Summary

• Fix remaining circuit breaker false positives after fix(circuit-breaker): make repetitive detection target-aware and add enabled escape hatch #2655 + refactor d48ea025, specifically when events arrive in flat format without state.input
• Unknown tool input (undefined/null) is now treated as non-comparable — consecutive counter resets to 1 each time, instead of accumulating on bare tool name signatures

Context

#2655 introduced target-aware signatures (toolName::sortedJson(input)) to fix false-positive cancellations. Then d48ea025 refactored detection from sliding-window to consecutive-count. Both changes are correct individually, but together they leave a gap:

createToolCallSignature() falls back to bare tool name when toolInput is undefined/null:

if (toolInput === undefined || toolInput === null) {
  return toolName  // bare "read" — all flat-format calls look identical
}

This matters because resolveMessagePartInfo() has two event paths:

With consecutive-count detection, 20 flat-format read events (reading 20 different files) all produce signature "read", hitting the threshold. This is the exact scenario reported in #2652 — agents like Oracle/Explore/Librarian doing legitimate multi-file reads get cancelled.

Changes

src/features/background-agent/loop-detector.ts

Early return in recordToolCall() when toolInput is undefined/null: returns a unique sentinel signature (toolName::__unknown-input__) with consecutiveCount: 1. This means:

• Unknown input → counter never accumulates → no false positive
• Known identical input → existing signature logic → still detects true loops
• Absolute maxToolCalls cap → unaffected, still catches runaway agents

src/features/background-agent/loop-detector.test.ts

• Updated consecutive-trigger tests to use known input (matching real loop behavior)
• Added tests: 20 calls with undefined input → does NOT trigger
• Added tests: 20 calls with null input → does NOT trigger

src/features/background-agent/manager-circuit-breaker.test.ts

• Updated flat-format event test: 20 consecutive read events without state.input → task keeps running
• Strengthened absolute cap test with repeated flat-format events hitting maxToolCalls

Testing

bun test src/features/background-agent/loop-detector.test.ts        # 18 pass
bun test src/features/background-agent/manager-circuit-breaker.test.ts  # 8 pass
bun run typecheck  # clean

Related Issues

Follow-up to #2655 / #2652

Summary by cubic

Fixes circuit breaker false positives on flat-format tool events with no input. Unknown tool input is now treated as non-comparable, so consecutive counts reset and valid multi-file reads aren’t cancelled.

• Bug Fixes
  • In recordToolCall(), undefined/null input returns toolName::__unknown-input__ with consecutiveCount = 1.
  • Keeps loop detection for identical known inputs; maxToolCalls backstop unchanged.
  • Updated tests for undefined/null input and flat-format events; threshold tests now use known input.

^{Written for commit b9fa2a3. Summary will update on new commits.}