feat: add streaming loop detection with sentence sampling

[avoidwork]

Summary

Add a streaming loop detection mechanism that samples sentences from the active message as it streams in, detects repetition via a sliding window, and issues a silent "You're looping." nudge to break the cycle.

Motivation

The agent can get stuck in repetitive loops — re-checking status, re-running the same commands, or repeating the same phrasing across multiple responses. This wastes time, creates noise, and can lead to catastrophic decisions. A lightweight, streaming-aware detection mechanism catches loops early and gives the agent a chance to self-correct.

Implementation

New Components

Sentence Detector (src/agent/sentence-detector.js)

• Buffers incoming streaming text chunks
• Emits complete sentences when boundary characters (., !, ?) are detected
• Handles partial chunks, multiple sentences per chunk, and edge cases (..., ?!, abbreviations)

Sliding Window Tracker (src/agent/sliding-window-tracker.js)

• Maintains a 30-second sliding window of recent sentences with timestamps
• Releases old sentences automatically
• Tracks sentence frequency with normalization (trim whitespace, lowercase)

Loop Detector (src/agent/loop-detector.js)

• Combines sentence detector + sliding window tracker into a pipeline
• Triggers when any sentence appears >3 times within the window
• Emits { type: 'loop_detected' } callback event

Integration Points

Streaming Pipeline (src/agent/react.js)

• Loop detector instantiated per-stream in callReactAgentStreaming()
• Text chunks processed through detector before emitting to callback
• Detector reset on abort, completion, and error

TUI Handling (src/tui/app.js)

• Added handler for loop_detected events
• Displays "You're looping." as a silent nudge in the tool call display area
• Non-disruptive — doesn't interrupt the stream

Testing

• Unit tests for sentence detector (boundary detection, buffering, edge cases)
• Unit tests for sliding window tracker (add, expire, frequency counting, normalization)
• Unit tests for loop detector (threshold trigger, no false positives)
• Integration tests for full pipeline (streaming chunks, loop detection, cleanup, concurrent streams)

Verification

• All tests pass (1199 pass, 0 fail)
• Lint passes (0 warnings, 0 errors)
• Coverage maintained
• Modules import without errors

Files Changed

• src/agent/sentence-detector.js — New sentence boundary detection component
• src/agent/sliding-window-tracker.js — New sliding window tracker
• src/agent/loop-detector.js — New loop detection pipeline
• src/agent/react.js — Integrated loop detector into streaming pipeline
• src/tui/app.js — Added loop_detected event handler
• tests/unit/agent/sentence-detector.test.js — Unit tests
• tests/unit/agent/sliding-window-tracker.test.js — Unit tests
• tests/unit/agent/loop-detector.test.js — Unit tests
• tests/integration/streaming-loop-detection.test.js — Integration tests
• openspec/changes/streaming-loop-detection/ — OpenSpec artifacts

[avoidwork]

Audit Results: Streaming Loop Detection

Coverage Audit

• Goal 1 (Sentence Boundary Detector): Covered by spec "Sentence Boundary Detection" ✓
• Goal 2 (Sliding Window Sentence Tracker): Covered by spec "Sliding Window Sentence Tracking" ✓
• Goal 3 (Loop Detection & Nudge Injection): Covered by spec "Loop Detection Threshold" + "Silent Loop Nudge" ✓
• Goal 4 (Integration with Streaming Pipeline): Covered by spec "Per-Stream Sampler State" + "Integration with Streaming Pipeline" ✓

Fidelity Audit

All specs faithfully represent the original intent of each goal. Key elements preserved:

• 30-second sliding window ✓

• 3 occurrence threshold ✓

• Silent "You're looping." nudge following RECURSION_LIMIT_MESSAGE pattern ✓
• Per-stream sampler state ✓
• Integration into callReactAgentStreaming() on_chat_model_stream handler ✓

Completeness Audit

All key requirements and acceptance criteria are captured in specs with scenarios. Minor notes:

• Edge cases (..., ?!, abbreviations) are covered in tasks 1.5 — acceptable deferral from spec
• "Debounce rapid successive loops" open question from design is correctly deferred as future enhancement
• All 4 goals have corresponding spec requirements with at least one scenario each

Consistency Audit

Tasks map correctly to spec requirements:

• Sentence Boundary Detection → Tasks 1.1-1.6 ✓
• Sliding Window Sentence Tracking → Tasks 2.1-2.5 ✓
• Loop Detection Threshold → Tasks 3.1-3.4 ✓
• Integration with Streaming Pipeline → Tasks 4.1-4.6 ✓
• Silent Loop Nudge → Tasks 5.1-5.4 ✓
• TUI handling → Tasks 6.1-6.4 ✓
• Integration testing → Tasks 7.1-7.5 ✓
• Verification → Tasks 8.1-8.4 ✓

Verdict

No errors found. Specs are comprehensive, faithful to goals, and tasks map correctly. Proceeding to Step 6.