fix(agent-challenge): stop public log redaction from wedging the event loop

[echobt]

Summary

• A catastrophic-backtracking regex in the public task-event redaction path pinned the validator event loop at 100% CPU, starving the evaluation worker and freezing the job queue in production.
• Redaction is rewritten as a single linear pass. Redaction decisions are unchanged.

Root cause

_PUBLIC_SECRET_SEGMENT_RE nested (?:[A-Za-z0-9_.]+[-_.])* over a character class that also contained the separators _ and .. Any dotted or underscored run that never reached secret/token could therefore be split in exponentially many ways before the match failed.

Measured on the live pattern:

input	time
37 chars	23 ms
61 chars	91 s

Terminal-Bench task events carry pip output (Successfully installed agent-challenge-1.0.1 aiohappyeyeballs-2.7.1 ...), which is exactly that shape, and get_submission_task_events replays it on an unauthenticated endpoint. One log line was enough to wedge the loop.

Observed in production: py-spy showed the main thread parked in _public_task_event_text (routes.py:5108) with wchan=0, validator at 101% CPU, a job stuck running for hours with no containers and no log events.

Because the endpoint is public and unauthenticated, this was also a remote denial-of-service vector.

Changes

• Replace the nested-quantifier regex with _PUBLIC_SECRET_WORD_RE (one linear pass over delimited words) plus two bounded marker checks.
• Redaction semantics preserved: whole-segment token redacted, count-dataset-tokens left alone, secret/raw-ref/broker-ref/pod- markers redacted, surrounding text untouched.

Test plan

• New regression test: pathological input must sanitize in under 1s. Before the fix this file hung past a 10-minute timeout; it now runs in 0.11s.
• New test: realistic pip output stays fast and is not redacted.
• Existing redaction contracts green (test_public_task_event_redaction.py, test_frontend_api_contract.py, test_submission_status.py — 21 passed).
• Full package suite: 2309 passed / 24 failed, and an A/B against clean main reproduces the same 24 failures, so they are pre-existing and unrelated.
• ruff check and ruff format --check clean.
• Verified on the production validator: pathological 241-char input went from unbounded to 0.30 ms; four own-runner-task-* containers started within seconds and CPU settled at 6.6%.

CI

• Waiting for required checks to go green on this head SHA.

Notes

• The fix is already hot-patched onto the live validator to unblock the queue; this PR makes it durable.
• Out of scope: construction fail-fast currently sits in the orchestrator, which never trips under the real one-task-per-runner topology. That belongs in the validator task loop and is tracked separately.

Summary by CodeRabbit

• Bug Fixes

  • Improved redaction of secrets and tokens in public task-event text.
  • Prevented slowdowns when processing long or complex text.
  • Reduced the risk of incorrectly redacting legitimate package and version information.

• Tests

  • Added coverage for fast processing of large and realistic command output.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 69e34c4f-690d-41ab-bf95-f2f92315f6ed

📥 Commits

Reviewing files that changed from the base of the PR and between 7b076d3 and 32ca672.

📒 Files selected for processing (2)

• packages/challenges/agent-challenge/src/agent_challenge/api/routes.py
• packages/challenges/agent-challenge/tests/test_public_task_event_redaction.py

---

📝 Walkthrough

Walkthrough

The task-event text redaction path replaces a monolithic secret-matching regex with bounded candidate and marker checks. New tests cover dotted-input performance and preservation of non-secret identifiers in realistic pip output.

Changes

Public secret redaction

Layer / File(s)	Summary
Bounded redaction implementation packages/challenges/agent-challenge/src/agent_challenge/api/routes.py	Adds candidate-word and marker regexes, introduces _redact_public_secret_segments(...), and routes _public_task_event_text(...) through it.
Performance and preservation validation packages/challenges/agent-challenge/tests/test_public_task_event_redaction.py	Adds timing tests for dotted/underscored input and repeated pip output, while checking that agent-challenge-1.0.1 remains unredacted.

Estimated code review effort: 3 (Moderate) | ~20 minutes

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly matches the main fix: replacing public log redaction logic to prevent event-loop wedging.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/public-task-event-redaction-redos

---

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.}