[ClaudiaFang]

Summary

Finish the existing Rage Loop detector end-to-end for #34 (part of #31). Firstsun and b08x already shared the same detector source, so this PR avoids copying duplicate code.

Fixes

• Remove the major false positive: two empty tool_calls: '[]' values no longer count as repeated tool invocations. Ignore trivial short messages and non-monotonic timestamps.
• Accept common input_tokens and total_tokens usage shapes.
• Treat detection as advisory (suspected stalled loop): the LLM still needs to verify semantic nonprogress.
• Add rage-loop to JSON Schema category enum (already present in prompt constants but missing from the schema).
• Pass detected signals into both CLI and Hono server session analysis paths, including chunked and facet-only paths.
• Regression tests for empty tool calls, invalid timestamps, low-context retries, positive loops, and prompt hint generation.

Compatibility

• No database migrations or full-fork merge.
• No changes to Firstsun multi-home, Personality, Kilo, Reports, Docker or Node/pnpm toolchain.
• Keep rage-loop distinct from repeated-mistakes: repetitive failures without saturated context must not be categorized as rage loop by the heuristic.

Validation checklist (keep Draft until complete)

• GitHub CI build + tests: https://github.com/firstsun-dev/code-insights/actions/runs/40624170710
• Verify mixed-agent/tool-call fixtures and recall/false-positive rate.
• Verify CLI/server/Reflect/Patterns presentations and no duplicate alerts.
• Human review.

Refs #31 and #34.
Source: https://github.com/b08x/code-insights/blob/development/cli/src/analysis/loop-detector.ts