Job hangs ~20-25 min with no output after a Task-delegated subagent reports completion (repository_dispatch trigger)

Describe the bug When the top-level agent delegates work to a custom project-defined subagent via Task (e.g. subagent_type: some-custom-agent), and that subagent's task reaches status: completed, the parent session appears to hang indefinitely afterward — no further JSON stream events, no final "type": "result" block, and no PR comment is posted despite the prompt instructing one. The GitHub Actions job eventually exits ~20–25 minutes later with conclusion: success but nothing was actually produced (no comment, no commit, no visible output).

This has reproduced twice in a row, in separate workflow runs. The second time, show_full_output: true was enabled to capture the raw stream, which pinpoints exactly where it stops.

A plain local_bash background task (not a Task-delegated subagent) completed and the session continued normally within ~2 seconds in the same run — so this is not "any Task-based flow hangs," it's specific to the subagent-delegation case.

To Reproduce

1. Set up a workflow triggered by repository_dispatch, running claude-code-action@v1 with claude_args: --allowedTools "Read,Grep,Glob,Bash,Edit,Write,Task".

2. Define a project .claude/agents/*.md custom subagent.

3. Prompt the top-level agent to delegate a real, multi-step piece of work to that subagent via Task (in our case: several Bash tool calls investigating repository state, taking ~11 minutes of real visible work).

4. Instruct the top-level agent (in the same prompt) to, after the subagent's work, post a comment on a target PR via gh.

5. Observe: the subagent's task_updated/task_notification (status: "completed") system events appear in the log, and then the JSON stream stops — no further assistant/user/result messages, no comment posted — until the job exits roughly 20–25 minutes later.

6. Go to '...'

7. Click on '....'

8. Scroll down to '....'

9. See error

Expected behavior The top-level session should continue after the Task-delegated subagent reports status: completed, produce a final "type": "result" message, and (per the prompt) post the PR comment via gh — completing within roughly the time the visible work actually took (a few minutes), not 20–25 minutes later with no output.

Screenshots T+~1m {"type":"system","subtype":"task_started","task_id":"bvnw44ydr","tool_use_id":"...","description":"Re-establish current HEAD SHAs for all ledger repos","is_backgrounded":false,"task_type":"local_bash", ...} T+~1m {"type":"system","subtype":"task_notification","task_id":"bvnw44ydr","tool_use_id":"...","status":"completed","output_file":"","summary":"Re-establish current HEAD SHAs for all ledger repos", ...} (session continues normally within ~2 seconds)

...~11 minutes of normal assistant/user/tool_use/tool_result messages inside a Task-delegated subagent (subagent_type: )...

T+~12m {"type":"system","subtype":"task_updated","task_id":"ad6906e5e3c847473","patch":{"status":"completed","end_time":<epoch_ms>}, ...} T+~12m {"type":"system","subtype":"task_notification","task_id":"ad6906e5e3c847473","tool_use_id":"...","status":"completed","output_file":"/tmp/claude-1001/.../tasks/ad6906e5e3c847473.output" <-- this is the LAST line in the entire log. No matching closing content for this JSON object was captured, no further stream lines at all. T+~35m (job log jumps directly to the action's own "Post job cleanup" steps — nothing in between)

Workflow yml file

name: Trigger Industrialisation

on:
  repository_dispatch:
    types: [some-custom-event]

permissions:
  contents: write

jobs:
  run-agent:
    name: Run custom orchestrator agent
    runs-on: ubuntu-latest
    timeout-minutes: 25
    steps:
      - name: Checkout
        uses: actions/checkout@v4
        with:
          token: ${{ secrets.GH_PAT }}
          fetch-depth: 0

      - name: Run agent
        uses: anthropics/claude-code-action@v1
        with:
          anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
          show_full_output: true
          github_token: ${{ secrets.GIT_PAT }}
          prompt: |
            Use the <custom-orchestrator-agent> agent.

            [... project-specific instructions describing the source
            event payload from github.event.client_payload.*, telling
            the agent to delegate analysis work to project-defined
            subagents via Task, and to post a final comment on a
            target PR via gh once done ...]
          claude_args: |
            --allowedTools "Read,Grep,Glob,Bash,Edit,Write,Task"

API Provider

[ X] Anthropic First-Party API (default) [ ] AWS Bedrock [ ] GCP Vertex

Additional context

• This is a project-specific (non-built-in) subagent defined under .claude/agents/*.md, invoked via Task with a matching subagent_type.
• Trigger type is repository_dispatch, not pull_request/issue_comment/workflow_dispatch — we have not tested whether this reproduces under those other trigger types.
• is_error: false in the one occurrence where a result block appeared, and no tool/permission errors anywhere in either log — this doesn't look like a crash, more like the process being alive but stuck (e.g. blocked reading/writing the subagent's output_file, or waiting on a stream/handle that never closes).
• We reduced our job's timeout-minutes from 90 to 25 to bound the wasted Actions runtime per occurrence — that's a mitigation for the symptom, not a fix.
• Happy to share the full (sanitized) JSON log for the second occurrence, or attempt a minimal from-scratch reproduction, if either would help narrow this down.