BOSS Agent MCP Tools + Skills Improvement Report (Actionable)

Goal

Translate the MCP/skills gap analysis into an execution plan, integrate concrete routing improvements, and provide an implementation runway for remaining tool wiring.

Current Gaps (validated)

1. KG-brain tool routing is underpowered - Category inference relied on narrow lexical hints (sql, deploy, incident) and did not natively prioritize KG/journey semantics.
2. Tool ranking did not reflect domain intent - Ranking was pure lexical overlap and could select a generic match before a more specific orchestration/journey tool.
3. No explicit execution backlog artifact - Existing docs describe target tools, but there was no concise execution matrix connecting “what to build” + “how to integrate” + “how to validate”.

Plan of Action

Phase 1 — Routing and Selection Hardening (implemented)

• Extend category inference to include:
• kg: journey/touchpoint/stage/rule/outcome intent
• cell, workflow, orchestrator: execution intent
• existing data, engineering, ops, marketing hints retained
• Introduce task-intent boosts in tool ranking for high-value MCP tools:
• kg_query_journey_state
• kg_query_orchestration_rules
• kg_record_touchpoint
• kg_record_outcome
• workflow_execute_pipeline
• learning_update_rule_confidence

Phase 2 — Registry + API exposure (next)

• Register P0/P1 tools in MCP registry (single source of truth):
• kg_query_journey_state, kg_query_orchestration_rules, cell_invoke, kg_record_outcome, kg_record_touchpoint, orchestrator_invoke, gcs_read_bucket
• Ensure each tool has:
• JSON schema, category/tags, timeout policy, retry policy, observability labels.
• Add /tools/list verification gate in CI to assert required tools are discoverable.

Phase 3 — /process end-to-end integration (next)

• Wire process flow to call BossAgentOrchestrator with:
• task, payload, optional preferred category
• Persist orchestration trail:
• selected category/tool, fallback attempts, success/failure, latency.
• Add learning-loop hook for outcomes where applicable.

Phase 4 — Skills governance and UX (next)

• Align Coding MOA skill detection with MCP categories and tool names.
• Add policy controls:
• approval requirements by risk tier
• denylist/allowlist for high-risk operations
• auditable reason metadata for infra-affecting actions.

Integration Blueprint

Integration points

1. Tool discovery layer: ToolCatalog.refresh_if_needed()
2. Category routing: ToolCatalog.pick_category()
3. Tool selection: BossAgentOrchestrator._rank_tools()
4. Execution and fallback: SubAgentExpert.execute() + candidate iteration in BossAgentOrchestrator.run()
5. Skill bridge: src/services/coding_moa/boss_agent_skills.py (future schema harmonization)

Validation strategy

• Unit tests for:
• category inference for journey/orchestration prompts
• ranking preference for domain-specific tools
• fallback behavior and payload sanitization
• Smoke tests against mock MCP client before enabling production flag.

Executed in this change set

• ✅ Expanded category inference for KG/cell/workflow/orchestrator intent.
• ✅ Added intent-based preferred-tool ranking boosts for MCP high-value tools.
• ✅ Added regression test to verify journey/orchestration tasks route to KG tools.

Remaining backlog

• [ ] Register P0/P1 tools in MCP registry and enforce discoverability checks
• [ ] Integrate orchestrator invocation into primary /process flow
• [ ] Add outcome telemetry and confidence update pathways
• [ ] Add policy/approval guardrails for infra-impacting skills

Definition of Done (target)

• Required P0/P1 tools discoverable at runtime via MCP list endpoint.
• Boss Agent route/tool decisions are deterministic for common journey prompts.
• End-to-end /process calls execute tool(s), capture outcome, and persist telemetry.
• Coding MOA skill calls are mapped to MCP tool contracts and policy-gated.