🔗 Integration Map — Greg · AAL · Ambient MCP

Status: ACTIVE · Wired: 2026-04-30 ET · Pipeline role: prompt-quality control plane attached to the Genesis Conductor mesh.

Architect vertex — Greg

• Spec changes to the prompt-version DAG, grader thresholds, retry budget, or model-candidate set route to Greg - Work for architect-vertex review before promotion.
• Greg returns a compressed review (Problem · Invariants · Interface · Risks · Next step) and names this worker as the implementation target. Greg does not author prompt content — only invariants and routing.
• Trigger: mention Greg on this page or on a pending-eval content page that requires architect adjudication. Greg's invariants override the lenient pass ratio.

Substrate — Ambient Agent Access Layer

• Per Ambient Agent Access Layer — Product Spec, Schema, and Skeleton, this worker is registered as an AAL task handler under request_type = "prompt_eval" and request_type = "prompt_optimize".
• Routing decision (additive to AAL §Router skeleton): prompt_eval → primaryAgent: "self-evolving-retraining", requiresApproval: false, policyTier: "standard" and prompt_optimize → primaryAgent: "self-evolving-retraining", requiresApproval: true, policyTier: "prod_sensitive" (promotion writes to the Prompt Ledger and is gated).
• Capsule emissions: every eval run appends a capsule_events row with record_type="prompt_eval", event_type ∈ {run_started, section_scored, prompt_promoted, clean_exit}, payload_json carrying {run_id, prompt_version, sections_evaluated, pass, fail, avg_score, action}. State projection at ws:{workspace_id} reflects open optimization retries.
• Approval flow: any prompt_promote action lands as an approvals row (approval_type="prompt_promotion") before the Prompt Ledger flips to ✅ ACTIVE. Greg is the default requested_from_user for architect-tier promotions; operator (Igor) for standard tier.

Surface — Ambient MCP Server

• Per ‣, three new MCP tools expose this worker to Cursor / Gemini CLI / Codex CLI clients:
  • retraining_run_eval — submits an prompt_eval task with {section_url, prompt_version_override?}; returns {run_id, status, scores, action}.
  • retraining_get_status — polls {run_id} → returns latest Eval Run Log entry shape.
  • retraining_promote_prompt — submits a prompt_optimize task; gated by AAL approval flow above.
• All three tools authenticate via the existing AAL Authorization: Bearer <api_key> header (KV API_KEYS); no new auth surface introduced.
• IDE wiring uses the same HTTP-transport JSON-RPC entry: https://ambient-mcp-server.iholt.workers.dev/mcp.

Worker fleet linkage

• Per ‣, this worker registers as W-14 gc-retraining-control under W-00 gc-conductor's orchestration tier.
• Reads from: queued/pending-eval content pages, Prompt Ledger on this page, AAL tasks + jobs tables.
• Writes to: Eval Run Log (this page), AAL capsule_events, AAL approvals (on promotion), Slack #gc-retraining (status summaries).
• Trigger sources: trigger-97 (hourly), trigger-99 (daily 9 AM), trigger-101 (hourly), AAL queue dispatch on prompt_eval tasks, agent-mention.