fix(stella-core,stella-pipeline): close the sub-agent bracket on cancellation (#1954)

[macanderson]

What

A caller that drops a sub-agent future mid-flight — a latency ceiling, a hard cancel — left Started open forever. Every ceiling-bearing caller therefore had to forge its own Finished, and could only guess steps: 0, because the committed-call count lived inside the dropped turn. The pre-plan research stage (#1778, merged as 1bdf2da) was the first pipeline caller to hit this and forged exactly such a bracket.

CancelBracket moves the obligation into the primitive (crates/stella-core/src/subagent.rs) — design 2 of the two the issue proposed, so the next caller with a ceiling inherits the fix instead of repeating the bug. This is the same argument that moved the goal verifier onto Engine::run_sub_agent.

Drop order does the sequencing, which is what makes the numbers honest: the in-flight turn future is declared after the guard, so it drops first — the engine's own cancel guard has already emitted the abandoned call's UsageIncomplete { Cancelled } envelope and SettleChildOnDrop has already folded the money back by the time the bracket closes. The bracket therefore reports only what was committed; the in-flight call's usage rides its own envelope and is never guessed at here.

The two committed numbers travel as one CommittedTally type rather than two loose Arcs — a bracket reporting a step count without the cost that produced it would be half an answer, and bundling them also keeps run_child_turn under the argument cap without an #[allow].

research_stage.rs drops its forged bracket accordingly.

Witness

Both fail on the old code and pass with the change — verified by checking out origin/main's copies of the two production files and re-running:

• subagent::tests::a_cancelled_child_closes_its_bracket_with_committed_steps_and_cost — on old code the event stream contains Started with no Finished at all. (The dump also shows UsageIncomplete { Cancelled } already present on old code, which is the evidence that only the bracket needed owning.)
• pipeline::tests::research::a_child_past_the_ceiling_closes_its_bracket_with_committed_steps — the issue's verbatim scenario: balanced bracket, UsageIncomplete with reason cancelled, and Finished.steps equal to the committed StepUsage count. Old code fails on "Finished.steps is the committed StepUsage count, not a forged zero".

No protocol change was needed — UsageIncompleteReason::Cancelled already exists, so there is no wire-schema regeneration.

Verification

main is currently red (see #1971), so these were run on a local merge of this branch with unbreak-main-pipeline-tests:

• cargo test -p stella-core -p stella-pipeline — 0 failures
• cargo clippy -p stella-core -p stella-pipeline --all-targets -- -D warnings — clean
• cargo fmt --all -- --check — clean
• scripts/check-file-size.sh on this branch alone — OK, none grew

This PR's CI will stay red until #1971 lands, for reasons that have nothing to do with this diff.

Closes #1954

Summary by Sourcery

Ensure sub-agent cancellations close their Started/Finished bracket with accurate committed steps and cost, and rely on the core primitive rather than callers to emit synthetic finishes.

Bug Fixes:

• Preserve a balanced Started/Finished sub-agent bracket on mid-flight cancellation, reporting the true committed step count and cost instead of leaving it open or forging zeros.

Enhancements:

• Track committed sub-agent step count and cost via a shared tally that survives cancellation, and introduce a cancel guard to emit the final Finished event when the sub-agent future is dropped.

Tests:

• Add core and pipeline tests that deterministically cancel hanging sub-agents to verify brackets close correctly, usage incomplete events are emitted, and budgets settle the committed spend.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
stella-cli-docs	Ignored	Preview	Aug 7, 2026 2:24am

[sourcery-ai]

Sorry @macanderson, you have reached your weekly rate limit of 500000 diff characters.

Please try again later or upgrade to continue using Sourcery

[sourcery-ai]

Reviewer's Guide

Adds a cancellation-aware sub-agent completion mechanism in stella-core so that dropped sub-agent futures still emit a truthful Finished bracket with committed steps and cost, replaces the pipeline’s locally-forged Finished on timeout with this primitive behavior, and adds focused tests plus helpers to witness the fix in both core and pipeline research stages.

Sequence diagram for sub-agent cancellation with CancelBracket

sequenceDiagram
    actor Pipeline
    participant TokioTimeout as tokio_time_timeout
    participant Engine as Engine
    participant CancelBracket as CancelBracket
    participant CommittedTally as CommittedTally
    participant Events as EventSender

    Pipeline->>TokioTimeout: timeout(ceiling, engine.run_sub_agent_with_sender(...))

    alt normal_completion
        TokioTimeout->>Engine: run_sub_agent_with_sender(host, spec, carve, budget, events)
        Engine->>CommittedTally: new CommittedTally::default()
        Engine->>CancelBracket: create CancelBracket{ events, agent_id, tally, armed=true }
        Engine->>Engine: run_child_turn(host, spec, carve, budget, events, &tally)
        Engine->>Events: send(AgentEvent::SubAgent Finished{ steps: tally.steps(), cost_usd: tally.cost_usd() })
        Engine->>CancelBracket: armed = false
        TokioTimeout-->>Pipeline: Ok(SubAgentOutcome)
    else timeout_or_cancellation
        TokioTimeout-->>Pipeline: Err(_elapsed)
        Note over Engine,CancelBracket: caller drops sub-agent future
        CancelBracket-->>Events: Drop::drop() -> send(AgentEvent::SubAgent Finished{ status: Incomplete, steps: tally.steps(), cost_usd: tally.cost_usd() })
    end

Loading

File-Level Changes

Change	Details	Files
Introduce CommittedTally and CancelBracket in the sub-agent engine so cancelled children still close their Started/Finished bracket with accurate committed steps and cost.	Add CommittedTally struct to track committed step count and USD cost outside the child turn, updated from StepUsage events. Change child_sender to update CommittedTally on StepUsage instead of only counting steps, and pass a shared tally Arc into run_child_turn. Wrap run_child_turn in a CancelBracket guard in Engine::run_sub_agent, arming it between Started and normal Finished and disarming it on successful completion. Implement Drop for CancelBracket to emit a SubAgentPhase::Finished with status Incomplete, committed steps, cost, and a cancellation reason when a caller drops the sub-agent future mid-flight. Update run_child_turn to read the step count from CommittedTally instead of a local AtomicUsize, and propagate the tally through the pipeline.	crates/stella-core/src/subagent.rs
Add deterministic cancellation tests for sub-agent cancellation semantics in stella-core.	Introduce HangAfterScript provider test double that serves a scripted sequence then hangs and signals when hang is reached. Add test a_cancelled_child_closes_its_bracket_with_committed_steps_and_cost to assert a balanced Started/Finished bracket, committed steps and cost, presence of UsageIncomplete::Cancelled, and correct budget settlement when the sub-agent future is dropped.	crates/stella-core/src/subagent/tests.rs
Make the pipeline research stage rely on the primitive CancelBracket behavior instead of forging Finished on research timeouts, and add an integration-style witness test.	Add HangScript enum, HangTailProvider provider, and OneHangProvider resolver test doubles in research tests to simulate a hanging research child while allowing later scripted calls. Add test a_child_past_the_ceiling_closes_its_bracket_with_committed_steps to assert pipeline-level behavior: balanced bracket, UsageIncomplete::Cancelled, and Finished.steps equal to committed StepUsage count when a research child exceeds its latency ceiling. Simplify research_stage timeout handling to treat Err(timeout) the same as non-usable outcomes (Ok(_)), removing manual emission of a forged Finished with steps: 0 and relying on sub-agent CancelBracket to close the bracket. Adjust research_stage comments to document that the primitive now owns the cancel bracket and UsageIncomplete emission.	crates/stella-pipeline/src/pipeline/tests/research.rs crates/stella-pipeline/src/pipeline/research_stage.rs

Assessment against linked issues

Issue	Objective	Addressed	Explanation
#1954	Ensure that sub-agent futures dropped mid-call (e.g., via latency ceilings or hard cancels) produce a balanced Started/Finished bracket whose Finished carries the true committed step count and cost, without requiring callers to forge a Finished event.	✅
#1954	Ensure that when a research-stage child exceeds its latency ceiling, the pipeline still completes while producing: (a) a balanced sub-agent Started/Finished bracket with Finished.steps equal to the committed StepUsage count, and (b) a UsageIncomplete envelope with reason Cancelled for the abandoned in-flight call.	✅

Possibly linked issues

• Engine-turn sub-agents cancelled mid-call leave no UsageIncomplete envelope (research stage's latency ceiling is the first pipeline caller) #1954: PR implements design 2 from the issue: primitive-level CancelBracket and tally to fix cancelled sub-agent accounting.

---

Tips and commands

Interacting with Sourcery

• Trigger a new review: Comment @sourcery-ai review on the pull request.
• Continue discussions: Reply directly to Sourcery's review comments.
• Generate a GitHub issue from a review comment: Ask Sourcery to create an
  issue from a review comment by replying to it. You can also reply to a
  review comment with @sourcery-ai issue to create an issue from it.
• Generate a pull request title: Write @sourcery-ai anywhere in the pull
  request title to generate a title at any time. You can also comment
  @sourcery-ai title on the pull request to (re-)generate the title at any time.
• Generate a pull request summary: Write @sourcery-ai summary anywhere in
  the pull request body to generate a PR summary at any time exactly where you
  want it. You can also comment @sourcery-ai summary on the pull request to
  (re-)generate the summary at any time.
• Generate reviewer's guide: Comment @sourcery-ai guide on the pull
  request to (re-)generate the reviewer's guide at any time.
• Resolve all Sourcery comments: Comment @sourcery-ai resolve on the
  pull request to resolve all Sourcery comments. Useful if you've already
  addressed all the comments and don't want to see them anymore.
• Dismiss all Sourcery reviews: Comment @sourcery-ai dismiss on the pull
  request to dismiss all existing Sourcery reviews. Especially useful if you
  want to start fresh with a new review - don't forget to comment
  @sourcery-ai review to trigger a new review!

Customizing Your Experience

Access your dashboard to:

• Enable or disable review features such as the Sourcery-generated pull request
  summary, the reviewer's guide, and others.
• Change the review language.
• Add, remove or edit custom review instructions.
• Adjust other review settings.

Getting Help

• Contact our support team for questions or feedback.
• Visit our documentation for detailed guides and information.
• Keep in touch with the Sourcery team by following us on X/Twitter, LinkedIn or GitHub.