Context Curation Command

You are orchestrating a 3-phase context curation pipeline that filters contexts by quality, extracts relevant information, and synthesizes a curated memory context.

Your Role

Coordinate three phases with dynamic scaling:

1. Phase 1: Frontmatter scan + quality gate (1 agent)
2. Phase 2: Full content extraction (N agents, parallel by component)
3. Phase 3: Memory synthesis (simple or parallel, based on scale)

Required Knowledge

This command requires the Context Engineering skill. If not loaded, invoke centauro:context-engineering before proceeding.

Arguments

task (required): Task description for curation quality-threshold (optional): Minimum quality score (0.00-1.00), default: 0.70

Examples:

# Use default threshold (0.70)
/centauro:curate "Add OAuth2 authentication"

# Custom strict threshold
/centauro:curate "Add OAuth2 authentication" --quality-threshold 0.85

# Custom lenient threshold
/centauro:curate "Add OAuth2 authentication" --quality-threshold 0.60

Pre-Flight Validation

Step 1: Validate Arguments

Validate quality-threshold:

If quality-threshold provided:
  - Must be number between 0.00 and 1.00
  - If outside range → Error message with correct range

If not provided:
  - Use default: 0.70 (Grade C minimum)

Example validation:

✅ Arguments validated:
- Task: "Add OAuth2 authentication"
- Quality threshold: 0.70 (Grade C minimum)

Step 2: Context Inventory

Count total contexts:

find .centauro/contexts -name "*.md" | wc -l

Display:

📊 Context Repository Status:
- Total contexts: [N] files
- Directory: .centauro/contexts/

---

PHASE 1: Frontmatter Scan + Quality Gate

Purpose: Token-efficient filtering

Step 1: Launch Context Scanner

Use the Task tool to launch context-scanner agent:

Analyze all contexts using frontmatter-only scan and apply quality gate.

TASK: [user's task description]
QUALITY_THRESHOLD: [validated threshold]
CONTEXT_DIRECTORY: .centauro/contexts/

Process:
1. Scan frontmatter of all .md files in context directories
2. Apply quality gate: Include only contexts where estimated_quality >= [threshold]
3. Assess task relevance for passing contexts
4. Return three lists: PASS, FAIL, NO_QUALITY

Return structured output with:
- PASS list (quality + relevance scores)
- FAIL list (with reasons)
- NO_QUALITY list (missing metadata)
- Machine-readable format for Phase 2 handoff

Step 2: Process Phase 1 Results

Extract from scanner output:

• PASS list: Contexts meeting quality gate
• FAIL list: Contexts below threshold
• NO_QUALITY list: Contexts missing quality metadata

Handle NO_QUALITY contexts:

If NO_QUALITY list is not empty:

⚠️ Quality Gate: [N] contexts missing quality metadata

Contexts without quality scores:
- [file1.md]
- [file2.md]
...

Options:
A) Skip them (safest, fastest) - Proceed without these contexts
B) Load anyway (ignore gate) - Include in Phase 2 without quality check
C) Auto-assess now (adds ~30s) - Run quality assessment first

Your choice? [A/B/C]

Wait for user input. If user chooses C, they should run /centauro:assess-quality first (future command).

Step 3: Display Phase 1 Summary

✅ Phase 1 Complete: Frontmatter Scan

**Quality Gate Results:**
- Threshold: [0.XX] (Grade [X] minimum)
- ✅ PASS: [N] contexts (meet quality + relevant)
- ❌ FAIL: [M] contexts (below threshold)
- ⚠️ NO_QUALITY: [P] contexts (missing metadata)

**Top Passing Contexts:**
1. [file1.md] - Quality: 0.XX, Relevance: 0.XX
2. [file2.md] - Quality: 0.XX, Relevance: 0.XX
3. [file3.md] - Quality: 0.XX, Relevance: 0.XX
... (top 5)

**Proceeding to Phase 2 with [N] contexts**

If PASS list is empty:

❌ No contexts passed quality gate

All contexts failed the quality threshold of [0.XX].

Suggestions:
1. Lower threshold: Try --quality-threshold 0.60
2. Check context quality: Run /centauro:assess-quality
3. Use simpler command: /centauro:load (no quality gate)

Stop execution if PASS list empty.

---

PHASE 2: Full Content Extraction (Parallel)

Purpose: Extract relevant information from quality-approved contexts

Step 1: Organize Contexts by Component

Group PASS list by component type:

c1-instructions: [file1.md, file2.md, ...]
c2-knowledge: [file3.md, file4.md, ...]
c4-memory: [file5.md, ...]
c5-state: [file6.md, ...]

Step 2: Token Budget Check

Estimate Phase 2 token cost:

tokens_per_file = 3000 (average)
total_tokens = PASS_count × 3000

If total_tokens > 50,000:

⚠️ High Token Usage Detected

Phase 2 will process [N] contexts (~[X]K tokens estimated).

This may take longer and consume more resources.

Options:
A) Load all (may be slow, ~[Y] seconds)
B) Load top 10 by relevance only
C) Raise threshold to 0.[higher] for fewer contexts
D) Load only CRITICAL priority contexts

Your choice? [A/B/C/D]

Wait for user input. Adjust PASS list based on choice.

Step 3: Launch Parallel Extraction

Launch N context-manager agents in parallel (one per component type with passing files).

IMPORTANT: Use a single message with multiple Task tool calls for true parallelism.

For each component type with passing files:

Extract relevant information from [component] contexts for the task.

TASK: [user's task description]
COMPONENT: [c1-instructions|c2-knowledge|c4-memory|c5-state]

FILES TO PROCESS:
- path/to/file1.md
- path/to/file2.md
...

For each file:
1. Read full content
2. Extract key methodologies/knowledge/memory/state relevant to task
3. Identify unique insights
4. Summarize in 3-5 bullet points

Return structured summary:
FILE: [path]
SUMMARY: [1-2 sentences]
KEY_POINTS:
- [Point 1]
- [Point 2]
...

Example for 3 component types:

Launch all 3 agents simultaneously in one message:

• Agent 1: c1-instructions contexts
• Agent 2: c2-knowledge contexts
• Agent 3: c5-state contexts

Step 4: Aggregate Phase 2 Results

Collect all extraction summaries from parallel agents.

Display Phase 2 Summary:

✅ Phase 2 Complete: Content Extraction

**Extracts by Component:**
- c1-instructions: [N] contexts extracted
- c2-knowledge: [M] contexts extracted
- c4-memory: [P] contexts extracted
- c5-state: [Q] contexts extracted

**Total:** [X] contexts processed
**Estimated tokens:** ~[Y]K
**Time:** ~[Z] seconds

**Sample extracts:**
→ c1-instructions/auth_workflow.md: "OAuth2 authorization code flow with PKCE..."
→ c2-knowledge/oauth2_guide.md: "4 grant types, auth code most secure..."
... (top 3)

**Proceeding to Phase 3: Memory Synthesis**

---

PHASE 3: Memory Synthesis (Dynamic)

Purpose: Create unified memory context

Step 1: Determine Synthesis Strategy

extracted_contexts = total number of contexts from Phase 2

If extracted_contexts < 15:
  → STRATEGY = "simple"
  → AGENTS = 1 (memory-meta-synthesizer)

If extracted_contexts >= 15:
  → STRATEGY = "parallel"
  → AGENTS = 4 + 1 (component-synthesizers + meta-synthesizer)

Display:

🔄 Phase 3 Strategy: [simple|parallel]
- Extracted contexts: [N]
- Synthesis mode: [simple (1 agent) | parallel (5 agents)]

Step 2a: Simple Synthesis (< 15 contexts)

Launch 1 memory-meta-synthesizer agent:

Create unified memory context from Phase 2 extracts.

MODE: simple
TASK: [user's task description]
QUALITY_THRESHOLD: [threshold used]

EXTRACTS: [All Phase 2 summaries]

Process:
1. Aggregate all extracts
2. Identify cross-component patterns
3. Detect dependencies and relationships
4. Synthesize unified memory context
5. Auto-generate filename: memory_curation_[task-slug]_[timestamp].md
6. Ensure .centauro/contexts/c4-memory/curated/ directory exists
7. Write synthesized memory to curated folder
8. Return file path and confirmation

Include in memory:
- Component summaries (c1, c2, c4, c5)
- Cross-component patterns
- Dependencies
- Contradictions (if any)
- Implementation guidance
- Provenance (source contexts)
- Quality self-assessment

Step 2b: Parallel Synthesis (≥ 15 contexts)

Phase 3a: Launch 4 component-synthesizer agents in parallel

IMPORTANT: Use a single message with multiple Task tool calls.

For each component with extracts:

Synthesize all [component] extracts into consolidated summary.

COMPONENT_TYPE: [c1-instructions|c2-knowledge|c4-memory|c5-state]
TASK: [user's task description]

EXTRACTS: [All Phase 2 summaries for this component]

Process:
1. Read all extracts for this component type
2. Identify patterns (consensus/complementary/unique)
3. Consolidate information
4. Flag contradictions
5. Provide source attribution
6. Return component-level summary

Follow component-specific synthesis structure.

Launch all 4 simultaneously.

Phase 3b: After 3a completes, launch 1 memory-meta-synthesizer agent:

Create unified memory context from component summaries.

MODE: meta
TASK: [user's task description]
QUALITY_THRESHOLD: [threshold used]

COMPONENT_SUMMARIES: [All Phase 3a outputs]

Process:
1. Take pre-synthesized component summaries
2. Identify cross-component patterns
3. Map dependencies and relationships
4. Synthesize holistic memory context
5. Auto-generate filename: memory_curation_[task-slug]_[timestamp].md
6. Ensure .centauro/contexts/c4-memory/curated/ directory exists
7. Write synthesized memory to curated folder
8. Return file path and confirmation

Include in memory:
- Component summaries (from Phase 3a)
- Cross-component patterns
- Dependencies
- Contradictions (if any)
- Implementation guidance
- Provenance (source contexts)
- Quality self-assessment

Step 3: Display Phase 3 Summary

✅ Phase 3 Complete: Memory Synthesis

**Synthesis Mode:** [simple|parallel]
**Component summaries:** [N if parallel, else "direct"]
**Cross-patterns identified:** [X]
**Dependencies mapped:** [Y]

**Memory Context Created:**
📄 `.centauro/contexts/c4-memory/curated/memory_curation_[slug]_[timestamp].md`

---

FINAL OUTPUT: Curation Complete

Present comprehensive summary:

# ✅ Context Curation Complete

## Curation Summary

**Task:** [user's task description]
**Quality Threshold:** [0.XX] (Grade [X] minimum)
**Total Time:** ~[X] seconds

## Pipeline Results

### Phase 1: Frontmatter Scan
- Contexts scanned: [N]
- Passed quality gate: [M]
- Failed quality gate: [P]
- Missing quality metadata: [Q]

### Phase 2: Content Extraction
- Contexts extracted: [M]
- Components involved: [c1, c2, c4, c5]
- Token usage: ~[X]K

### Phase 3: Memory Synthesis
- Mode: [simple|parallel]
- Agents used: [1|5]
- Cross-patterns identified: [Y]

## Curated Memory Context

**Location:** `.centauro/contexts/c4-memory/curated/memory_curation_[slug]_[timestamp].md`

**Content includes:**
- Methodologies from [N] c1 contexts
- Knowledge from [M] c2 contexts
- Historical decisions from [P] c4 contexts
- Current state from [Q] c5 contexts

**Estimated Quality:** Grade [A-F] (self-assessed)

## Next Steps

**You can now:**
1. **View the curated memory:**
   ```bash
   Read .centauro/contexts/c4-memory/curated/memory_curation_[slug]_[timestamp].md

2. Use in future sessions: The curated memory is now a c4-memory context. It will be scanned in future curations.

3. Load for current task: The synthesized information is available for your task: "[task]"

4. Share with team: Share the curated memory file with teammates working on similar tasks.

---

Ready to proceed with your task!

---

## Error Handling

### Invalid Quality Threshold
```markdown
❌ Invalid quality threshold: [value]

Quality threshold must be a number between 0.00 and 1.00.

Examples:
- 0.60 (Grade D+ minimum)
- 0.70 (Grade C minimum) ← default
- 0.80 (Grade B- minimum)
- 0.90 (Grade A- minimum)

Usage: /centauro:curate "task" --quality-threshold 0.80

Context Directory Not Found

❌ Context directory not found: .centauro/contexts/

This command requires a context repository.

Action:
1. Initialize: /centauro:setup
2. Or check directory path

Phase 1 Agent Failed

❌ Phase 1 failed: Context scanner error

Error: [error message from agent]

Action:
- Check .centauro/contexts/ directory exists
- Check contexts have valid YAML frontmatter
- Try: /centauro:load (simpler alternative)

Phase 2 Agent Failed

❌ Phase 2 failed: Extraction error in [component]

Error: [error message from agent]

Partial results available from:
- ✅ c1-instructions: [N] contexts
- ✅ c2-knowledge: [M] contexts
- ❌ c4-memory: Failed

Continue with partial results? (yes/no)

Phase 3 Agent Failed

❌ Phase 3 failed: Synthesis error

Error: [error message from agent]

Extraction completed successfully ([N] contexts).
Synthesis could not complete.

Action:
- Check .centauro/contexts/c4-memory/curated/ permissions
- Phase 2 extracts are available but not synthesized
- You can manually review Phase 2 results

---

Key Principles

Automation First: This command should be highly automated. Only prompt user for critical decisions (NO_QUALITY handling, high token usage).

Clear Communication: Always explain what's happening at each phase. Include progress indicators.

Adaptive Scaling: Automatically choose synthesis strategy based on workload. User doesn't need to think about it.

Graceful Degradation: If some contexts fail, continue with what works. Don't fail completely.

Token Efficiency: Phase 1 reads only frontmatter. Only Phase 2 reads full content of quality-approved contexts.

Parallel Execution: Launch multiple agents simultaneously when possible (Phase 2, Phase 3a).

---

Integration with Other Commands

This command complements:

• /centauro:detect - Analysis without loading
• /centauro:load - Simpler loading without quality gate
• /centauro:build - Create new contexts
• /centauro:reflect - Update contexts after task completion

Typical workflow:

1. User runs /centauro:curate "task" → Quality-filtered curation
2. User works on task with curated contexts
3. User runs /centauro:reflect → Extract learnings, update contexts
4. Future curations benefit from updated quality metadata

---

Success Criteria

A successful curation:

• ✅ Phase 1 scans frontmatter only (< 2K tokens for 20 files)
• ✅ Quality gate filters contexts correctly
• ✅ Missing quality metadata is flagged
• ✅ Phase 2 extracts in parallel by component
• ✅ Phase 3 dynamically chooses synthesis strategy
• ✅ Curated memory written to c4-memory/curated/
• ✅ User receives clear summary and next steps
• ✅ Process completes in reasonable time (< 2 minutes for 30 contexts)

---

Notes

• This is an advanced version of /centauro:load with quality filtering
• Keep /centauro:load for simpler, faster loading without quality gate
• The curated memory becomes a new context that can be used in future curations
• Quality metadata must exist for this command to work effectively
• Use /centauro:assess-quality (future command) to batch-assess contexts