[Wenbo11]

Problem

--config-override (or a config_override: block in a run-config YAML) is merged into the task config in Rollout.setup().
The host hard deadline (rollout/_deadline.py, added in #949) never sees that merge.
_computed_deadline_sec re-reads Task(cfg.task_path).config and sums the task file's agent.timeout_sec and sandbox.build_timeout_sec.
So if a run raises the agent budget with --config-override '{"agent":{"timeout_sec":...}}', the agent may legitimately run until the raised budget, but the backstop fires at the sum of the task file's budgets plus the fixed 1800 s margin.
The kill is reported as INFRA_ERROR ("transport or teardown wedged"), so the default retry policy reruns the attempt, and the rerun is killed at the same point.
This is the first item of #1134.

Change

_computed_deadline_sec passes the task config through apply_config_override(..., cfg.config_override) before reading budgets, the same call Rollout.setup() makes.
cfg.timeout still takes precedence, as it does in setup.
With no override, apply_config_override is a no-op.

Tests

tests/test_rollout_hard_deadline.py::TestDeadlineComputation::test_config_override_budgets_dominate_task_budget[agent-timeout|build-timeout] checks that the deadline grows by exactly the raised budget.
Both cases fail on main and pass here.

ruff check, ruff format --check and ty check are clean, and pytest tests/ passes (6502 passed, 69 skipped).

End-to-end repro (local Docker, oracle agent)

slow-oracle/task.md:

---
agent:
  timeout_sec: 60
verifier:
  timeout_sec: 30
sandbox:
  build_timeout_sec: 120
---

Write `done` to `/app/result.txt`.

environment/Dockerfile is FROM ubuntu:24.04 plus WORKDIR /app, oracle/solve.sh is sleep 2100; echo done > /app/result.txt, and verifier/test.sh writes reward 1 when /app/result.txt says done.

bench eval run --tasks-dir slow-oracle --agent oracle --sandbox docker \
  --config-override '{"agent":{"timeout_sec":3000}}' --retry-attempts 0

• main: [HARD-DEADLINE] slow-oracle: rollout exceeded host deadline (2010s); abandoning sandbox (60 + 30 + 120 + 1800), 0/1 with errors=1, although the rollout's config.json records timeout_sec: 3000.
• This branch: [PASS] slow-oracle (reward=1.00).

With the default retry policy, main reruns the attempt twice and each rerun is killed at the same point.

Notes

• Only the first item of Host hard deadline ignores the agent.timeout_sec override and reports the kill as INFRA_ERROR #1134 is covered. Classifying a deadline kill as timeout rather than INFRA_ERROR, and the reviewer deadline, are separate changes.
• Other places that also ignore the override, left for follow-ups:
  • bench continue (continue_run/orchestrator.py::build_rollout_config) rebuilds the RolloutConfig without the recorded config_override.
  • The resume mismatch check does not compare config_override.
  • Self-gen derives its default generated-skills root from the on-disk sandbox.workdir.

Refs #1134

---

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.

[bingran-you]

Verified against current main: both new cases fail on main and pass here; ruff, ty and the full suite pass, and release/0.8 carries the same line. Fresh CI test and fixture-scenarios pass; parity and pip-audit fail only on main's existing openclaw manifest drift and dnspython advisory.