Hacktoberfest 2026: những issue maintainer đã đánh dấu cho tháng Mười, đang mở và phù hợp người mới. Xem issue Hacktoberfest

fw termlink dispatch: CLAUDE.md's check-in rule terminates headless workers at their first commit (exit 0, mandate unfinished)

Đang mở
  • Có gì không ổn?
#44 1 bình luận 0 reaction 0 người được giao Xem trên GitHub

Chưa có ai nhận issue này.

Đánh giá

Độ khó
4/5
Thời gian dự kiến
3-5 ngày
Mức phù hợp với người mới
43/100
Loại issue
Tính năng
Độ rõ ràng
Khá rõ ràng
Mức độ hoạt động
Sôi nổi
Công nghệ
cli, shell
Lĩnh vực
cli, developer-experience

Hướng nghiên cứu

Start by locating the fw termlink dispatch implementation and the shipped CLAUDE.md template; the issue does not name their file paths or tests. Read how dispatch prompts are composed and how run completion is recorded, then check the existing dispatch tests. Done means headless workers receive the narrowly scoped stop-condition guidance, the template distinguishes interactive sessions, and missing expected artifacts can be reported as incomplete without bypassing gates.

Do mô hình lập chỉ mục viết ra từ nội dung của issue.

Mô tả

Summary

fw termlink dispatch spawns a headless claude -p worker into a project whose CLAUDE.md
tells it, in §Agent Behavioral Rules:

Commit Cadence and Check-In
After every commit, briefly report what was done and ask if the user wants to continue.
Do not chain multiple commits without user interaction.

There is no user attached to a dispatched worker. A claude -p process has exactly one way
to "ask and wait": end its turn. Ending its turn ends the process. So a well-behaved worker
that makes a commit terminates itself at its first commit, with exit_code 0 and a
partially-finished mandate.

The worker is not misbehaving. It is following the governance it was given, and the
governance is what kills it.

What we measured

Dispatch r3-audit, one of nine in a three-round orchestrated sequence (task T-1102):

fw termlink dispatch --task T-1102 --name r3-audit \
  --prompt-file <composed> --project /opt/1409-sprind --timeout 7200 \
  --permission-mode acceptEdits --allowed-tools "Read Write Edit Bash Glob Grep TodoWrite"

meta.json: "status": "done", "exit_code": 0, 14:41:56Z → 14:59:25Z.

Exit 0, 17½ minutes, and the mandate was a 3-cycle audit run with a required handback
deliverable. It completed cycle 1 of 3 and wrote no handback. Its final message:

Committed. Per the commit cadence protocol, let me check in before continuing.
[…] Given the scale of the remaining mandate […] do you want me to continue autonomously
through the full 3-cycle run, or pause here so you can review/redirect?

It cited the rule by name. Cycle 1 itself was good work — four new WARNs triaged, three
closed with evidence, one correctly parked as a Sovereign question, one filed upstream. Then
it stopped, because it was told to.

Why this is easy to miss: exit_code 0 and status: done. Nothing in meta.json
distinguishes "finished the mandate" from "asked a question into the void and exited". An
orchestrator that checks exit codes — the obvious thing to check — records a successful run.
We only caught it because the deliverable file was absent.

Two earlier dispatches in the same sequence hit a related failure: both REVIEW workers exited
without the standalone handback their dispatch required, and their stop conditions had to be
reconstructed from report bodies and meta.json afterwards.

Why this belongs in the framework, not in every prompt

fw termlink dispatch knows the worker is headless — it constructs the claude -p
invocation. The project's CLAUDE.md does not, and cannot: it is one file serving both
interactive sessions and dispatched workers, and the rule is correct for the first.

So the knowledge exists at exactly the point where it can be acted on, and is not acted on.
Every dispatch made by any project carries this latent.

Asks, in order of leverage

  1. fw termlink dispatch should inject a headless-context preamble that narrowly
    suspends the check-in rule and names the stop conditions that replace it (mandate
    complete / budget / blocking gate / Sovereign decision). Narrow: it must not read as
    permission to skip gates, tick Human ACs, --force, or decide Sovereign questions. We
    did this by hand in the re-dispatch and the worker resumed correctly.

  2. Make the rule conditional in the shipped CLAUDE.md template, e.g. "…ask if the user
    wants to continue. This applies to interactive sessions only; a dispatched or headless
    worker continues to its stop condition and records what it would have asked as a Sovereign
    question in its handback.    "     Fixes the source rather than patching each dispatch.

  3. Distinguish "mandate complete" from "exited early" in meta.json. exit_code 0 +
    status: done currently covers both. A --expect-artifact <path> flag that marks the run
    incomplete when the deliverable is missing would have caught all three instances here
    automatically, instead of relying on an orchestrator noticing an absent file.

Environment

  • Framework 1.6.769, shared-tooling mode, tmux backend
  • 9 dispatches across 3 rounds; 3 affected (1 hard stop at cycle 1/3, 2 missing handbacks)
Ngôn ngữ chính
Shell
Star
16
Fork
6
Merge trung bình
6 giờ 24 phút
Pull request đã merge (30 ngày)
5

Chuẩn bị môi trường

Bắt đầu từ đâu

  1. Đọc hết issue, rồi đọc hướng dẫn đóng góp của dự án.
  2. Bình luận trên issue rằng bạn sẽ nhận — tránh hai người làm cùng một việc.
  3. Fork repository và làm thay đổi trên một nhánh.
  4. Mở pull request có tham chiếu số hiệu của issue.

Issue khác của DimitriGeelen/agentic-engineering-framework

Tất cả issue của DimitriGeelen/agentic-engineering-framework

Issue tương tự

Thêm issue về Shell/Bash

Nhận issue mới trong hộp thư của bạn

Bản tóm tắt ngắn những issue GitHub phù hợp với người mới.