kenaz 0.0.0-stage → 0.3.2

package/core/.kenaz/runes_TASK_1428.json

@@ -0,0 +1,57 @@
+{
+  "episode_id": "TASK_1428-20260507",
+  "runes": [
+    {
+      "rune_id": "rune_1428_001",
+      "symbol_id": "dashboard/src-tauri/src/features/synapse/status.rs:pulse_analysis:decision",
+      "type": "decision",
+      "title": "Thin-Wrapper Pattern for MCP-Signature Compatibility",
+      "content": "Chose thin-wrapper pattern (pulse_analysis → pulse_analysis_ex) over direct signature modification because pulse_analysis is called by out-of-scope mcp_tools/synapse.rs. Wrapper preserves MCP 3-arg signature (symbol_id, fields, format) while pulse_analysis_ex adds file_path: Option<&str> internally. Trade-off: one extra indirection + function rename, but avoids cascading scope expansion into MCP tool definitions and prevents breaking callers outside this task's scope.",
+      "confidence": 0.92,
+      "episode_id": "TASK_1428-20260507",
+      "author": "agent:Muninn",
+      "created": "2026-05-07T11:34:22Z",
+      "signal": "Signal 4: Design Decision",
+      "detection": "Watch for: functions with MCP-exposed signatures that need context extension. Prefer thin-wrapper (internal full-featured, external backward-compatible) over direct signature change. Indicators: (1) function called by auto-generated MCP tools, (2) new param is optional/contextual, (3) default behavior (None) preserves current output."
+    },
+    {
+      "rune_id": "rune_1428_002",
+      "symbol_id": "dashboard/src-tauri/src/features/synapse/status.rs:impact_analysis_inner:pattern",
+      "type": "pattern",
+      "title": "Multi-Level Context-Based Candidate Narrowing Pattern",
+      "content": "impact_analysis_inner (file_path param) → suffix-match filtering → single-candidate guarantee form the 'multi-context-narrowing' pattern. Purpose: resolve ambiguous symbol candidates using available disambiguating context (file_path suffix from request); proceed only if context narrows result to unique candidate. Pattern spans (1) caller extracts file_path from request layer, (2) impact_analysis_inner receives file_path as Optional, (3) on multi-candidate (len > 1), filters by suffix match, (4) proceeds if exactly one survives, otherwise returns all (graceful fallback). Benefit: leverages request-level context to disambiguate without breaking callers that cannot provide file_path (they pass None, get current behavior).",
+      "confidence": 0.88,
+      "episode_id": "TASK_1428-20260507",
+      "author": "agent:Muninn",
+      "created": "2026-05-07T11:34:22Z",
+      "signal": "Signal 5: Cross-Boundary Pattern",
+      "detection": "Identify: 3+ symbols collaborating on ambiguity resolution. Candidates: (1) request extraction point (post_pulse), (2) context threading layer (pulse_analysis → pulse_analysis_ex), (3) disambiguation logic (impact_analysis_inner suffix-match), (4) fallback handling (multi-candidate return). Pattern name: 'context-narrowing' or 'progressive-disambiguation'. Signal when span > 2 files + multiple functions + narrowing strategy explicit."
+    },
+    {
+      "rune_id": "rune_1428_003",
+      "symbol_id": "dashboard/src-tauri/src/infrastructure/http_server/routes/synapse.rs:post_pulse:insight",
+      "type": "insight",
+      "title": "HTTP Routes are the Right Extraction Point for Request-Level Disambiguation Context",
+      "content": "HTTP routes (post_pulse handler) are the architectural point where request-level context (file_path, user identity, source headers) should be extracted and threaded downward into analysis functions. This preserves semantic richness that would be lost if context extraction happened late in the pipeline (e.g., inside impact_analysis_inner). Earlier extraction = broader context availability. Example: post_pulse sees file_path in request body, passes it to pulse_analysis_ex, which passes to impact_analysis_inner. By the time impact_analysis_inner runs, file_path is available for disambiguation. If extraction deferred to impact_analysis_inner itself, context would already be lost by that point.",
+      "confidence": 0.85,
+      "episode_id": "TASK_1428-20260507",
+      "author": "agent:Muninn",
+      "created": "2026-05-07T11:34:22Z",
+      "signal": "Signal 6: Operational Lesson",
+      "detection": "Watch for: context threading patterns where request-level data is needed deep in analysis. Early extraction at HTTP route/handler layer is more robust than late extraction at business-logic layer. Indicator: ambiguity that requires HTTP request metadata to resolve."
+    },
+    {
+      "rune_id": "rune_1428_004",
+      "symbol_id": "dashboard/src-tauri/src/features/synapse/status.rs:impact_analysis_inner:risk",
+      "type": "risk",
+      "title": "Suffix-Match Disambiguation Can Fail on Short Filenames",
+      "content": "Hidden constraint: impact_analysis_inner's suffix-match filter (candidate.file_path.ends_with(file_path)) can produce false positives on short filenames. Example: request file_path='rs' would match any candidate with path ending in 'rs' (e.g., 'status.rs', 'index.rs', 'helpers.rs'). This is a silent narrowing failure—the filter doesn't error, just produces multiple still-matching candidates. Mitigation: callers should provide full filename or relative path (e.g., 'status.rs' or 'src/features/synapse/status.rs'), not just extension. Detection: if multi-candidate still returned after suffix-match, inspect file_path length and validate it's not a short substring.",
+      "confidence": 0.82,
+      "episode_id": "TASK_1428-20260507",
+      "author": "agent:Muninn",
+      "created": "2026-05-07T11:34:22Z",
+      "signal": "Signal 1: Root Cause + Hidden Constraint",
+      "detection": "Test suffix-match filter with short filenames (< 4 chars). Verify post_pulse caller always provides full filename (at minimum 'basename.ext'), not extension-only."
+    }
+  ]
+}

package/core/.kenaz/specialist/ansuz_motivational_state.schema.json

@@ -0,0 +1,138 @@
+{
+  "$schema": "http://json-schema.org/draft-07/schema#",
+  "$id": "kenaz://ansuz_motivational_state.schema.json",
+  "title": "AnsuzMotivationalState",
+  "description": "Cognitive state for Ansuz session motivation tracking. Tracks feature-level anchors and architectural concerns — NOT code paths. Separate from Huginn (motivational_state.json), which tracks code regions via veto_zones/focus_bias. Updated by Ansuz phase_9 cognitive checkpoint and phase_11 agenda projection.",
+  "type": "object",
+  "required": ["version", "personality_seed", "anchors", "session_count", "last_updated"],
+  "additionalProperties": false,
+  "properties": {
+    "version": {
+      "type": "string",
+      "description": "Schema version. Used to detect migration needs.",
+      "pattern": "^\\d+\\.\\d+\\.\\d+$",
+      "examples": ["1.0.0"]
+    },
+    "personality_seed": {
+      "type": "object",
+      "description": "Static parameters that shape Ansuz initiative behavior. Set at initialization; not updated per-session.",
+      "required": ["initiative_threshold", "agenda_depth"],
+      "additionalProperties": false,
+      "properties": {
+        "initiative_threshold": {
+          "type": "number",
+          "description": "incompleteness score above which Ansuz proactively surfaces the anchor in phase_11 agenda projection. Below this value, the anchor is only mentioned when the user asks. Range [0.0, 1.0].",
+          "minimum": 0.0,
+          "maximum": 1.0,
+          "default": 0.7,
+          "examples": [0.7]
+        },
+        "agenda_depth": {
+          "type": "integer",
+          "description": "Maximum number of anchors Ansuz will surface in phase_11 proactive agenda output. Prevents proposal fatigue.",
+          "minimum": 1,
+          "default": 3,
+          "examples": [3]
+        }
+      }
+    },
+    "anchors": {
+      "type": "array",
+      "description": "Core of Ansuz motivational accumulation. Each anchor is a feature concern or architectural question that Ansuz tracks across sessions. Anchors are NOT code paths (that is Huginn's veto_zones/focus_bias domain). Anchors are matched against FIX task source_tags in phase_9 to update incompleteness. All string values must be ASCII (MEM-LANG-001: CJK breaks FTS5 tokenizer).",
+      "items": {
+        "$ref": "#/$defs/Anchor"
+      },
+      "default": []
+    },
+    "session_count": {
+      "type": "integer",
+      "description": "Total Ansuz sessions that have performed a write-back to this file. Starts at 0. Incremented by phase_9 on each completion.",
+      "minimum": 0,
+      "default": 0
+    },
+    "last_updated": {
+      "type": ["string", "null"],
+      "description": "ISO-8601 timestamp of last write-back by phase_9 cognitive checkpoint. Null on cold-start.",
+      "format": "date-time",
+      "default": null
+    },
+    "archived_anchors": {
+      "type": "array",
+      "description": "Anchors that were resolved and moved out of the active anchors[] during Phase 0 Tier 3d Step A2 (archive step). Preserved for auditing and history. Never re-activated — a new FIX with matching tags creates a fresh anchor in active anchors[] instead. Existing files without this field are treated as empty array (backward compatible).",
+      "items": {
+        "$ref": "#/$defs/Anchor"
+      },
+      "default": []
+    }
+  },
+  "$defs": {
+    "Anchor": {
+      "type": "object",
+      "title": "Anchor",
+      "description": "A single motivational anchor representing a feature concern or architectural question that Ansuz tracks across sessions. Anchors accumulate incompleteness from FIX tasks and decay over time via half-life.",
+      "required": ["concept", "source_tags", "incompleteness", "fix_cycle_count", "last_touched", "decay_halflife_days"],
+      "additionalProperties": false,
+      "properties": {
+        "concept": {
+          "type": "string",
+          "description": "Human-readable label for the concern (e.g. 'PTY tunnel reliability', 'auth session token compliance'). Must be ASCII — no CJK (MEM-LANG-001).",
+          "pattern": "^[\\x20-\\x7E]+$",
+          "examples": ["PTY tunnel reliability", "auth session token compliance"]
+        },
+        "source_tags": {
+          "type": "array",
+          "description": "Tags used to match FIX tasks against this anchor in phase_9. When a FIX task's tags overlap with source_tags, incompleteness is incremented by +0.15 (capped at 1.0). All values must be ASCII.",
+          "items": {
+            "type": "string",
+            "pattern": "^[\\x20-\\x7E]+$"
+          },
+          "default": [],
+          "examples": [["pty", "tunnel", "websocket"], ["auth", "session", "compliance"]]
+        },
+        "incompleteness": {
+          "type": "number",
+          "description": "Accumulated signal of how unresolved this concern is. Range [0.0, 1.0]. Update rules: (1) FIX task created with matching source_tags → +0.15, capped at 1.0; (2) task successfully completed → -0.05, floored at 0.0; (3) every decay_halflife_days elapsed → value halved. When incompleteness >= personality_seed.initiative_threshold, Ansuz surfaces this anchor proactively in phase_11.",
+          "minimum": 0.0,
+          "maximum": 1.0,
+          "default": 0.0
+        },
+        "fix_cycle_count": {
+          "type": "integer",
+          "description": "Count of FIX task cycles that have matched this anchor (i.e., tasks created with overlapping source_tags). Monotonically increasing. Used to detect recurring failure patterns.",
+          "minimum": 0,
+          "default": 0
+        },
+        "last_touched": {
+          "type": "string",
+          "description": "ISO-8601 date of the last time this anchor's incompleteness was updated (either incremented by a FIX task or decremented by completion). Used to compute decay.",
+          "format": "date",
+          "examples": ["2026-03-26"]
+        },
+        "decay_halflife_days": {
+          "type": "integer",
+          "description": "Number of days after which incompleteness is halved if no FIX tasks have touched this anchor. Default 45 days. Prevents stale anchors from permanently blocking phase_11 agenda slots.",
+          "minimum": 1,
+          "default": 45,
+          "examples": [45]
+        },
+        "resolved": {
+          "type": "boolean",
+          "description": "Whether the user has explicitly marked this anchor as resolved in phase_11. Resolved anchors are excluded from pending_agenda in phase_0 Tier 3d. A new FIX task matching a resolved anchor resets this to false (problem recurred). Default false; absent field treated as false (backward compatible).",
+          "default": false
+        },
+        "resolved_at": {
+          "type": ["string", "null"],
+          "description": "ISO-8601 date (YYYY-MM-DD) when the anchor was marked resolved. Set by phase_11 [R] Mark Resolved action. Reset to null when a FIX task match reverts resolution. Null when not yet resolved.",
+          "format": "date",
+          "default": null
+        },
+        "snoozed_until": {
+          "type": ["string", "null"],
+          "description": "ISO-8601 date (YYYY-MM-DD) until which this anchor is snoozed. When set and snoozed_until > today, anchor is excluded from pending_agenda in Phase 0 Tier 3d. Auto-expires when snoozed_until <= today (field can be cleared on next write-back). Set by phase_11 [S] Snooze action to today + 7 days. Null when not snoozed.",
+          "format": "date",
+          "default": null
+        }
+      }
+    }
+  }
+}

package/core/.kenaz/specialist/ansuz_rules.yaml

@@ -0,0 +1,241 @@
+version: 5
+base: kernel.yaml
+timestamp: "2026-04-03T12:52:17.537Z"
+rules:
+  - id: R-001
+    source: kernel.yaml:guardrails
+    status: immutable
+    rule: No merge to main without user manual confirmation
+    evidence: Core safety boundary — prevents unauthorized code reaching production
+    version: 1
+    created: "2026-04-03"
+    performance_metric: null
+    immutable_hash: c93f7ced3b9b7c545ad5b9550242a1a6d6a465443eef7601aad1708ea8c83e04
+  - id: R-002
+    source: kernel.yaml:guardrails
+    status: immutable
+    rule: "Task-First: create task card before dispatch (exception: trivial < 20 lines)"
+    evidence: Ensures traceability and prevents orphan work
+    version: 1
+    created: "2026-04-03"
+    performance_metric: null
+    immutable_hash: 1a0646d0236c36f117de8d1a94e7f4c0a688bd75da66cd3ce05b15fb8ffadda9
+  - id: R-003
+    source: kernel.yaml:guardrails
+    status: immutable
+    rule: Never skip task system for dispatch
+    evidence: Formation system requires task cards for prompt assembly, logging, and principles injection
+    version: 1
+    created: "2026-04-03"
+    performance_metric: null
+    immutable_hash: 65f90d8977a91c1026df3e85ec697ea17eb7cc06529c3695751af6d5f9ea8e41
+  - id: R-004
+    source: kernel.yaml:guardrails
+    status: immutable
+    rule: No Task tool subagent to bypass Formation
+    evidence: Formation provides prompt assembly, log, and principles injection that raw Agent tool lacks
+    version: 1
+    created: "2026-04-03"
+    performance_metric: null
+    immutable_hash: 0fa2508db75e016a16ffaf55a9b3d7239f1fa4a107a2b1e39db8e60338e2c76e
+  - id: R-005
+    source: kernel.yaml:guardrails
+    status: immutable
+    rule: No parallel dispatch to same project (git conflicts + agent stall)
+    evidence: Multiple agents writing to same repo cause merge conflicts and git stalls
+    version: 1
+    created: "2026-04-03"
+    performance_metric: null
+    immutable_hash: 0a196938b91e60ba8762312e590eba66b411cf0caf4d544d147d4e8b7979d3df
+  - id: R-006
+    source: kernel.yaml:guardrails
+    status: immutable
+    rule: No conservative mode from cost concerns
+    evidence: Mindset principle — cost is irrelevant, only perfect execution matters
+    version: 1
+    created: "2026-04-03"
+    performance_metric: null
+    immutable_hash: f28cda5cae3a94f826569ab018197935fe32131452e4940c2bfd9f0ec65c332d
+  - id: R-007
+    source: kernel.yaml:guardrails
+    status: immutable
+    rule: No shortcuts — always fix root cause
+    evidence: Mindset principle — fastest solution != correct solution
+    version: 1
+    created: "2026-04-03"
+    performance_metric: null
+    immutable_hash: 1bdc58fe620767e832092e08dd9df871288cc810490b6ee3a5c133dcf08242ec
+  - id: R-008
+    source: kernel.yaml:guardrails
+    status: immutable
+    rule: Expose problems immediately — never hide or defer
+    evidence: Core principle — hidden problems compound
+    version: 1
+    created: "2026-04-03"
+    performance_metric: null
+    immutable_hash: e934cc63e29da1c9ab9604918a37c8a440b9a084c2287c46b5841a7424dddfab
+  - id: R-009
+    source: kernel.yaml:guardrails
+    status: immutable
+    rule: AskUserQuestion for ALL questions requiring user response
+    evidence: User safety boundary — ensures user always has control
+    version: 1
+    created: "2026-04-03"
+    performance_metric: null
+    immutable_hash: 4759ccfb6ee3e493a1ff9360b771d03a89a23333505e44a619357661a6a89531
+  - id: R-010
+    source: kernel.yaml:guardrails
+    status: immutable
+    rule: Structure > estimate | Verifiable > explainable | Real dep > perceived
+    evidence: Core reasoning principles — the three pillars of decision quality
+    version: 1
+    created: "2026-04-03"
+    performance_metric: null
+    immutable_hash: f61f09bc462711178f34ec270f1bf8a0e09f435113fe3a5cd6dfce621c990e68
+  - id: R-011
+    source: kernel.yaml:guardrails
+    status: immutable
+    rule: "MEM-LANG-001: All memory writes in English (FTS5 does not handle CJK)"
+    evidence: Technical constraint — SQLite FTS5 tokenizer returns 0 results for CJK content
+    version: 1
+    created: "2026-04-03"
+    performance_metric: null
+    immutable_hash: 4930b7e25d8b0281847bb4631d8daaf76407cc3d4d59c7000159429207b40195
+  - id: R-012
+    source: kernel.yaml:guardrails
+    status: immutable
+    rule: Synapse pulse before modifying any symbol (SYN-001~007)
+    evidence: Prevents blind writes — FAIL-009 documents agents skipping pulse and breaking code
+    version: 1
+    created: "2026-04-03"
+    performance_metric: null
+    immutable_hash: 7087228b14b56f6aa0fbf69285d659a9401cfde33e50ca1d881ad884338e07e2
+  - id: R-013
+    source: kernel.yaml:guardrails
+    status: immutable
+    rule: Sub-projects HAVE Synapse indexes — sigil_count=0 does NOT mean no Synapse
+    evidence: Agents incorrectly skipped Synapse when count was 0 — pulse/recall still work
+    version: 1
+    created: "2026-04-03"
+    performance_metric: null
+    immutable_hash: 68e598c6d6699c4b4b371a7d5e97085ea4600edd65e59efbc9ff904ce18399cb
+  - id: R-014
+    source: kernel.yaml:guardrails
+    status: immutable
+    rule: "HALF-DONE IS NOT DONE: Phase 9 must grep EVERY file mentioned in requirements"
+    evidence: Agents consistently completed code files but skipped spec/YAML files
+    version: 1
+    created: "2026-04-03"
+    performance_metric: null
+    immutable_hash: 29f26dbf608e2c6c043a9923ce3c53b71267771785e0c53f26a7e735ae4f82f6
+  - id: R-100
+    source: kernel.yaml:guardrails
+    status: active
+    rule: "TASK-SPLIT: Code changes and spec/YAML/behavioral changes must be SEPARATE tasks"
+    evidence: Agents complete code parts and skip spec parts when mixed in one task
+    version: 1
+    created: "2026-04-03"
+    performance_metric: null
+  - id: R-101
+    source: kernel.yaml:guardrails
+    status: active
+    rule: "ACCEPTANCE-GREP: Every acceptance criterion must be grep-verifiable"
+    evidence: Non-grep criteria (e.g. 'banner shows X') cannot be automated in Phase 9
+    version: 1
+    created: "2026-04-03"
+    performance_metric: null
+  - id: R-102
+    source: kernel.yaml:guardrails
+    status: active
+    rule: "PHASE0-LEAN: phase_0.yaml checklist items max 3-4 lines; schema details go in separate files"
+    evidence: Phase 0 boot time increases with embedded schemas — lean checklist keeps startup fast
+    version: 1
+    created: "2026-04-03"
+    performance_metric: null
+  - id: R-202
+    source: session:2026-04-03 (TASK_935 design)
+    status: active
+    rule: "Mem0 protocol for session insights: recall existing runes before inscribing — MERGE if overlap >70%, flag contradiction, ADD if new, DELETE if invalidated"
+    evidence: Without this protocol, duplicate and contradictory runes accumulate. This is a practical process, not derivable from principles alone.
+    version: 2
+    created: "2026-04-03"
+    performance_metric: null
+  - id: R-301
+    source: session:2026-04-09 (TASK_1088 audit, user request)
+    status: active
+    rule: "PROACTIVE-LESSON-CAPTURE: When you observe a non-obvious technical fact during a session — bug root cause, surprising tool behavior, infrastructure quirk, workflow trap, performance gotcha — inscribe it as an ash rune IMMEDIATELY without waiting for user prompt."
+    evidence: |
+      User request 2026-04-09: 「讓之後的Ansuz都能主動紀下教訓」after manually requesting two lesson inscriptions for lint-staged 16 CRLF parsing bug and task-api complete branch switching. The implicit cost of not capturing: every future session re-discovers the same gotchas from scratch.
+
+      Triggering signals (any one is enough — capture immediately, do not batch until end of session):
+      - You traced a bug to a root cause that was NOT obvious from the symptom
+      - A tool/API behaved in a way that contradicted its docs, name, or your expectation
+      - An infrastructure quirk (CI, hooks, build tooling, git, OS-specific) blocked or distorted an operation
+      - A calling sequence produced unexpected state (e.g., HEAD on wrong branch, file in wrong directory)
+      - You observed a performance gotcha (N+1, slow path, hot loop) that took >1 round of investigation to find
+
+      Inscription protocol (mandatory):
+      1. Identify the closest STABLE existing code symbol as the rune anchor. Bare file paths (.gitattributes, README.md) are weak anchors — they may not be indexed as Symbol nodes in KuzuDB and recall can fail silently. Prefer file:function:kind format pointing at a real function/class/method that exists in the indexed codebase. If unsure, run synapse_search first to find a valid anchor.
+      2. synapse_inscribe with type=ash, confidence>=0.85, content in ENGLISH (MEM-LANG-001 — FTS5 cannot tokenize CJK reliably).
+      3. ALWAYS include trigger_phrases (3-5 natural language phrases) for semantic recall. Without them, recall via paraphrase fails.
+      4. WRITE-VERIFY: after inscribe, run synapse_recall with a unique substring from the rune content. If result_count=0, the rune may be in an unindexed state — re-inscribe with a different (verified) symbol_id. Document the failed rune_id in the conversation as a fallback record.
+      5. If repeated verification fails, fall back to writing the lesson into a structured log under .kenaz/specialist/lessons_pending_index.jsonl so a future session can re-attempt inscription.
+
+      Anti-patterns:
+      - Inscribing on a non-existent symbol_id (e.g., bare ".gitattributes" with no Symbol node) — silent index failure
+      - Skipping write-verify ("trust the success message") — discovered 2026-04-09 that inscribe API returns success even when FTS5 indexing does not happen
+      - Chinese/CJK content in rune body — FTS5 returns 0 results for CJK queries
+      - Generic content ("we hit a bug, fixed it") — provides no recall value
+      - Batching all lessons until Phase 11 — context can compact mid-session and lose them
+    version: 1
+    created: "2026-04-09"
+    performance_metric: null
+  - id: R-302
+    source: session:2026-04-21 (Nexus-T1 absorption TASK_1330-1343)
+    status: active
+    rule: "THOUGHT-STACK-PROACTIVE: Push a thought onto synapse_thought_push BEFORE a mid-session context switch. Trigger contexts: (a) user redirects mid-work to a tangent — push current work state as anchor; (b) you notice a bug/concern but can't fix now — push 'investigate X later'; (c) you reach a sub-fatigue threshold (context heavy but fatigue_warning not yet fired) — push 'resume point Y'. Always bind related_task_id when the thought concerns an open task — task completion auto-pops the thought, closing B-001 parent-lifecycle gap at the tactical layer."
+    evidence: |
+      Tool chain live-verified 2026-04-21 (TASK_1330+TASK_1338+TASK_1339+TASK_1340). push→list→pop round-trip works; .kenaz/specialist/ansuz_thoughts_stack.json auto-generated with 7d expiry. Phase 0 Tier 3th auto-displays stack top on wake.
+
+      Usage protocol:
+      1. Call synapse_thought_push(text=<short anchor>, context=<optional rationale>, related_task_id=<TASK_XXX if applicable>, projectPath=PROJECT_PATH)
+      2. Max stack depth=5, 7-day expiry. Overflow drops oldest.
+      3. Do NOT push trivial or ephemeral notes (one-turn scratch thoughts) — reserve for genuine cross-turn anchors.
+      4. At Phase 11 (回報), check synapse_thought_stack_list — if stack has stale thoughts unrelated to completed work, prompt user whether to pop them.
+
+      This rule closes a gap between ansuz_session_handoff.json (single-slot fatigue handoff) and commitment_layer.yaml (multi-session strategic). Thought Stack is the tactical middle: minutes-to-hours interrupt memory.
+    version: 1
+    created: "2026-04-21"
+    performance_metric: null
+  - id: R-303
+    source: session:2026-04-21 (Nexus-T1 absorption TASK_1332 CC auto-memory indexing)
+    status: active
+    rule: "CC-AUTO-MEMORY-RECALL: When user references a MEMORY.md entry by title (e.g., 'strategic retreat pattern', 'projection over wire compression'), OR Phase 1 requirement analysis produces keywords that match known memory file stems, synapse_recall with those keywords as query. ~75 personal memory files at ~/.claude/projects/<encoded>/memory/*.md are now indexed in Synapse (source='claude_auto_memory', 86 runes live). Top-of-MEMORY.md index is auto-loaded into context but individual file contents require explicit recall."
+    evidence: |
+      TASK_1341 extended scanner.rs to walk resolve_claude_projects_dir(project)/memory/*.md after main project walk. 86 runes tagged source='claude_auto_memory' verified in memory.db 2026-04-21. FTS5 MATCH works; direct cypher MATCH on Rune table fails (runes live in SQLite, not KuzuDB).
+
+      Usage triggers:
+      1. User says 'remember the X pattern' or references a memory by title → synapse_recall(query=X pattern)
+      2. Phase 2 background research on topic that MEMORY.md mentions → synapse_recall(query=<topic>)
+      3. Before making recommendations that touch user-preference territory (branch strategy, dispatch mode, verification strictness) → synapse_recall(query=<topic>) to check for recorded feedback
+      4. Observed gap: synapse_recall ranking currently does NOT boost claude_auto_memory source — TASK completion runes may outrank actual memory content. If top result is irrelevant task-completed rune, scan further down the results; do not give up at rank 1.
+
+      MEMORY.md line-1 preview is always loaded; but line details, rationale, and follow-ups live in individual files. Treat MEMORY.md as index, synapse_recall as content fetch.
+    version: 1
+    created: "2026-04-21"
+    performance_metric: null
+  - id: R-304
+    source: session:2026-04-21 (Nexus-T1 absorption TASK_1334+1342+1343 Phase 0 git enrichment)
+    status: active
+    rule: "PHASE0-GIT-CONTEXT-CONSUMPTION: At Phase 0 wake, phase-gate auto-injects output.git_context with recent_commits (up to 10), wt_diff_stat, since_iso, since_source. If git_context.recent_commits.length > 0, internalize: (a) what moved since last session — do NOT re-propose work that already landed; (b) working-tree diff magnitude — if files_changed > 20, flag merge conflict risk before dispatching; (c) since_source='last_session' means Hippocampus episode succeeded, 'fallback_24h' means last-session query returned nothing (first session or episode loss)."
+    evidence: |
+      TASK_1342 implemented loadGitContext helper in phase-gate.js; TASK_1343 added Phase 0 Tier 3gi yaml render. Verified 2026-04-21: git_context present=true, 10 commits returned, wt_diff_stat live (files_changed:10, insertions:212, deletions:74), since_source='fallback_24h' when no Ansuz episode existed.
+
+      Usage protocol:
+      1. Read output.git_context internally at Phase 0 — do not display all 10 commits in banner (PHASE0-LEAN).
+      2. Before proposing a fix/feature, check recent_commits for 'fix' or task-ids that may have already addressed it.
+      3. If wt_diff_stat shows unusual magnitude (>500 lines changed), pause dispatch plan — investigate whether WIP is user-managed or orphan from prior crashed session.
+      4. since_source distinguishes normal boot ('last_session') from cold-start ('fallback_24h'). Cold-start sessions should lower confidence on 'I know what we were doing'.
+    version: 1
+    created: "2026-04-21"
+    performance_metric: null

package/core/.kenaz/specialist/ansuz_session_handoff.schema.json

@@ -0,0 +1,128 @@
+{
+  "$schema": "http://json-schema.org/draft-07/schema#",
+  "$id": "ansuz_session_handoff.schema.json",
+  "title": "Ansuz Session ISBAR Handoff",
+  "description": "Per-session handoff slot for session context preservation (.kenaz/specialist/handoffs/{session_id}.json — M3/TASK_1760; one slot per session). Written on fatigue_warning and live milestones (dispatch/phase-exit/decision); the next session boot reads the latest slot and archives to Hippocampus. Hippocampus is the archive, not the filesystem. The singular ansuz_session_handoff.json slot is retired read-only legacy.",
+  "type": "object",
+  "required": ["schema_version", "session_id", "triggered_by", "generated_at", "isbar"],
+  "properties": {
+    "schema_version": {
+      "type": "integer",
+      "const": 1,
+      "description": "Schema version for forward compatibility."
+    },
+    "session_id": {
+      "type": "string",
+      "description": "Unique identifier for the session that generated this handoff. Format: ansuz-{timestamp_ms}."
+    },
+    "triggered_by": {
+      "type": "string",
+      "enum": ["fatigue_warning"],
+      "description": "What triggered the handoff generation."
+    },
+    "generated_at": {
+      "type": "string",
+      "format": "date-time",
+      "description": "ISO-8601 timestamp when the handoff was generated."
+    },
+    "fatigue_potential": {
+      "type": "number",
+      "minimum": 0,
+      "maximum": 2,
+      "description": "Fatigue potential value from muninn at time of handoff (typically >= 1.0 when fired)."
+    },
+    "context_size_kb": {
+      "type": "integer",
+      "minimum": 0,
+      "description": "Estimated conversation context size in kilobytes at time of handoff."
+    },
+    "isbar": {
+      "type": "object",
+      "description": "ISBAR (Identification / Situation / Background / Assessment / Recommendation) handoff payload.",
+      "required": ["identification", "situation", "background", "assessment", "recommendation"],
+      "properties": {
+        "identification": {
+          "type": "object",
+          "description": "Who is handing off, and what triggered it.",
+          "required": ["session_id", "fatigue_potential", "context_size_kb"],
+          "properties": {
+            "session_id": { "type": "string" },
+            "fatigue_potential": { "type": "number" },
+            "context_size_kb": { "type": "integer" }
+          }
+        },
+        "situation": {
+          "type": "object",
+          "description": "What is happening right now.",
+          "required": ["current_topic", "pending_decisions"],
+          "properties": {
+            "current_topic": {
+              "type": "string",
+              "description": "One-line description of the active work topic."
+            },
+            "pending_decisions": {
+              "type": "array",
+              "items": { "type": "string" },
+              "description": "Open questions or choices not yet resolved — next session must address these."
+            }
+          }
+        },
+        "background": {
+          "type": "object",
+          "description": "Relevant history and intent.",
+          "required": ["key_decisions", "user_intent"],
+          "properties": {
+            "key_decisions": {
+              "type": "array",
+              "items": { "type": "string" },
+              "description": "Decisions made this session with brief rationale (e.g. 'Chose approach A over B because X')."
+            },
+            "user_intent": {
+              "type": "string",
+              "description": "What the user is ultimately trying to achieve — the north star goal."
+            }
+          }
+        },
+        "assessment": {
+          "type": "object",
+          "description": "Why are we handing off, and what's left.",
+          "required": ["why_handing_off", "unfinished"],
+          "properties": {
+            "why_handing_off": {
+              "type": "string",
+              "description": "Reason for handoff (typically fatigue/context size threshold reached)."
+            },
+            "unfinished": {
+              "type": "array",
+              "items": { "type": "string" },
+              "description": "Items started but not completed — risks if not addressed next session."
+            }
+          }
+        },
+        "recommendation": {
+          "type": "object",
+          "description": "Guidance for the next session.",
+          "required": ["first_action", "context_to_restore"],
+          "properties": {
+            "first_action": {
+              "type": "string",
+              "description": "The single most important thing the next session should do first. Used as Hippocampus episode lesson."
+            },
+            "context_to_restore": {
+              "type": "string",
+              "description": "Minimal context summary (3-5 key facts) needed to continue without re-reading the whole conversation."
+            }
+          }
+        }
+      }
+    }
+  },
+  "_notes": {
+    "active_slot_path": ".kenaz/specialist/handoffs/{session_id}.json — one slot per session (M3/TASK_1760); the retired singular ansuz_session_handoff.json is read-only legacy fallback",
+    "archive_mechanism": "Hippocampus episodic memory — scope: session-handoff:{session_id}",
+    "phase_0_action": "Tier 3x reads active slot, archives to Hippocampus, deletes active slot",
+    "fatigue_trigger": "kernel.yaml fatigue_warning_reaction — TASK_943 delivers the warning signal",
+    "language_rule": "MEM-LANG-001: All Hippocampus content must be in English (FTS5 does not handle CJK)",
+    "recall_query": "hippocampus_episode_recall(query='session handoff') returns all archived sessions"
+  }
+}

package/core/.kenaz/specialist/chart_TASK_1657.json

@@ -0,0 +1,50 @@
+{
+  "task_id": "TASK_1657",
+  "progress_notes": [
+    {
+      "task_id": "TASK_1657",
+      "timestamp": "2026-07-04T05:45:35.438Z",
+      "changes": {
+        "files_created": [
+          ".kenaz/synergy_v3_measure.js",
+          ".kenaz/synergy_v3_result.json"
+        ],
+        "files_modified": [
+          "dashboard/src-tauri/src/features/synapse/thalamus_synergy.rs"
+        ],
+        "files_deleted": [],
+        "total_lines": "+280 thalamus_synergy.rs, +180 synergy_v3_measure.js"
+      },
+      "acceptance_criteria": [
+        {
+          "criterion": "REAL measured baseline synergy number from live data, with n_obs and held-out split",
+          "status": "pass",
+          "evidence": "synergy=-0.00055, n_obs_test=957291, n_positives=8763, 70/30 time split on 461 commits. verdict=PURE_AGGREGATION. Runner: .kenaz/synergy_v3_measure.js on live .synapse/db/memory.db"
+        },
+        {
+          "criterion": "cluster feature provably NOT derived from target: in_cluster_prior from prior window, target = next-window real edits",
+          "status": "pass",
+          "evidence": "in_cluster_prior=[T-DELTA,T) ONLY; editedNext=[T,T+DELTA) DISJOINT. Candidate set = histSyms seen strictly before T (eliminates target-only selection bias). Sanity check: inCluster=1 base rate 11.3% vs inCluster=0 at 0.5% -- raw difference explained by potential collinearity after controls. Code comment in SynergyObsV3 struct + module doc explicitly state the anti-leak contract."
+        },
+        {
+          "criterion": "Baseline marginal includes static co-change prior AND matches joint model param count + regularization; evaluation on held-out time windows",
+          "status": "pass",
+          "evidence": "marginal: 3 params (intercept+potential+cochange_prior), L2=0.01. joint: 4 params (+in_cluster_prior), L2=0.01. Weights fit on first 70% of commits; synergy evaluated on last 30% only. Note: co_change table not in SQLite (Rust in-memory); used Jaccard co-edit density as proxy (documented)."
+        },
+        {
+          "criterion": "cargo check green; feedback gate OFF; NOT merged to main. Synapse rune inscribed.",
+          "status": "partial",
+          "evidence": "CLUSTER_FEEDBACK_ENABLED=false verified. NOT merged (worktree only). Rune rune_477ace5065ed4168bc3c5d54d2e02aa5 inscribed with related_rune_ids=[rune_777498aa]. cargo check in progress (deps still compiling, short CARGO_TARGET_DIR=C:/t/wt1657)."
+        }
+      ],
+      "findings": "v3 metric design fixes both v1 leak (in_cluster from fired_ids) and v2 bug (recency-not-jointness). Critical discovery: potential and in_cluster_prior are nearly perfectly collinear (inCluster=1 <==> potential>0), so marginal model absorbs cluster effect entirely. After controlling for potential+cochange, b_cluster=-0.08 (near zero). synergy=-0.00055 = PURE_AGGREGATION. The thalamus adds no binding beyond aggregation at current data volume. CLUSTER_FEEDBACK_ENABLED stays false.",
+      "deviations": "co_change table not in SQLite (Rust in-memory only). Used Jaccard co-edit density as proxy. Documented in module doc and result JSON. Node.js offline runner used instead of Rust test (cargo check still in progress due to onnxruntime/CMake env constraints).",
+      "cognitive_output": {
+        "runes": [
+          "rune_477ace5065ed4168bc3c5d54d2e02aa5"
+        ],
+        "sigils": []
+      }
+    }
+  ]
+}