cdeust/ai-architect-mcp/orchestrator

cortex:recall(query="pipeline lessons failures decisions for repo {target_repo}", limit=15)
cortex:recall(query="pipeline retry patterns verification failures for {task_description}", limit=10)

cortex:remember(
  content="Pipeline run on repo {target_repo}: task={task_description}, findings_processed={count}, stages_completed={list}, total_retries={sum}, final_compound_scores={scores}, key_decisions={decisions}, failure_patterns={failures}",
  tags=["pipeline-run", "repo:{target_repo_name}", "task:{task_type}", "stage:orchestrator"]
)

ai_architect_fs_read(path="{data_dir}/artifacts/blueprint-signoff.json")
→ BLOCK if file missing or signoff.approved == false
 
ai_architect_fs_read(path="{data_dir}/artifacts/skill-version-manifest.json")
→ Compare each skill version against SKILL.md frontmatter version fields
→ BLOCK if any version mismatch (stale skill = stale logic)

ai_architect_load_session_state(session_id="{sessionID}")
→ Extract: currentStage, activeFindingID, retryCount, loadedSkillVersions
→ If resuming: skip to the stage indicated by currentStage

ai_architect_query_context(finding_id="{findingID}", query="findings queue sorted by compound impact score descending")
→ Returns ranked list of findings awaiting processing
→ Each finding gets its own StageContext namespace keyed by findingID

ai_architect_git_worktree_add(branch_name="pipeline/{findingID}", base="main")
→ Returns worktree path — one worktree per finding, context isolation enforced
→ No two agents share a worktree or StageContext namespace

Stage 0: skills/stage-0-health/SKILL.md      → stage-0-health-report.json
Stage 1: skills/stage-1-discovery/SKILL.md    → stage-1-findings.json
Stage 2: skills/stage-2-impact/SKILL.md       → stage-2-impact-map.json
Stage 3: skills/stage-3-integration/SKILL.md  → stage-3-integration-design.md
Stage 4: skills/stage-4-prd/SKILL.md          → 9-file PRD package
  ↕ (Stage 4.5: skills/stage-4-5-interview/SKILL.md → plan-interview-report.json)
Stage 5: skills/stage-5-review/SKILL.md       → stage-5-verification-report.json
Stage 6: skills/stage-6-implementation/SKILL.md → stage-6-implementation-manifest.json
Stage 7: skills/stage-7-verification/SKILL.md → stage-7-hor-report.json
Stage 8: skills/stage-8-benchmark/SKILL.md    → stage-8-benchmark-report.json
Stage 9: skills/stage-9-deployment/SKILL.md   → stage-9-test-report.json
Stage 10: skills/stage-10-pr/SKILL.md         → stage-10-pr-manifest.json

# Load ALL upstream artifacts
stage_0 = ai_architect_load_context(stage_id=0, finding_id="{findingID}")
stage_1 = ai_architect_load_context(stage_id=1, finding_id="{findingID}")
stage_2 = ai_architect_load_context(stage_id=2, finding_id="{findingID}")
stage_3 = ai_architect_load_context(stage_id=3, finding_id="{findingID}")
 
# Load codebase intelligence (if available)
codebase_query = ai_architect_codebase_query(query="{finding_keywords}", repo="{target_repo}")
codebase_symbols = ai_architect_codebase_context(name="{primary_symbol}", include_content=true, repo="{target_repo}")
 
# Package as upstream_context — this IS the input to Stage 4
upstream_context = {
  "stage_0_health": stage_0,
  "stage_1_findings": stage_1,
  "stage_2_impact": stage_2,
  "stage_3_integration": stage_3,
  "codebase_intelligence": {
    "query_results": codebase_query,
    "symbol_details": codebase_symbols
  }
}
 
# NOW execute Stage 4 with full context
→ Read skills/stage-4-prd/SKILL.md
→ Execute with upstream_context available

ai_architect_save_session_state(state_data={
  "session_id": "{sessionID}",
  "finding_id": "{findingID}",
  "current_stage": N+1,
  "status": "running",
  "completed_stages": [0, 1, ..., N]
})
 
ai_architect_append_audit_event(event_data={
  "event_id": "stage-transition-{findingID}-{N}-to-{N+1}",
  "session_id": "{sessionID}",
  "stage_id": N,
  "tool_name": "orchestrator",
  "outcome": "pass",
  "message": "Stage transition for finding {findingID}: stage {N} → stage {N+1}",
  "metadata": {"from_stage": "{N}", "to_stage": "{N+1}"}
})

ai_architect_save_session_state(state_data={
  "session_id": "{sessionID}",
  "finding_id": "{findingID}",
  "current_stage": {retry_target_stage},
  "status": "running",
  "completed_stages": [0, 1, ..., {retry_target_stage - 1}],
  "metadata": {"retry_count": "{current + 1}", "retry_reason": "{failed_rule_or_gate}"}
})

ai_architect_check_context_budget(total_tokens={model_context_limit}, used_tokens={current_usage})
→ If usage ≥ 70%: switch to L2 summaries (load config + summaries only)
→ If usage ≥ 93%: auto-create handoff and compact
 
ai_architect_create_handoff(
  completed=[...],
  in_progress=[...],
  blocked=[...],
  next_actions=[...],
  session_id="{sessionID}"
)

ai_architect_git_worktree_remove(worktree_path="{worktree_path_from_add}")
→ Only after Stage 10 succeeds (PR created)

ai_architect_emit_ooda_checkpoint(stage_id=0, phase="observe", decision="Blueprint signoff present: {true/false}", confidence=1.0, session_id="{sessionID}")
ai_architect_emit_ooda_checkpoint(stage_id=0, phase="observe", decision="All skill versions match: {true/false}", confidence=1.0, session_id="{sessionID}")
ai_architect_emit_ooda_checkpoint(stage_id=0, phase="observe", decision="Each finding has one agent: {true/false}", confidence=1.0, session_id="{sessionID}")
ai_architect_emit_ooda_checkpoint(stage_id=0, phase="observe", decision="No shared stage context: {true/false}", confidence=1.0, session_id="{sessionID}")
ai_architect_emit_ooda_checkpoint(stage_id=0, phase="observe", decision="All active findings have worktree: {true/false}", confidence=1.0, session_id="{sessionID}")
ai_architect_emit_ooda_checkpoint(stage_id=0, phase="observe", decision="Context budget under 70 percent: {true/false}", confidence=1.0, session_id="{sessionID}")
ai_architect_emit_ooda_checkpoint(stage_id=0, phase="decide", decision="All retry counts under max: {true/false}", confidence=1.0, session_id="{sessionID}")