murillodutt/cellm/skills/knowledge-funnel

Knowledge Routing — Before Moving On

Classify discoveries and route only reusable/systemic signals to the SCE learning loop.

Intent

• Prevent loss of high-value discoveries during execution.
• Keep ephemeral noise out of memory stores.
• Actively persist reusable insights via knowledge_add — not just expertise and observations.

Policy

• Classify into EPHEMERAL, REUSABLE, ARCHITECTURAL.
• Use knowledge_add for reusable atomic facts that benefit future sessions.
• Use context_record_outcome for reusable/architectural outcomes.
• Promote evidence only when confidence and repetition justify it (context_promote).

When to call knowledge_add

Call knowledge_add when you encounter any of these during work:

| Trigger | Example | |---------|---------| | Non-obvious behavior discovered | "libsql requires explicit readYourWrites: true for WAL mode consistency" | | Gotcha resolved after debugging | "bun:sqlite readonly: false is not valid — omit the option instead" | | Pattern confirmed across sessions | "FTS corruption on UPDATE requires try/catch with autoRepairFts" | | Cross-session insight | "spec_create_node rejects dependsOnPhase: null — omit the field" | | API contract nuance | "knowledge_add MCP deduplicates by title+scope+project — returns existing ID" | | Architecture decision recorded | "hookGenerateKnowledgeAtoms is a no-op — atoms come from expertise + dot23 + MCP" |

Example calls

knowledge_add({
  title: "bun:sqlite Database constructor rejects readonly:false — omit option for read-write",
  scope: "oracle",
  project: "<project-name>",
  source: "session",
  detail: "Bun v1.3.10 throws SQLITE_MISUSE when readonly:false is passed. Use new Database(path) without options for read-write.",
  confidence: 0.9
})

knowledge_add({
  title: "spec_create_node approval gate rejects phases under pending checks",
  scope: "oracle",
  project: "<project-name>",
  source: "session",
  detail: "CHECK_NOT_APPROVED error when adding phase/task under a check in pending/cancelled/deferred/failed state. Transition to active first.",
  confidence: 0.95
})

When NOT to call knowledge_add

• Routine edits, file reads, or standard operations
• Information already captured in this session's expertise
• Temporary state or progress updates
• Facts specific to one session with no reuse value

Routing

1. EPHEMERAL: continue work, no persistent write.
2. REUSABLE: knowledge_add (atomic fact) + context_record_outcome + record_expertise.
3. ARCHITECTURAL: knowledge_add + record_observation + record_expertise + context_record_outcome.
4. Periodically run context_promote for evidence-based progression.

NEVER

• Let a discovery die unclassified — if you learned something reusable, capture it
• Capture ephemeral context — session progress, temp state, and routine edits are noise
• Duplicate existing knowledge — knowledge_add deduplicates automatically by title+scope+project
• Skip SCE write-back — reusable/architectural signals must go through context_record_outcome
• Block work to capture — classify inline, call MCP, continue working
• Skip knowledge_add for REUSABLE discoveries — expertise alone is not searchable cross-session; atoms are