Let agent workers park on tasks and resume

[praveenperera]

Why

An agent worker that needs a long command (GPU run, release build, CI watch) had no good option:

• Waiting in the foreground burns its turn and context budget.
• Backgrounding the command dies with the turn in headless claude -p. A worker did this, lost two builds, and exited 0 without a report. That counted as TASK_SUCCEEDED, so anything held after it would have started on unfinished work.
• Handing it off by convention (blocked with WAITING <id>, plus an orchestrator-queued resume) needed manual bookkeeping. It used after, so the resume was cancelled whenever the waited task didn't succeed.

What changes

• waiting report outcome. homebased task report --outcome waiting --on <task> --notes-file - parks the worker. Notes are stored in homebased, so no RESUME.md goes in the repository. Validation runs in the report transaction and covers unknown, self, remote, resource-queue, and transitive-cycle targets, plus a 20-run chain limit.
• Chains and continuations. Run tasks stay immutable. A chain owns one unit of work across its runs. When the worker exits, the daemon atomically parks the chain, creates a held continuation that is released when every target ends with any outcome, and sends TASK_WAITING. Codex continuations resume the known thread with only the continuation block. Other agents start a fresh session with the original prompt plus the block. A parked chain reserves its Codex thread.
• Dependencies follow the chain. A task held after a parked worker stays held and resolves on the chain's final outcome. Cancelling any run cancels the chain's current run.
• no_report. An agent given the report trailer that exits 0 without reporting is TASK_FAILED with reason: "no_report". Task, container, and trailer-off agents are unchanged. Readers derive the reason, so it is never sent between Fleet machines.
• No Claude background work. Claude workers run with CLAUDE_CODE_DISABLE_BACKGROUND_TASKS=1. Verified on Claude Code 2.1.288: Bash and Agent lose run_in_background, and a foreground command that hits its timeout is killed instead of being moved to the background. Monitor keeps claude -p alive until it finishes, so it stays enabled.
• Explicit exit classification table in callback.rs, plus schema v2 (waiting targets, notes, after rule, task_chains, chain_runs).
• Skill and docs. Report trailer, worker.md, events.md, submit.md, errors.md, inspect.md, README, and a dashboard parked badge.

The second commit fixes three review findings: the Codex continuation thread fallback, cancel-requested runs parking, and doc wording. A separate commit fixes a pre-existing race in process_group_cleaned_when_child_leaves_descendant.

Not in v1

Waiting across machines, waiting on resource-queue jobs, persistent Claude sessions, suppressing root events for waited tasks, and dashboard chain navigation.

Testing

just fmt --check, just clippy, and just test pass: 632 lib, 87 integration, 54 fleet, plus cli smoke and message tests. New integration tests in tests/integration/waiting.rs cover parking and resume for Codex and Claude, release on failure, dependents through a chain, restart and duplicate-exit recovery, validation errors, cancellation, thread reservation, and no_report. A v1 database migration test and a v0.14 Fleet decoder test are included.

Summary by CodeRabbit

• New Features
  • Agents can pause work while long-running tasks complete, then continue in a new run with their notes and the tasks’ results.
  • Task details and lists show when work is parked, along with its continuation status.
  • Task reports now support a waiting outcome with specified tasks and notes.
• Bug Fixes
  • Task status and dependency handling now distinguish parked work from completed or failed work.
  • Missing reports from agents required to report are identified as failures.

[coderabbitai]

Navigate logical layers of code changes, visualize relationships, and explore their blast radius.

Note

Currently processing new changes in this PR. This may take a few minutes, please wait...

⚙️ Run configuration

• Configuration used: defaults
• Review profile: CHILL
• Plan: Advanced
• Run ID: c5eefa9a-d06b-4f43-9905-e01c7c55aa7c

📥 Commits

Reviewing files that changed from the base of the PR and between a65d660 and d8d88b4.

📒 Files selected for processing (49)

• .agents/skills/homebased/SKILL.md
• .agents/skills/homebased/references/errors.md
• .agents/skills/homebased/references/events.md
• .agents/skills/homebased/references/inspect.md
• .agents/skills/homebased/references/submit.md
• .agents/skills/homebased/references/worker.md
• README.md
• src/agents.rs
• src/callback.rs
• src/callback/tests.rs
• src/cli/task.rs
• src/daemon.rs
• src/daemon/actors/store.rs
• src/daemon/api.rs
• src/daemon/api/views.rs
• src/daemon/chains.rs
• src/daemon/dependencies.rs
• src/dependency.rs
• src/dispatcher_tests.rs
• src/domain.rs
• src/error.rs
• src/events.rs
• src/lib.rs
• src/report.rs
• src/store.rs
• src/store/admission.rs
• src/store/chain.rs
• src/store/dependency.rs
• src/store/events.rs
• src/store/events/inbox.rs
• src/store/events/tests.rs
• src/store/lifecycle.rs
• src/store/presentation.rs
• src/store/schema.rs
• src/store/task.rs
• src/store/tests.rs
• src/submission.rs
• src/waiting.rs
• tests/fixtures/fake-agent
• tests/fixtures/fake-opencode
• tests/integration.rs
• tests/integration/lifecycle.rs
• tests/integration/reports.rs
• tests/integration/waiting.rs
• tests/integration/workers.rs
• web/src/lib/api.ts
• web/src/lib/components/ParkedBadge.svelte
• web/src/lib/components/TaskList.svelte
• web/src/routes/tasks/[id]/+page.svelte

 ________________________________________________________
< Code review is a dish best served cold. Like a carrot. >
 --------------------------------------------------------
  \
   \   (\__/)
       (•ㅅ•)
       / 　 づ

✨ Finishing Touches

📝 Generate docstrings

• Commit to this branch
• Create a new PR

🧪 Generate unit tests (beta)

• Commit to this branch
• Create a new PR

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[greptile-apps]

Confidence Score: 5/5

[High risk] Adds task parking and continuation logic to the daemon.

The PR appears safe to merge; no actionable failure was established.

Summary

The PR adds waiting reports that park agent workers, persist held continuations, and make dependencies follow the final outcome of a run chain. It also classifies reporting agents that exit without a report as failed, updates worker guidance, and displays parked state in the dashboard.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[Agent run] -->|waiting report and exit 0| B[Parked chain]
  B --> C[Held continuation]
  D[Waited tasks end] --> C
  C --> E[Continued agent run]
  E -->|final outcome| F[Resolve tasks held after the chain]

Loading

_{Reviews (1) · Last reviewed commit: "Fix continuation thread and cancel handl..."}