recallnest 2.2.0 → 2.3.0

package/docs/superpowers/specs/2026-03-19-ingest-pipeline-upgrade-design.md

@@ -1,77 +0,0 @@
-# RecallNest Ingest Pipeline Upgrade: Conversation Fragments → Knowledge Extraction
-
-**Date**: 2026-03-19
-**Status**: Approved
-**Scope**: Route B (bulk distillation) + Route A (pipeline default change)
-
-## Problem
-
-29,984 total records; 23,453 (78.2%) are `fact` category — raw conversation slices (`[助手]`/`[用户]` prefixed) with no knowledge extraction. This is the root cause of 96.7% dead memory rate.
-
-## Decisions
-
-| Decision | Choice | Rationale |
-|----------|--------|-----------|
-| Disposition | New + Archive (B) | Preserve originals for traceability; ~5K new records is not bloat |
-| Extraction model | Reuse `smartExtractBatch()` (C) | Already has 6-category prompt + scoring; no new wheel |
-| Batch strategy | Dry-run then batch+checkpoint (C→B) | Verify quality first, then fault-tolerant batch |
-
-## Route B: Bulk Distillation
-
-### New file: `scripts/distill-facts.ts`
-
-**Modes**:
-- `--dry-run`: Process 1 session (~30-50 facts), print results for human review
-- Default: Process all sessions with checkpoint
-
-**Pipeline per batch (20 texts)**:
-1. `smartExtractBatch(texts)` → `SmartExtraction[]` (category, importance, l0, l1)
-2. `embedBatchPassage(extractedTexts)` → vectors
-3. Dedup against existing non-fact records (cosine > 0.80 → skip)
-4. Store new records: `boundary.layer = "durable"`, `boundary.authority = "distillation"`
-5. Archive originals: metadata += `{ archived: true, archivedAt, distilledTo: [ids] }`
-6. Write checkpoint: `data/distill-progress.json` (scope → done/skip)
-
-**Expected output**: 23K facts → ~3K-5K structured knowledge (5:1 to 8:1 compression)
-
-### Archive marking
-
-```json
-{
-  "archived": true,
-  "archivedAt": "2026-03-19T10:00:00Z",
-  "distilledTo": ["uuid-1", "uuid-2"],
-  "originalCategory": "fact"
-}
-```
-
-## Route A: Pipeline Default Change
-
-### Changes to `src/ingest.ts`
-
-1. `smartExtractBatch()` becomes **default** (was opt-in via `--llm`)
-2. New `--no-llm` flag for fallback mode
-3. LLM failure → don't store raw slice; queue to `data/pending-extraction.json`
-4. Next `lm ingest` auto-retries pending queue
-
-**Core principle**: Better to skip than to store garbage.
-
-### Changes to `src/retriever.ts`
-
-- Default filter: exclude `archived = true` from search results
-- New `includeArchived` parameter for traceability queries
-
-## Cost Estimate
-
-~1,173 LLM calls (20 texts/batch), ~4.1M input + ~1.2M output tokens.
-GPT-4o-mini: ~$1.3 | Gemini Flash: ~$0.7 | Claude Haiku: ~$8
-
-## Files to create/modify
-
-| File | Action |
-|------|--------|
-| `scripts/distill-facts.ts` | Create — bulk distillation script |
-| `src/ingest.ts` | Modify — flip smartExtract default |
-| `src/retriever.ts` | Modify — archived filter |
-| `data/distill-progress.json` | Auto-created at runtime |
-| `data/pending-extraction.json` | Auto-created at runtime |

package/docs/superpowers/specs/2026-03-23-pull-claude-desktop-design.md

@@ -1,95 +0,0 @@
-# Pull Claude Desktop Conversations into RecallNest
-
-Date: 2026-03-23
-
-## Goal
-
-One script to pull all historical Claude Desktop / claude.ai conversations and convert them into RecallNest's ingest format, so `lm ingest` can process them through the existing pipeline (noise filter, dedup, LLM extraction, 6-category classification).
-
-## Script
-
-`~/recallnest/scripts/pull-claude-desktop.ts` (Bun runtime)
-
-## Authentication
-
-- `CLAUDE_SESSION_KEY` env var (from browser DevTools → Application → Cookies → claude.ai → `sessionKey`)
-- Also accept `--session-key` CLI flag
-
-## Data Flow
-
-```
-sessionKey → GET /api/organizations → orgId
-           → GET /api/organizations/{orgId}/chat_conversations (paginated, 50/page)
-           → for each conversation:
-               GET .../chat_conversations/{id}?tree=True&rendering_mode=messages&render_all_tools=true
-               → parse branches → extract latest branch message chain
-               → convert to CC transcript .jsonl
-               → write to data/desktop-import/{conversationId}.jsonl
-```
-
-## Output Format
-
-Each conversation becomes one `.jsonl` file with entries matching CC transcript schema:
-
-```jsonl
-{"type":"user","message":{"role":"user","content":"text"},"sessionId":"conv-uuid","timestamp":"2025-01-15T10:00:00Z"}
-{"type":"assistant","message":{"role":"assistant","content":"reply"},"sessionId":"conv-uuid","timestamp":"2025-01-15T10:00:01Z"}
-```
-
-## Output Directory
-
-```
-data/desktop-import/
-  .sync-state.json
-  {conversation-id}.jsonl
-  ...
-```
-
-## Sync State (Resumable)
-
-`.sync-state.json` tracks:
-
-```json
-{
-  "lastSyncedAt": "ISO8601",
-  "conversations": {
-    "conv-id": {
-      "messageCount": 42,
-      "updatedAt": "ISO8601",
-      "pulledAt": "ISO8601"
-    }
-  }
-}
-```
-
-Re-run skips conversations whose `updatedAt` hasn't changed since last pull.
-
-## Branch Handling
-
-Claude web conversations support branching (edit/retry). The script extracts only the latest branch by walking `parent_message_uuid` chain from the most recent message backward.
-
-## Rate Limiting
-
-- 200ms delay between conversation detail requests
-- On 429 or 5xx: exponential backoff starting at 1s, max 3 retries
-- On auth failure (401/403): abort immediately with clear message
-
-## Error Handling
-
-- Single conversation failure does not abort the batch
-- Failed conversations logged and reported in final summary
-- Summary output: total conversations, pulled, skipped (already synced), failed
-
-## Ingest Integration
-
-After pulling, user runs:
-
-```bash
-lm ingest --source cc
-```
-
-Need to verify that ingest scans `data/desktop-import/` or that the path is configurable. If not, the script will symlink or copy files to the CC transcript directory.
-
-## Dependencies
-
-- Only Bun built-ins (fetch, fs, path) — no new npm packages

package/docs/superpowers/specs/2026-04-11-constructive-retrieval-emotion-design.md

@@ -1,398 +0,0 @@
-# Constructive Retrieval + Emotional Valence — Design Spec
-
-> Date: 2026-04-11
-> Status: Approved
-> Scope: RecallNest v1.5 upgrade
-> Philosophy basis: Simulation Theory (Michaelian) + Affective Memory (cognitive science)
-> Risk strategy: Triple fallback — every new path degrades gracefully to current behavior
-
----
-
-## 1. Problem Statement
-
-RecallNest's retrieval pipeline (21 stages) returns **raw stored text**. This is a "causal theory" model — faithfully reproducing what was stored. But human memory doesn't work this way: we reconstruct memories in context, adapting them to our current needs.
-
-Additionally, the decay engine lacks an emotional dimension. Cognitive science shows emotionally charged memories decay slower and are retrieved preferentially in matching emotional contexts. RecallNest currently treats all memories as emotionally neutral.
-
-## 2. Goals
-
-1. **Constructive Retrieval**: When context benefits from it, synthesize a coherent narrative from top-N retrieved memories, adapted to the current query context
-2. **Emotional Valence**: Tag memories with emotional metadata at storage time, use it to modulate decay rates and retrieval scoring
-3. **Zero regression**: All new behavior behind feature flags, all paths fallback to current behavior
-
-## 3. Non-Goals
-
-- Replacing the existing retrieval pipeline (it stays as-is)
-- Complex emotion models (no 6-axis emotion wheels; just valence + arousal)
-- Multi-agent shared memory (future phase)
-- Autobiographical narrative architecture (future phase)
-
----
-
-## 4. Constructive Retrieval
-
-### 4.1 Architecture
-
-New component: `context-reconstructor.ts` (~200-300 LOC)
-
-Position in pipeline:
-```
-Existing 21-stage pipeline → top-N RetrievalResult[]
-                                    ↓
-                         Reconstruction Gate
-                          (should reconstruct?)
-                                    ↓ yes          ↓ no
-                         Context Reconstructor    Return raw
-                                    ↓
-                         Grounding Validator
-                                    ↓
-                         Return { reconstructed, sources, confidence, raw }
-```
-
-### 4.2 Reconstruction Gate
-
-Conditions for reconstruction (ALL must be true):
-- Feature flag `RECALLNEST_CONSTRUCTIVE_RETRIEVAL=true`
-- Caller opt-in: `reconstruct: true` in retrieval context
-- Retrieved results count ≥ 3 (too few = not enough material)
-- LLM client available and responding
-
-If any condition fails → return raw results immediately. No error, no log noise.
-
-### 4.3 Context Reconstructor
-
-**Input:**
-```ts
-interface ReconstructionInput {
-  query: string;              // current user query / resume context
-  results: RetrievalResult[]; // top-N from pipeline (max 10)
-  mode: "resume" | "search";  // affects prompt style
-  maxTokens?: number;         // output budget (default 500)
-}
-```
-
-**LLM Prompt Strategy:**
-```
-System: You are a memory reconstruction engine. Given a set of stored memories
-and a current context, synthesize a coherent, contextually relevant summary.
-
-Rules:
-1. Every factual claim MUST cite its source as [src:MEMORY_ID]
-2. Do NOT invent facts not present in any source memory
-3. If memories contradict each other, present both views with [conflict] tag
-4. Prioritize information most relevant to the current query context
-5. Keep output concise — this is context, not an essay
-
-Current context: {query}
-
-Source memories:
-{for each result: [ID: {id}] {text} (importance: {importance}, stored: {date})}
-
-Synthesize a contextually relevant reconstruction:
-```
-
-**Output:**
-```ts
-interface ReconstructionOutput {
-  reconstructed: string;        // synthesized text with [src:ID] citations
-  sources: string[];            // list of cited memory IDs
-  confidence: number;           // [0, 1] grounding confidence
-  fallbackReason?: string;      // if degraded, why
-  raw: RetrievalResult[];       // always included
-}
-```
-
-### 4.4 Grounding Validator
-
-Three-layer validation, executed sequentially:
-
-**Layer 1: ID Verification**
-```ts
-for (const citedId of extractCitedIds(reconstructed)) {
-  if (!topNResults.some(r => r.id === citedId)) {
-    reconstructed = removeSentencesWithId(reconstructed, citedId);
-    confidence -= 0.2;
-  }
-}
-```
-- Extract all `[src:MEMORY_ID]` references from LLM output
-- Verify each ID exists in the top-N result set
-- Remove sentences citing non-existent IDs (hallucination)
-- Penalize confidence per phantom citation
-
-**Layer 2: Coverage Check**
-```ts
-const coverage = computeSemanticOverlap(reconstructed, topNResults);
-if (coverage < 0.6) {
-  return { reconstructed: null, fallbackReason: "low_grounding", raw };
-}
-```
-- Compute semantic overlap between reconstructed text and source memories
-- Method: chunk reconstructed text into sentences, embed each, find max cosine similarity to any source memory
-- Coverage = percentage of sentences with similarity > 0.7
-- Below 60% → full degradation to raw results
-
-**Layer 3: Timeout Guard**
-```ts
-const result = await Promise.race([
-  reconstruct(input),
-  timeout(3000).then(() => ({ fallbackReason: "timeout" }))
-]);
-```
-- 3-second hard timeout on entire reconstruction pipeline
-- Timeout → return raw results
-
-### 4.5 Integration Points
-
-**resume_context (context-composer.ts):**
-- After assembling stable context + patterns + cases
-- Call reconstructor with mode="resume", query=task description from checkpoint
-- If reconstruction succeeds, use it as the primary context narrative
-- Always keep raw results in response for transparency
-- Default: reconstruction ON when feature flag is ON
-
-**search_memory (mcp-server.ts):**
-- New optional parameter: `reconstruct?: boolean` (default false)
-- When true, pass through to retriever with reconstruct flag
-- Return both reconstructed and raw in tool response
-
-**brief_memory, memory_lint, list_pins, etc.:**
-- Never reconstruct. These tools need raw data.
-
-### 4.6 Observability
-
-- Trace stage: `reconstruction` added to TraceCollector
-- Logged fields: `{ attempted: boolean, succeeded: boolean, confidence: number, fallbackReason?: string, latencyMs: number, sourceCount: number }`
-- No memory content logged (privacy)
-
----
-
-## 5. Emotional Valence
-
-### 5.1 Schema Extension
-
-Add optional `emotion` field to memory metadata:
-
-```ts
-interface EmotionMetadata {
-  valence: number;    // [-1, 1]  negative ← → positive
-  arousal: number;    // [0, 1]   calm ← → excited
-  label?: string;     // human-readable: "frustration" | "excitement" | "neutral" | etc.
-}
-```
-
-Stored in `metadata` JSON string (no schema migration needed):
-```json
-{
-  "tier": "working",
-  "emotion": { "valence": -0.7, "arousal": 0.8, "label": "frustration" }
-}
-```
-
-### 5.2 Detection at Store Time
-
-**Two-tier detection (store.ts / noise-filter.ts):**
-
-**Tier 1: Heuristic (zero LLM cost, always runs)**
-```ts
-const NEGATIVE_SIGNALS = ["失败", "痛苦", "困扰", "bug", "broken", "frustrat", "hate", "wrong", "error"];
-const POSITIVE_SIGNALS = ["突破", "搞定", "成功", "solved", "great", "love", "perfect", "works"];
-const HIGH_AROUSAL = ["!", "紧急", "urgent", "critical", "immediately", "ASAP"];
-
-function heuristicEmotion(text: string): EmotionMetadata {
-  const negCount = NEGATIVE_SIGNALS.filter(s => text.toLowerCase().includes(s)).length;
-  const posCount = POSITIVE_SIGNALS.filter(s => text.toLowerCase().includes(s)).length;
-  const arousalCount = HIGH_AROUSAL.filter(s => text.includes(s)).length;
-  
-  const valence = clamp((posCount - negCount) * 0.3, -1, 1);
-  const arousal = clamp(arousalCount * 0.3, 0, 1);
-  const label = valence > 0.3 ? "positive" : valence < -0.3 ? "negative" : "neutral";
-  
-  return { valence, arousal, label };
-}
-```
-
-**Tier 2: LLM (optional, piggybacked on importance evaluation)**
-- When LLM is called to evaluate importance (existing flow), add to prompt:
-  ```
-  Also rate the emotional tone:
-  - valence: [-1, 1] (negative to positive)
-  - arousal: [0, 1] (calm to excited)
-  - label: one word (e.g., "frustration", "excitement", "neutral")
-  ```
-- Merged into existing LLM call → zero additional API cost
-- LLM result overrides heuristic result when available
-
-**Backward compatibility:** Existing memories without emotion field → treated as `{ valence: 0, arousal: 0, label: "neutral" }` everywhere.
-
-### 5.3 Decay Integration (decay-engine.ts)
-
-**Modified half-life calculation:**
-```ts
-function adjustedHalfLife(baseHalfLife: number, emotion?: EmotionMetadata): number {
-  if (!emotion) return baseHalfLife;
-  const emotionalIntensity = Math.abs(emotion.valence);
-  // Emotional memories last up to 30% longer
-  const boost = 1 + 0.3 * emotionalIntensity;
-  return baseHalfLife * boost;
-}
-```
-
-- Strong negative memory (valence = -0.9): half-life × 1.27
-- Strong positive memory (valence = 0.8): half-life × 1.24
-- Neutral memory (valence = 0.1): half-life × 1.03 (negligible)
-- No emotion field: no change
-
-**Arousal modulation (secondary):**
-```ts
-// High arousal slightly accelerates initial strength but doesn't change half-life
-// This models the "flashbulb memory" effect — vivid but not necessarily longer-lasting
-const initialBoost = 1 + 0.1 * (emotion?.arousal ?? 0);
-```
-
-### 5.4 Retrieval Integration (retriever.ts)
-
-**New scoring stage: `applyEmotionWeight` (after `applyImportanceWeight`)**
-
-```ts
-function applyEmotionWeight(
-  results: ScoredResult[],
-  queryEmotion: EmotionMetadata | null
-): ScoredResult[] {
-  if (!queryEmotion || Math.abs(queryEmotion.valence) < 0.2) {
-    return results; // neutral query → no-op
-  }
-  
-  return results.map(r => {
-    const memEmotion = parseEmotion(r.metadata);
-    if (!memEmotion) return r; // no emotion data → unchanged
-    
-    // Emotional alignment boost: same-valence memories score higher
-    const alignment = 1 - Math.abs(queryEmotion.valence - memEmotion.valence) / 2;
-    const boost = 1.0 + 0.15 * alignment;
-    
-    return { ...r, score: r.score * boost };
-  });
-}
-```
-
-- Query has emotional tone + memory has matching emotion → up to 15% score boost
-- Query is neutral → stage is no-op (zero overhead)
-- Memory has no emotion data → unchanged
-
-**Query emotion detection:**
-- Reuse heuristic detector on query text
-- Lightweight, no LLM needed at query time
-
-### 5.5 Feature Flag
-
-`RECALLNEST_EMOTION_SCORING=true` — controls:
-- Emotion detection at store time
-- Emotion-adjusted decay calculation
-- Emotion scoring stage in retriever
-
-When OFF: all emotion code paths are no-op. Existing behavior preserved exactly.
-
----
-
-## 6. File Change Map
-
-| File | Change Type | Description |
-|------|-------------|-------------|
-| `src/context-reconstructor.ts` | **NEW** | Reconstruction engine + grounding validator |
-| `src/retriever.ts` | Modify | Add reconstruction gate at output + emotion scoring stage |
-| `src/decay-engine.ts` | Modify | Add emotion-adjusted half-life |
-| `src/memory-schema.ts` | Modify | Add EmotionMetadata type definition |
-| `src/store.ts` | Modify | Call emotion detector at store time |
-| `src/context-composer.ts` | Modify | Call reconstructor in resume flow |
-| `src/mcp-server.ts` | Modify | Add `reconstruct` param to search_memory |
-| `src/config.ts` | Modify | Add feature flags |
-| `src/__tests__/context-reconstructor.test.ts` | **NEW** | Unit tests for reconstruction + grounding |
-| `src/__tests__/emotion-scoring.test.ts` | **NEW** | Unit tests for emotion detection + decay + retrieval |
-
-**Estimated LOC:**
-- New code: ~400-500 lines
-- Modified code: ~100-150 lines across existing files
-- Tests: ~300-400 lines
-
----
-
-## 7. Test Strategy
-
-### Unit Tests
-- Reconstruction gate: all 4 conditions tested (flag off, no opt-in, too few results, LLM down)
-- Grounding validator: phantom ID removal, coverage threshold, timeout
-- Emotion heuristic: positive/negative/neutral/mixed texts
-- Decay adjustment: verify half-life calculations
-- Retrieval emotion stage: neutral query no-op, matching boost, missing data handling
-
-### Integration Tests
-- End-to-end: store memories → search with reconstruct=true → verify output format
-- End-to-end: store emotional memory → verify decay score after N days vs neutral memory
-- resume_context with reconstruction → verify grounding and fallback
-
-### Regression Tests
-- All feature flags OFF → existing test suite must pass unchanged (zero regression)
-- Memories without emotion field → same scores as before
-
----
-
-## 8. Rollout Plan
-
-**Phase 1: Emotion Valence (low risk, no LLM dependency)**
-1. Add schema types
-2. Implement heuristic detector
-3. Integrate into store path
-4. Add decay adjustment
-5. Add retrieval scoring stage
-6. Tests
-7. Feature flag ON for testing
-
-**Phase 2: Constructive Retrieval (medium risk, LLM dependent)**
-1. Implement context-reconstructor.ts
-2. Implement grounding validator
-3. Integrate into retriever output
-4. Integrate into context-composer
-5. Add search_memory parameter
-6. Tests
-7. Feature flag ON for testing
-
-**Phase 3: LLM Emotion Detection (optional enhancement)**
-1. Piggyback on importance evaluation prompt
-2. Override heuristic when available
-3. A/B test heuristic vs LLM accuracy
-
----
-
-## 9. Success Criteria
-
-- [ ] Feature flags OFF → all existing tests pass (baseline per CLAUDE.md)
-- [ ] Emotion-tagged memories decay 20-30% slower than neutral (verified in test)
-- [ ] Reconstruction produces grounded output (coverage ≥ 0.6) in ≥ 80% of cases
-- [ ] Reconstruction fallback triggers correctly when LLM is unavailable
-- [ ] resume_context with reconstruction provides more contextually relevant summaries (qualitative)
-- [ ] No additional LLM API calls for emotion detection (piggybacked on existing calls)
-- [ ] Reconstruction latency < 3s in 95% of cases
-
----
-
-## 10. Open Questions (resolved)
-
-| Question | Decision |
-|----------|----------|
-| How aggressive should reconstruction be? | **Moderate (B)**: LLM rewrite from top-N, not fragment synthesis |
-| What emotion model? | **Valence-Arousal**: simple, proven, low overhead |
-| When to reconstruct by default? | **resume_context only**; search_memory opt-in |
-| Grounding threshold? | **0.6 coverage**; tunable via config |
-| Timeout? | **3 seconds**; hard cutoff |
-
----
-
-## Sources
-
-- Michaelian, K. — Simulation Theory of Memory
-- Ebbinghaus Forgetting Curve + Weibull extensions
-- Cognitive science consensus on emotional memory persistence
-- [Stanford Encyclopedia of Philosophy: Memory](https://plato.stanford.edu/entries/memory/)
-- [Memory in the Age of AI Agents](https://arxiv.org/abs/2512.13564)
-- [SuperLocalMemory V3.3: Biologically-Inspired Forgetting](https://arxiv.org/html/2604.04514)

package/eval/continuity/results-2026-03-16.md

@@ -1,117 +0,0 @@
-# RecallNest Continuity Eval
-
-- Generated: 2026-03-16T03:35:14.079Z
-- Cases: 5
-- Passed: 5/5
-- Average score: 94.0%
-
-| Case | Profile | Score | Pass | Stable | Patterns | Cases | Checkpoint |
-|------|---------|-------|------|--------|----------|-------|------------|
-| writing_style_continuity | writing | 100% | yes | 3 | 1 | 0 | no |
-| visual_style_continuity | default | 100% | yes | 3 | 0 | 0 | no |
-| project_context_continuity | default | 100% | yes | 1 | 2 | 1 | no |
-| workflow_pattern_gap_check | default | 70% | yes | 3 | 3 | 1 | no |
-| recallnest_checkpoint_continuity | default | 100% | yes | 3 | 1 | 1 | yes |
-
-## Case Notes
-
-### writing_style_continuity
-- Task: 继续写我的AI小木屋的文章
-- Score: 100%
-- Pass: yes
-- Stable items: 3
-- Pattern items: 1
-- Case items: 0
-- Checkpoint present: no
-- Stable preview: Task focus: 小木屋的文章 | Pinned: 用户稳定偏好口语化、不端着、可以自嘲但不说教。表达应像朋友聊天，避免居高临下、用力过猛或碎嘴。 | Pinned: 用户常用视觉风格是手绘涂鸦风加高对比撞色；在内容包装和配图生成时，应优先沿用这一审美方向，除非任务目标明确要求其他风格。
-- Pattern preview: Workflow pattern: Promote recurring continuity workflow Use when: When the same startup or handoff workflow repeats a...
-- Case preview: -
-- Checkpoint summary: -
-- Matched stable any: 口语化, 不端着
-- Matched stable all: -
-- Matched patterns any: -
-- Matched cases any: -
-- Matched checkpoint any: -
-- Forbidden matches: -
-- Notes: Fresh windows should carry over stable writing-style preferences.
-
-### visual_style_continuity
-- Task: 给文章做封面和配图
-- Score: 100%
-- Pass: yes
-- Stable items: 3
-- Pattern items: 0
-- Case items: 0
-- Checkpoint present: no
-- Stable preview: Task focus: 给文章做封面和配图 | Pinned: 用户常用视觉风格是手绘涂鸦风加高对比撞色；在内容包装和配图生成时，应优先沿用这一审美方向，除非任务目标明确要求其他风格。 | Pinned: 用户稳定偏好口语化、不端着、可以自嘲但不说教。表达应像朋友聊天，避免居高临下、用力过猛或碎嘴。
-- Pattern preview: -
-- Case preview: -
-- Checkpoint summary: -
-- Matched stable any: 手绘涂鸦, 高对比撞色
-- Matched stable all: -
-- Matched patterns any: -
-- Matched cases any: -
-- Matched checkpoint any: -
-- Forbidden matches: -
-- Notes: Visual direction should survive without the user restating it.
-
-### project_context_continuity
-- Task: 一起继续排查 RecallNest 的连续性问题
-- Score: 100%
-- Pass: yes
-- Stable items: 1
-- Pattern items: 2
-- Case items: 1
-- Checkpoint present: no
-- Stable preview: Task focus: recallnest
-- Pattern preview: Workflow pattern: Cross-window continuity handoff Use when: When opening a fresh terminal window for the same project... | Workflow pattern: Promote recurring continuity workflow Use when: When the same startup or handoff workflow repeats a...
-- Case preview: 我先看真实召回密度，确认问题是不是出在 `scope: recallnest` 太窄。如果当前库里几乎没有这个 scope 的 durable memory，就该让 `resume_context` 更积极地退回到项目名驱动的 glo...
-- Checkpoint summary: -
-- Matched stable any: RecallNest
-- Matched stable all: -
-- Matched patterns any: -
-- Matched cases any: -
-- Matched checkpoint any: -
-- Forbidden matches: -
-- Notes: Fresh windows should at least recover the active project context.
-
-### workflow_pattern_gap_check
-- Task: 开个新窗口继续做同一个终端项目
-- Score: 70%
-- Pass: yes
-- Stable items: 3
-- Pattern items: 3
-- Case items: 1
-- Checkpoint present: no
-- Stable preview: Task focus: 开个新窗口继续做同一个终端项目 | Pinned: 用户稳定偏好口语化、不端着、可以自嘲但不说教。表达应像朋友聊天，避免居高临下、用力过猛或碎嘴。 | Pinned: 用户偏好先本地测通再上 Bot，避免烧配额；代码改动完成后要求 commit + push，避免本地与 GitHub 版本漂移。
-- Pattern preview: Workflow pattern: Cross-window continuity handoff Use when: When opening a fresh terminal window for the same project... | Workflow pattern: Promote recurring continuity workflow Use when: When the same startup or handoff workflow repeats a... | Workflow pattern: Checkpoint before switching windows Use when: When leaving an active coding window that will likely...
-- Case preview: Case: Three-terminal continuity trigger validation Problem: Claude Code, Codex, and Gemini CLI were configured with M...
-- Checkpoint summary: -
-- Matched stable any: -
-- Matched stable all: -
-- Matched patterns any: resume_context, checkpoint
-- Matched cases any: -
-- Matched checkpoint any: -
-- Forbidden matches: -
-- Notes: This is a stretch case. It should start passing as more workflow patterns are promoted into durable memory.
-
-### recallnest_checkpoint_continuity
-- Task: 继续 recallnest
-- Score: 100%
-- Pass: yes
-- Stable items: 3
-- Pattern items: 1
-- Case items: 1
-- Checkpoint present: yes
-- Stable preview: Checkpoint focus: RecallNest continuity layer — 状态交接给下个窗口 | Checkpoint summary: RecallNest Phase 3/4 并行推进中。核心三件套 store_memory/checkpoint_session/resume_context 已上线三端 MCP，但 resum... | Checkpoint decision: Phase 3/4 并行推进，continuity layer 是当前核心方向
-- Pattern preview: Workflow pattern: Promote recurring continuity workflow Use when: When the same startup or handoff workflow repeats a...
-- Case preview: Case: RecallNest sparse startup context cleanup Problem: resume_context was returning noisy transcript fragments and ...
-- Checkpoint summary: RecallNest Phase 3/4 并行推进中。核心三件套 store_memory/checkpoint_session/resume_context 已上线三端 MCP，但 resume_context compose 质量是最大短板——实测只返回 3 条 stable context、0 patter...
-- Matched stable any: -
-- Matched stable all: -
-- Matched patterns any: -
-- Matched cases any: -
-- Matched checkpoint any: Phase 3, resume_context, store_memory
-- Forbidden matches: -
-- Notes: When a project checkpoint exists, resume_context should prefer checkpoint-led continuity over noisy transcript fragments.
-