intelli-verse-x Sid Orchestra
One Master. Specialists. Blackboard + leases. Human gates. Evals as quality gates. Loop until verified.
Say this
| Phrase | Action |
|---|---|
sid orchestra / @sid-orchestra |
Full pipeline |
sid swarm |
Fan out (only if parallelizable) |
sid orchestra loop |
Arm recurring loop |
sid evals |
Load eval rubrics / grade a batch |
solo / no swarm |
Single agent — required for sequential-only |
Read first (mandatory)
- This skill
.cursor/loops/sid-orchestra-loop.md.cursor/sid-orchestra/BLACKBOARD.md.cursor/sid-orchestra/LANE_LOG.md.cursor/sid-orchestra/EVALS.md- Role briefs
.cursor/agents/sid-*.mdas needed - Mem0 + Hindsight at intake
0) Triage: swarm vs solo (anti-slop)
Research (2026): multi-agent can hurt sequential reasoning by ~40–70%.
Do not swarm when the job is one linear thread.
| Signal | Mode |
|---|---|
| One file / one bug / one rename / one clear sequential fix | solo — Master only |
| Parallelizable paths, research fan-out, multi-feature | swarm |
| Unsure | Start solo research, then swarm only the parallel build |
Master must write on BLACKBOARD: Mode: solo | swarm and why.
Cast
| Role | Count | Job |
|---|---|---|
| Master | 1 | Intake, triage, gates, merge, user channel |
| Research | 1–3 | Code + Firecrawl/web; evidence-only |
| Planning | 1–2 | Lane cards + deps |
| Plan Critic | 1 | Eval the plan before build (Review-lite) |
| Working | N (2–6) | Code in leased paths |
| Generation | 0–2 | Scaffolds / assets |
| Review | 1–3 | Eval code; ≠ Working |
| Harness | 1–2 | Canary → then full tests |
| Bus | 1 | Blackboard, leases, re-orient |
| Memory / Integration | 0–1 | Retain / merge conflicts |
| Eval Judge | 0–1 | PASS/FAIL rubrics (can be Review) |
Hard rule: Working ≠ Review ≠ Plan Critic (for the same artifact).
Pipeline
USER → MASTER triage (solo vs swarm)
→ RESEARCH (evidence paths required)
→ GATE A (user approves understanding)
→ PLANNING
→ PLAN CRITIC + plan eval (PASS/FAIL) ← NEW
→ GATE B if high-risk or plan FAIL
→ BUS: run_id, locks+leases, LANE_LOG
→ WORKING(N) + GENERATION (parallel only)
→ CODE REVIEW + code evals (PASS/FAIL)
→ FIX ≤2
→ CANARY HARNESS ← NEW
→ FULL HARNESS (only if canary PASS)
→ REPORT + RETAIN + eval alignment note
Run ID + lane log (mandatory)
On intake Master creates:
run_id = sid-<YYYYMMDD>-<HHMM>-<short>
Example: sid-20260717-1310-streak
Every Task prompt and return must stamp:
RUN_ID: sid-...
LANE_ID: research-01 | plan-01 | plan-critic-01 | work-ui | review-01 | harness-canary | ...
ROLE: ...
STARTED_AT: ISO-8601
ENDED_AT: ISO-8601
STATUS: ok | fail | blocked | missing
EVIDENCE_PATHS:
- path/that/exists.dart
CLAIMS_WITHOUT_EVIDENCE: none | <list — Master discards these>
Append each return to .cursor/sid-orchestra/LANE_LOG.md.
Master discards any claim with no evidence path (anti-hallucination).
Lock leases (30 min TTL)
BLACKBOARD file locks use leases:
| Field | Rule |
|---|---|
path |
Exact file or directory globs |
owner_lane |
e.g. work-ui |
leased_at |
ISO time |
ttl_min |
30 default |
expires_at |
leased_at + ttl |
Bus steals stale locks: if now > expires_at and lane not heartbeat-updated, Bus reassigns or frees.
Workers must heartbeat (touch leased_at) every wave or when still writing.
One writer per path. Conflict → pause both → Integration/Master.
Plan critic (before build)
After Planning, launch Plan Critic (not the planner):
- Rubric:
.cursor/sid-orchestra/EVALS.md→ Plan Eval - PASS → Bus arms locks + build wave
- FAIL → Planning revises (≤2) or Gate B to user
Checks: path non-overlap, deps sound, sequential edges not parallelized, done-when testable, scope ≤ Gate A.
Canary harness → full harness
- Canary (required first)
dart analyzeon touched packages /flutter analyzescoped if possible- One focused test file or
flutter test test/<relevant>_test.dart - Or compile/smoke the changed feature entrypoint
- Full
flutter test/ broader suite only if canary PASS - Canary FAIL → Fix wave; do not burn full suite
Evals (stop slop)
Full rubrics: .cursor/sid-orchestra/EVALS.md and skill @sid-evals.
Rules:
- Prefer PASS/FAIL (not 1–10) for gates
- Codify what PASS means before generating
- LLM Judge grades → optional Human Judge → alignment %
- Mismatches → improve rubric → re-batch
- Flip to gate: discard / rework any output that FAILs
Gates in this orchestra:
| Gate | Eval | On FAIL |
|---|---|---|
| After Research | Understanding Eval | More research / ask user |
| After Plan | Plan Eval (Plan Critic) | Revise plan |
| After Build | Code Eval + Review scorecard | Fix ≤2 |
| After Gen assets | Asset/Copy Eval | Regenerate |
| After Harness | Harness Eval | Fix or BLOCKED |
Anti-hallucination (reduce agent fiction)
- Evidence or MISSING — no path → claim dropped
- No invented APIs/files — say
MISSING:explicitly - Quotes from repo for architectural claims when possible
- Eval gates throw away failing generations
- Solo for sequential — fewer agents inventing conflicts
- Blackboard contracts — Workers implement contracts, not vibes
- Discard unaudited peer claims in Master merge
Launch protocol (Task tool)
RUN_ID: sid-...
LANE_ID: ...
ROLE: ...
GOAL: <one line>
BLACKBOARD: .cursor/sid-orchestra/BLACKBOARD.md
LANE_LOG: append stamped return to .cursor/sid-orchestra/LANE_LOG.md
EVAL: <rubric name from EVALS.md or none>
IN SCOPE: <exact paths>
OUT OF SCOPE: <forbidden>
DEPS: <lanes or none>
LEASES: <paths you will lock>
DONE WHEN:
- [ ] ...
RETURN: RUN_ID, LANE_ID, STATUS, evidence paths, blockers, eval PASS/FAIL if judged
ANTI-GUESS: Do not invent files. If missing, say MISSING.
Swarm size (swarm mode only)
| Job | Agents excl. Master |
|---|---|
| Small parallel feature | 5–6 |
| Medium | 7–9 |
| Large port | ≤12 |
Default QuizVerse: Research(2)+Plan(1)+PlanCritic(1)+Work(3)+Review(2)+Harness(1)+Bus(1).
Approval gates
Gate A — Understanding (non-trivial): goal, assumptions, out-of-scope, ≤5 questions → Approve to plan?
Gate B — Plan: high-risk or Plan Eval FAIL after 2 revises → Approve to build?
Review scorecard + eval
BLOCKER / MAJOR / MINOR / NIT + overall Code Eval PASS/FAIL.
No BLOCKERs and Code Eval PASS → Canary.
Done criteria
- Mode solo|swarm recorded
-
run_idset; LANE_LOG has every lane - Gate A approved (or skip)
- Plan Eval PASS
- Leases respected / stale stolen
- Code Eval PASS; no BLOCKERs
- Canary PASS then full harness (or documented skip)
- Master report + memory retain
- Hallucinated claims discarded
Related
- Loop · Blackboard · Lane log · Evals
@sid-evals·@qv-skill-router- Agents:
.cursor/agents/sid-*.md