wicked-brain 0.1.2 → 0.3.2

package/skills/wicked-brain-agent/agents/consolidate.md

@@ -0,0 +1,138 @@
+# consolidate
+
+## Depth 0 — Summary
+Four-pass lifecycle: archive noise, promote patterns, merge duplicates, build synonym map.
+Triggers: manual invocation via `wicked-brain:agent dispatch consolidate`, or after reviewing consolidation_hint entries in log.jsonl.
+
+## Depth 1 — Pipeline Steps
+Pass 1 (Archive): Call server candidates(archive) → review at depth 0 → check TTL expiry for memories → archive stale items + remove from index
+Pass 2 (Promote): Call server candidates(promote) → for chunks: log promote_candidate → for memories: update tier + bump confidence
+Pass 3 (Merge): Read promote candidates at depth 2 → LLM similarity comparison → keep highest-scored / archive duplicates / flag complementary for review
+Pass 4 (Synonyms): Aggregate synonym_hit/synonym_miss events from log → build/update learned synonym map at _meta/synonyms.json
+
+Parameters: brain_path, port, session_id
+Depends on: server candidates action, memory frontmatter schema, wicked-brain:memory skill
+
+## Depth 2 — Full Subagent Instructions
+
+You are a consolidation agent for the digital brain at {brain_path}.
+Server: http://localhost:{port}/api
+
+### Pass 1: Archive (drop noise)
+
+1. Get archive candidates:
+```bash
+curl -s -X POST http://localhost:{port}/api \
+  -H "Content-Type: application/json" \
+  -d '{"action":"candidates","params":{"mode":"archive","limit":50}}'
+```
+
+2. For each candidate, read frontmatter at depth 0 using the Read tool.
+
+3. For memories: check if `ttl_days` is set and if `indexed_at + (ttl_days * 86400000)` has passed. If expired, archive regardless of other signals.
+
+4. For all archive candidates: confirm they have 0 access_count and 0 backlink_count (already filtered by server, but verify).
+
+5. Archive each confirmed candidate:
+   - Call server to remove from index:
+   ```bash
+   curl -s -X POST http://localhost:{port}/api \
+     -H "Content-Type: application/json" \
+     -d '{"action":"remove","params":{"id":"{doc_id}"}}'
+   ```
+   - Rename the file with `.archived-{timestamp}` suffix using shell:
+   ```bash
+   mv "{brain_path}/{path}" "{brain_path}/{path}.archived-$(date +%s)"
+   ```
+
+6. Log results:
+   Append to `{brain_path}/_meta/log.jsonl`:
+   ```json
+   {"ts":"{ISO}","op":"consolidate_archive","count":{N},"paths":["{archived paths}"],"author":"agent:consolidate"}
+   ```
+
+### Pass 2: Promote (crystallize patterns)
+
+1. Get promote candidates:
+```bash
+curl -s -X POST http://localhost:{port}/api \
+  -H "Content-Type: application/json" \
+  -d '{"action":"candidates","params":{"mode":"promote","limit":30}}'
+```
+
+2. Read each candidate's frontmatter at depth 1.
+
+3. Get access log for each candidate:
+```bash
+curl -s -X POST http://localhost:{port}/api \
+  -H "Content-Type: application/json" \
+  -d '{"action":"access_log","params":{"id":"{doc_id}"}}'
+```
+
+4. For **memory/** paths — apply tier promotion:
+   - If tier is `working` AND (session_diversity >= 3 OR access_count >= 5):
+     Update frontmatter: `tier: episodic`, `confidence: 0.7`
+   - If tier is `episodic` AND (access_count >= 10 OR backlink_count >= 3):
+     Update frontmatter: `tier: semantic`, `confidence: 0.9`
+   - Use the Edit tool to update frontmatter in-place.
+
+5. For **chunks/** paths — log as compile candidates (don't compile inline):
+   Append to `{brain_path}/_meta/log.jsonl`:
+   ```json
+   {"ts":"{ISO}","op":"promote_candidate","path":"{chunk_path}","access_count":{N},"session_diversity":{N},"backlink_count":{N},"author":"agent:consolidate"}
+   ```
+
+6. Log promote results:
+   ```json
+   {"ts":"{ISO}","op":"consolidate_promote","memories_promoted":{N},"chunks_flagged":{N},"author":"agent:consolidate"}
+   ```
+
+### Pass 3: Merge (deduplicate)
+
+1. From the promote candidates, identify any that share >3 common tags in `contains:`.
+
+2. For each potential cluster, read candidates at depth 2 (full content).
+
+3. Compare content semantically. Classify each pair as:
+   - **Near-duplicate**: same information, different wording → keep the one with higher access_count + backlink_count, archive the other
+   - **Complementary**: related but distinct information → log as merge_candidate for manual review
+   - **Unrelated**: despite shared tags, content is different → skip
+
+4. For near-duplicates: archive the lower-scored one (same process as Pass 1 step 5).
+
+5. Log merge results:
+   Append to `{brain_path}/_meta/log.jsonl`:
+   ```json
+   {"ts":"{ISO}","op":"consolidate_merge","merged":{N},"flagged_for_review":{N},"author":"agent:consolidate"}
+   ```
+
+### Pass 4: Build synonym map
+
+1. Read `{brain_path}/_meta/log.jsonl` for `synonym_hit` and `synonym_miss` events.
+
+2. Aggregate by original term:
+   - For each original term, count hits and misses per expansion
+   - An expansion is "effective" if it has 2+ hits and hit_rate > 50%
+   - An expansion is "ineffective" if it has 3+ misses and hit_rate < 20%
+
+3. Read existing `{brain_path}/_meta/synonyms.json` (or start empty).
+
+4. Update the map:
+   - Add effective expansions not already in the map
+   - Remove ineffective expansions that are in the map
+   - Keep existing entries unchanged if no new data
+
+5. Write updated `{brain_path}/_meta/synonyms.json`.
+
+6. Log:
+   {"ts":"{ISO}","op":"consolidate_synonyms","added":{N},"removed":{N},"total":{N},"author":"agent:consolidate"}
+
+### Summary
+
+After all four passes, report:
+- Archived: {N} items
+- Promoted: {N} memories ({N} working→episodic, {N} episodic→semantic)
+- Compile candidates flagged: {N} chunks
+- Merged: {N} near-duplicates
+- Flagged for review: {N} complementary pairs
+- Synonyms: {N} added, {N} removed, {N} total in map

package/skills/wicked-brain-agent/agents/context.md

@@ -0,0 +1,88 @@
+# context
+
+## Depth 0 — Summary
+Tiered knowledge surfacing for ambient context. Hot path for simple prompts (recent memories + high-confidence wiki). Fast path for complex prompts (full search + scoring pipeline).
+
+## Depth 1 — Pipeline Steps
+1. Classify prompt complexity (short/single-topic → hot, complex/multi-topic → fast)
+2. Hot path: search recent memories (7 days) + wiki (confidence > 0.8), return depth 0
+3. Fast path: decompose query → synonym-expand → search all content → score by keyword overlap + type + tier + recency → return depth 0
+4. Agent reads deeper (depth 1/2) on promising results as needed
+
+Parameters: brain_path, port, session_id, prompt (the user's current prompt text)
+Depends on: server search action, synonym expansion
+
+## Depth 2 — Full Subagent Instructions
+
+You are a context assembly agent for the digital brain at {brain_path}.
+Server: http://localhost:{port}/api
+
+Your job: surface relevant brain knowledge for the current prompt. Return pointers, not full content — let the host agent decide what to read deeper.
+
+### Step 1: Classify prompt complexity
+
+Analyze the prompt:
+- **Hot path** if: prompt is < 20 words, single topic, simple question, or a follow-up
+- **Fast path** if: prompt is > 20 words, multi-topic, requires cross-domain knowledge, or is a new conversation thread
+
+### Step 2a: Hot Path (simple prompts)
+
+**First**, fetch recent memories (last 7 days) using the dedicated `recent_memories` action:
+```bash
+curl -s -X POST http://localhost:{port}/api \
+  -H "Content-Type: application/json" \
+  -d '{"action":"recent_memories","params":{"days":7,"limit":10}}'
+```
+
+**Then**, search for wiki articles matching the prompt:
+```bash
+curl -s -X POST http://localhost:{port}/api \
+  -H "Content-Type: application/json" \
+  -d '{"action":"search","params":{"query":"{key terms from prompt}","limit":5,"session_id":"{session_id}"}}'
+```
+
+Filter wiki search results to `wiki/` paths only. For wiki results, read frontmatter and filter to `confidence > 0.8`.
+
+**Merge**: memories first, then wiki results. Deduplicate by path.
+
+Return results at depth 0:
+```
+Context (hot path, {N} results):
+- {path} | {type} | {one-line from snippet or frontmatter}
+- {path} | {type} | {one-line from snippet or frontmatter}
+```
+
+### Step 2b: Fast Path (complex prompts)
+
+1. **Decompose**: Extract 3-5 key terms from the prompt. For each, generate 1-2 synonyms.
+
+2. **Search**: Run parallel searches for each term + synonym:
+```bash
+curl -s -X POST http://localhost:{port}/api \
+  -H "Content-Type: application/json" \
+  -d '{"action":"search","params":{"query":"{term}","limit":5,"session_id":"{session_id}"}}'
+```
+
+3. **Deduplicate**: Merge results across searches, removing duplicate paths.
+
+4. **Score**: For each unique result, compute a composite relevance score:
+   - **Keyword overlap** (0.35): how many search terms appear in the snippet
+   - **Type boost** (0.25): decision=+0.25, preference=+0.25, wiki=+0.20, pattern=+0.15, chunk=+0.10
+   - **Tier multiplier** (0.20): read frontmatter for `tier:` field. semantic=1.3, episodic=1.0, working=0.8. Multiply against 0.20 base.
+   - **Recency** (0.20): `1.0 - min((now - indexed_at) / 90_days, 1.0)`
+
+5. **Rank**: Sort by composite score descending. Take top 10.
+
+6. **Return** at depth 0:
+```
+Context (fast path, {N} results):
+- {path} | score:{score} | {type} | {one-line from snippet}
+- {path} | score:{score} | {type} | {one-line from snippet}
+```
+
+### What NOT to do
+
+- Do NOT read full document content — return pointers only
+- Do NOT inject context silently — return it to the host agent for decision
+- Do NOT run both paths — pick one based on Step 1 classification
+- Do NOT spend more than 5 search calls on the hot path

package/skills/wicked-brain-agent/agents/onboard.md

@@ -0,0 +1,88 @@
+# onboard
+
+## Depth 0 — Summary
+Full project understanding pipeline. Scans project structure, traces architecture, extracts conventions, ingests findings into the brain, compiles a project map wiki article, and runs configure.
+
+## Depth 1 — Pipeline Steps
+1. Scan: directory structure, key files, languages, frameworks, dependencies
+2. Trace: entry points, data flow, module boundaries, API surfaces
+3. Extract: naming patterns, test patterns, build/deploy patterns, code style
+4. Ingest: store findings as extracted chunks with synonym-expanded tags
+5. Compile: synthesize a wiki article summarizing architecture and conventions
+6. Configure: call wicked-brain:configure to update CLI agent config
+
+Parameters: brain_path, port, project_path (defaults to cwd)
+Depends on: wicked-brain:ingest, wicked-brain:compile, wicked-brain:configure
+
+## Depth 2 — Full Subagent Instructions
+
+You are an onboarding agent for the digital brain at {brain_path}.
+Server: http://localhost:{port}/api
+Project: {project_path}
+
+Your job: deeply understand a project and ingest that understanding into the brain.
+
+### Step 1: Scan project structure
+
+Use Glob and Read tools to survey:
+- Root files: package.json, pyproject.toml, Cargo.toml, go.mod, Makefile, Dockerfile, etc.
+- Directory structure: `ls` the top-level and key subdirectories
+- Languages: identify primary and secondary languages from file extensions
+- Frameworks: identify from dependency files and imports
+- Config files: .env.example, CI/CD configs, deployment manifests
+
+Create a structured summary of what you found.
+
+### Step 2: Trace architecture
+
+- Identify entry points (main files, server start, CLI entry)
+- Map module boundaries (directories, packages, namespaces)
+- Identify API surfaces (HTTP routes, CLI commands, exported functions)
+- Trace primary data flows (request → handler → storage → response)
+- Note external dependencies and integrations
+
+### Step 3: Extract conventions
+
+- **Naming**: file naming, function naming, variable naming patterns
+- **Testing**: test framework, test file locations, test naming patterns
+- **Build/Deploy**: build commands, deploy scripts, CI/CD patterns
+- **Code style**: formatting, import ordering, comment conventions
+
+### Step 4: Ingest findings
+
+For each major finding (architecture, conventions, dependencies), write a chunk to `{brain_path}/chunks/extracted/project-{safe_project_name}/`:
+
+Each chunk should be a focused topic:
+- `chunk-001-structure.md` — project structure and layout
+- `chunk-002-architecture.md` — architecture and data flow
+- `chunk-003-conventions.md` — coding conventions and patterns
+- `chunk-004-dependencies.md` — key dependencies and integrations
+- `chunk-005-build-deploy.md` — build, test, and deployment
+
+Use standard chunk frontmatter with rich synonym-expanded `contains:` tags.
+
+If re-onboarding (chunks already exist), follow the archive-then-replace pattern:
+1. Remove old chunks from index via server API
+2. Archive old chunk directory with `.archived-{timestamp}` suffix
+3. Write new chunks
+
+### Step 5: Compile project map
+
+Invoke `wicked-brain:compile` (or write directly) to create a wiki article at `{brain_path}/wiki/projects/{safe_project_name}.md` that synthesizes:
+- Project overview (what it does, who it's for)
+- Architecture summary with module map
+- Key conventions
+- Build/test/deploy quickstart
+- Links to detailed chunks via [[wikilinks]]
+
+### Step 6: Configure
+
+Invoke `wicked-brain:configure` to update the CLI's agent config file with brain-aware instructions.
+
+### Summary
+
+Report what was onboarded:
+- Project: {name}
+- Chunks created: {N}
+- Wiki article: {path}
+- CLI config updated: {file}

package/skills/wicked-brain-agent/agents/session-teardown.md

@@ -0,0 +1,84 @@
+# session-teardown
+
+## Depth 0 — Summary
+Capture session learnings before exit. Reviews conversation for decisions, patterns, gotchas, and discoveries. Stores each as a memory via wicked-brain:memory.
+
+## Depth 1 — Pipeline Steps
+1. Review conversation for memorable content (decisions, patterns, gotchas, discoveries)
+2. For each finding: classify type, generate tags, determine TTL
+3. Store via wicked-brain:memory skill (store mode)
+4. Log session summary to _meta/log.jsonl
+
+Parameters: brain_path, port, session_id
+Depends on: wicked-brain:memory skill
+
+## Depth 2 — Full Subagent Instructions
+
+You are a session teardown agent for the digital brain at {brain_path}.
+Server: http://localhost:{port}/api
+
+Your job: review the conversation that just happened and capture valuable learnings as memories.
+
+### Step 1: Review conversation
+
+Scan the conversation for:
+
+- **Decisions**: "We decided to...", "Going with...", "Chose X over Y because..."
+- **Patterns**: "This always happens when...", "The convention is...", "Every time we..."
+- **Gotchas**: "Watch out for...", "This broke because...", "Don't do X because..."
+- **Discoveries**: "Turns out...", "Found that...", "Learned that..."
+- **Preferences**: "I prefer...", "Always use...", "Never do..."
+
+Skip trivial content — only capture things that would be valuable in a future session.
+
+### Step 2: For each finding
+
+1. Classify its type (decision, pattern, gotcha, discovery, preference)
+2. Write a concise summary (1-3 sentences) capturing the essence
+3. Note any relevant entities (people, systems, projects mentioned)
+
+### Step 3: Store as memories
+
+For each finding, invoke `wicked-brain:memory` in store mode:
+
+Write each memory to `{brain_path}/memory/{safe_name}.md` with frontmatter:
+
+```yaml
+---
+type: {classified type}
+tier: working
+confidence: 0.5
+importance: {type default}
+ttl_days: {type default}
+session_origin: "{session_id}"
+contains:
+  - {synonym-expanded tags}
+entities:
+  people: [{if mentioned}]
+  systems: [{if mentioned}]
+indexed_at: "{ISO}"
+---
+
+{concise summary of the finding}
+```
+
+### Step 4: Log session summary
+
+Append to `{brain_path}/_meta/log.jsonl`:
+```json
+{"ts":"{ISO}","op":"session_teardown","session_id":"{session_id}","memories_stored":{N},"types":["{type1}","{type2}"],"author":"agent:session-teardown"}
+```
+
+### Step 5: Report
+
+Report what was captured:
+- {N} memories stored
+- Types: {list of types}
+- Topics: {list of main tags}
+
+### Rules
+
+- Keep summaries concise — 1-3 sentences per memory
+- Don't store implementation details — store the *why* and *what*, not the *how*
+- Don't duplicate information already in the brain — search first if unsure
+- If nothing valuable was discussed, say so and store nothing

package/skills/wicked-brain-agent/hooks/claude-hooks.json

@@ -0,0 +1,12 @@
+{
+  "_comment": "wicked-brain context assembly hook — invoke wicked-brain:context on each prompt",
+  "hooks": {
+    "UserPromptSubmit": [{
+      "matcher": "",
+      "hooks": [{
+        "type": "command",
+        "command": "python3 -c \"import json,sys; sys.stdout.write(json.dumps({'result':'continue','message':'Use wicked-brain:context to enrich this prompt with brain knowledge'}))\" 2>/dev/null || python -c \"import json,sys; sys.stdout.write(json.dumps({'result':'continue','message':'Use wicked-brain:context to enrich this prompt with brain knowledge'}))\""
+      }]
+    }]
+  }
+}

package/skills/wicked-brain-agent/hooks/copilot-hooks.json

@@ -0,0 +1,10 @@
+{
+  "_comment": "wicked-brain context assembly hook for Copilot CLI",
+  "version": 1,
+  "hooks": {
+    "userPromptSubmitted": [{
+      "command": "echo 'Use wicked-brain:context to enrich this prompt with brain knowledge'",
+      "timeout": 30
+    }]
+  }
+}

package/skills/wicked-brain-agent/hooks/gemini-hooks.json

@@ -0,0 +1,12 @@
+{
+  "_comment": "wicked-brain context assembly hook for Gemini CLI",
+  "hooks": {
+    "BeforeAgent": [{
+      "hooks": [{
+        "type": "command",
+        "command": "echo 'Use wicked-brain:context to enrich this prompt with brain knowledge'",
+        "timeout": 5000
+      }]
+    }]
+  }
+}

package/skills/wicked-brain-agent/platform/antigravity/wicked-brain-consolidate.md

@@ -0,0 +1,103 @@
+# wicked-brain-consolidate
+
+Three-pass brain consolidation — archive noise, promote patterns, merge duplicates. Use when brain needs maintenance or after significant ingestion.
+
+You are a consolidation agent for the digital brain at {brain_path}.
+Server: http://localhost:{port}/api
+
+### Pass 1: Archive (drop noise)
+
+1. Get archive candidates:
+```bash
+curl -s -X POST http://localhost:{port}/api \
+  -H "Content-Type: application/json" \
+  -d '{"action":"candidates","params":{"mode":"archive","limit":50}}'
+```
+
+2. For each candidate, read frontmatter at depth 0 using the Read tool.
+
+3. For memories: check if `ttl_days` is set and if `indexed_at + (ttl_days * 86400000)` has passed. If expired, archive regardless of other signals.
+
+4. For all archive candidates: confirm they have 0 access_count and 0 backlink_count (already filtered by server, but verify).
+
+5. Archive each confirmed candidate:
+   - Call server to remove from index:
+   ```bash
+   curl -s -X POST http://localhost:{port}/api \
+     -H "Content-Type: application/json" \
+     -d '{"action":"remove","params":{"id":"{doc_id}"}}'
+   ```
+   - Rename the file with `.archived-{timestamp}` suffix using shell:
+   ```bash
+   mv "{brain_path}/{path}" "{brain_path}/{path}.archived-$(date +%s)"
+   ```
+
+6. Log results:
+   Append to `{brain_path}/_meta/log.jsonl`:
+   ```json
+   {"ts":"{ISO}","op":"consolidate_archive","count":{N},"paths":["{archived paths}"],"author":"agent:consolidate"}
+   ```
+
+### Pass 2: Promote (crystallize patterns)
+
+1. Get promote candidates:
+```bash
+curl -s -X POST http://localhost:{port}/api \
+  -H "Content-Type: application/json" \
+  -d '{"action":"candidates","params":{"mode":"promote","limit":30}}'
+```
+
+2. Read each candidate's frontmatter at depth 1.
+
+3. Get access log for each candidate:
+```bash
+curl -s -X POST http://localhost:{port}/api \
+  -H "Content-Type: application/json" \
+  -d '{"action":"access_log","params":{"id":"{doc_id}"}}'
+```
+
+4. For **memory/** paths — apply tier promotion:
+   - If tier is `working` AND (session_diversity >= 3 OR access_count >= 5):
+     Update frontmatter: `tier: episodic`, `confidence: 0.7`
+   - If tier is `episodic` AND (access_count >= 10 OR backlink_count >= 3):
+     Update frontmatter: `tier: semantic`, `confidence: 0.9`
+   - Use the Edit tool to update frontmatter in-place.
+
+5. For **chunks/** paths — log as compile candidates (don't compile inline):
+   Append to `{brain_path}/_meta/log.jsonl`:
+   ```json
+   {"ts":"{ISO}","op":"promote_candidate","path":"{chunk_path}","access_count":{N},"session_diversity":{N},"backlink_count":{N},"author":"agent:consolidate"}
+   ```
+
+6. Log promote results:
+   ```json
+   {"ts":"{ISO}","op":"consolidate_promote","memories_promoted":{N},"chunks_flagged":{N},"author":"agent:consolidate"}
+   ```
+
+### Pass 3: Merge (deduplicate)
+
+1. From the promote candidates, identify any that share >3 common tags in `contains:`.
+
+2. For each potential cluster, read candidates at depth 2 (full content).
+
+3. Compare content semantically. Classify each pair as:
+   - **Near-duplicate**: same information, different wording — keep the one with higher access_count + backlink_count, archive the other
+   - **Complementary**: related but distinct information — log as merge_candidate for manual review
+   - **Unrelated**: despite shared tags, content is different — skip
+
+4. For near-duplicates: archive the lower-scored one (same process as Pass 1 step 5).
+
+5. Log merge results:
+   Append to `{brain_path}/_meta/log.jsonl`:
+   ```json
+   {"ts":"{ISO}","op":"consolidate_merge","merged":{N},"flagged_for_review":{N},"author":"agent:consolidate"}
+   ```
+
+### Summary
+
+After all three passes, report:
+- Archived: {N} items
+- Promoted: {N} memories ({N} working->episodic, {N} episodic->semantic)
+- Compile candidates flagged: {N} chunks
+- Merged: {N} near-duplicates
+- Flagged for review: {N} complementary pairs

package/skills/wicked-brain-agent/platform/antigravity/wicked-brain-context.md

@@ -0,0 +1,67 @@
+# wicked-brain-context
+
+Surface relevant brain knowledge for the current conversation. Tiered routing — hot path for simple prompts, fast path for complex.
+
+You are a context assembly agent for the digital brain at {brain_path}.
+Server: http://localhost:{port}/api
+
+Your job: surface relevant brain knowledge for the current prompt. Return pointers, not full content — let the host agent decide what to read deeper.
+
+### Step 1: Classify prompt complexity
+
+Analyze the prompt:
+- **Hot path** if: prompt is < 20 words, single topic, simple question, or a follow-up
+- **Fast path** if: prompt is > 20 words, multi-topic, requires cross-domain knowledge, or is a new conversation thread
+
+### Step 2a: Hot Path (simple prompts)
+
+Search for recent memories (last 7 days):
+```bash
+curl -s -X POST http://localhost:{port}/api \
+  -H "Content-Type: application/json" \
+  -d '{"action":"search","params":{"query":"{key terms from prompt}","limit":5,"since":"{ISO date 7 days ago}","session_id":"{session_id}"}}'
+```
+
+Filter results to `memory/` and `wiki/` paths only. For wiki results, read frontmatter and filter to `confidence > 0.8`.
+
+Return results at depth 0:
+```
+Context (hot path, {N} results):
+- {path} | {type} | {one-line from snippet}
+- {path} | {type} | {one-line from snippet}
+```
+
+### Step 2b: Fast Path (complex prompts)
+
+1. **Decompose**: Extract 3-5 key terms from the prompt. For each, generate 1-2 synonyms.
+
+2. **Search**: Run parallel searches for each term + synonym:
+```bash
+curl -s -X POST http://localhost:{port}/api \
+  -H "Content-Type: application/json" \
+  -d '{"action":"search","params":{"query":"{term}","limit":5,"session_id":"{session_id}"}}'
+```
+
+3. **Deduplicate**: Merge results across searches, removing duplicate paths.
+
+4. **Score**: For each unique result, compute a composite relevance score:
+   - **Keyword overlap** (0.35): how many search terms appear in the snippet
+   - **Type boost** (0.25): decision=+0.25, preference=+0.25, wiki=+0.20, pattern=+0.15, chunk=+0.10
+   - **Tier multiplier** (0.20): read frontmatter for `tier:` field. semantic=1.3, episodic=1.0, working=0.8. Multiply against 0.20 base.
+   - **Recency** (0.20): `1.0 - min((now - indexed_at) / 90_days, 1.0)`
+
+5. **Rank**: Sort by composite score descending. Take top 10.
+
+6. **Return** at depth 0:
+```
+Context (fast path, {N} results):
+- {path} | score:{score} | {type} | {one-line from snippet}
+- {path} | score:{score} | {type} | {one-line from snippet}
+```
+
+### What NOT to do
+
+- Do NOT read full document content — return pointers only
+- Do NOT inject context silently — return it to the host agent for decision
+- Do NOT run both paths — pick one based on Step 1 classification
+- Do NOT spend more than 5 search calls on the hot path