fix(analysis): finish rage-loop and repeated-failure detection end-to-end #34

狀態更新（2026-10-28）

• Draft PR fix(analysis): complete rage-loop detection and avoid false positives (#34) #38 已建立正式 Development 關聯，CI Build/Test 有通過紀錄。
• 尚待混合 Agent／Tool-call fixtures、CLI/Server/Reflect/Patterns 呈現及誤判率驗證、Human Review。
• 保持 Open 至 PR 驗收與合併；不要把 Rage Loop 與 repeated-mistakes 混為一談。Parent epic: selectively port b08x session intelligence, long-context retrieval, RAG chat and source adapters #31。

Parent: #31 · Phase A (P1)

Verified baseline

firstsun-dev/main:cli/src/analysis/loop-detector.ts has the same blob SHA (4ff49ea4b6f7060fd59c138c9355bf1a75213b8b) as b08x; do not copy it again. Firstsun's prompts.ts already passes a <rage_loop> hint, but classification coverage must be verified. Its JSON Schema currently lacks the rage-loop enum value present in b08x, potentially causing rejected output.

Scope

• Audit both CLI and server analysis paths: detector -> hint -> schema enum -> response normalization -> storage -> Dashboard/Patterns and export.
• Ensure the rage-loop category and repeated-mistakes remain distinguishable and consistently supported in prompt constants, JSON Schema, parser, persistence and UI.
• Improve heuristic if warranted: do not count repeated null tool-call fields as evidence; handle invalid/out-of-order timestamps, >50k-token threshold variability, compactions and different provider usage shapes.
• Show supporting turn range and confidence as an advisory rather than asserting wasted effort from high token count alone; minimize false positives.

Acceptance

• Positive and negative fixture tests for repeated failed tool actions, compaction, high-context but productive iterative work, missing usage and very short sessions.
• No changes to historical DB migration numbers; no duplicate detectors.
• CLI + server results agree and no regression in existing Reflect/Patterns.

Source: https://github.com/b08x/code-insights/blob/development/cli/src/analysis/loop-detector.ts