  • No pull request template
  • No contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from docker/docker-agent

All issues in docker/docker-agent

Similar issues

More Go issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.