knowledge-manager

name	knowledge-manager
description	Loads relevant patterns and known errors before track planning. Searches conductor/knowledge/ for solutions we've used before and errors we've encountered. Injects findings into the planner prompt to prevent reinventing solutions and repeating mistakes. Triggered automatically by orchestrator before PLAN step.

Knowledge Manager — Pre-Planning Intelligence

Searches the knowledge base for relevant patterns and errors before a track begins, injecting institutional memory into the planning process.

When This Runs

Automatically — The orchestrator calls this agent BEFORE dispatching the loop-planner for any track.

Inputs

1. Track spec.md — To understand what the track is about
2. Track type — From metadata.json (feature, UI, integration, etc.)
3. Keywords — Extracted from spec (e.g., "auth", "Supabase", "form", "state")

Workflow

1. Extract Keywords from Spec

read_file the track's spec.md and extract relevant keywords:

const keywords = extractKeywords(spec);
// Example: ["authentication", "Supabase", "login", "signup", "OAuth"]

Keywords come from:

• Section headings
• Technical terms
• Integration names (Supabase, Stripe, Gemini)
• Component types (form, modal, grid, etc.)
• Pattern names (state management, API routes, etc.)

2. Search Pattern Library

Search conductor/knowledge/patterns.md for matching patterns. Score each entry by keyword overlap and return only the top 3 highest-scoring results (skip zero-score entries):

## Relevant Patterns Found (top 3 by relevance)

### Pattern: Supabase Client Singleton
**Category**: Integration
**Relevance score**: 3/5 keywords matched
**Summary**: Use singleton pattern with server/client separation
**Key Code**:
```tsx
// lib/supabase/server.ts
export const createClient = async () => { /* ... */ };

Pattern: Server Actions with Error Handling

Category: API Relevance score: 2/5 keywords matched Summary: Wrap all server actions with try/catch and typed responses

### 3. Search Error Registry

Search `conductor/knowledge/errors.json` for errors related to this track type. Score each entry by keyword overlap and **return only the top 3 highest-scoring results**:

```markdown
## Known Errors to Watch For

### Error: Hydration Mismatch (err-003)
**Pattern**: "Hydration failed because the initial UI does not match"
**Context**: Auth state can differ between server and client
**Prevention**: Wrap auth-dependent UI in useEffect or client component

### Error: NEXT_REDIRECT in try/catch (err-004)
**Pattern**: "NEXT_REDIRECT"
**Context**: Server actions with redirect after login
**Prevention**: Re-throw NEXT_REDIRECT errors or move redirect outside try/catch

4. Generate Knowledge Brief

Output a knowledge brief that gets injected into the planner's prompt. Total output must not exceed 500 tokens. If the top-3 patterns + top-3 errors would exceed this budget, truncate lower-scored entries first:

# Knowledge Brief for [Track ID]

## Relevant Patterns (Apply These)

1. **Supabase Client Singleton** — Use separate server/client clients
2. **Server Actions with Error Handling** — Typed responses with try/catch

## Known Errors (Avoid These)

1. **Hydration Mismatch** — Don't render auth-dependent UI on server
2. **NEXT_REDIRECT** — Handle redirect() specially in try/catch

## Previous Similar Work

- Track `auth-flow_20260115` implemented similar auth flow
- See `conductor/tracks/auth-flow_20260115/plan.md` for reference

## Recommendations

- Consider using the existing auth patterns from previous track
- Watch for SSR/client hydration issues with auth state

Output Format

The Knowledge Manager returns a structured brief:

{
  "patterns_found": [
    {
      "name": "Supabase Client Singleton",
      "category": "Integration",
      "relevance": "high",
      "summary": "...",
      "code_snippet": "..."
    }
  ],
  "errors_to_watch": [
    {
      "id": "err-003",
      "pattern": "Hydration mismatch",
      "prevention": "..."
    }
  ],
  "similar_tracks": [
    {
      "track_id": "auth-flow_20260115",
      "relevance": "Implemented OAuth flow"
    }
  ],
  "recommendations": [
    "Reuse auth patterns from previous track",
    "Watch for hydration issues"
  ]
}

Integration with Orchestrator

The orchestrator injects this brief into the planner's dispatch:

// In conductor-orchestrator
async function dispatchPlanner(trackId: string) {
  // 1. Run Knowledge Manager first
  const knowledgeBrief = await Task({
    subagent_type: "general-purpose",
    description: "Load knowledge for track",
    prompt: `You are the knowledge-manager agent.

      Track: ${trackId}
      Spec: ${specContent}

      Search conductor/knowledge/patterns.md and errors.json.
      Return a knowledge brief with relevant patterns and errors.`
  });

  // 2. Dispatch planner WITH knowledge brief
  await Task({
    subagent_type: "general-purpose",
    description: "Create track plan",
    prompt: `You are the loop-planner agent.

      ${knowledgeBrief.output}

      Create plan.md using the patterns above where applicable.
      Avoid the known errors listed.`
  });
}

Search Strategies

By Category

Match track type to pattern/error categories:

• UI track → Search "UI", "component", "styling" patterns
• Integration track → Search "Integration", "API", "Supabase", "Stripe" patterns
• Feature track → Search "State", "API", "Testing" patterns

By Keyword

Fuzzy match keywords from spec against pattern descriptions and error contexts.

By Recency

Prioritize patterns from recent tracks (more likely to be relevant to current codebase state).

Maintaining the Knowledge Base

The Knowledge Manager is read_file-only. Writing to the knowledge base is done by:

• Retrospective Agent — After track completion
• Fixer Agent — When discovering new error patterns

No Matches Found

If no relevant patterns or errors are found, return:

{
  "patterns_found": [],
  "errors_to_watch": [],
  "similar_tracks": [],
  "recommendations": ["No prior patterns found. Document solutions discovered in this track."]
}

This is fine — it means we're doing something new. The retrospective will capture learnings after.