monkilabs/team-lead-reference

Team Lead Reference

Delegation Sequence

1. Score task complexity (table below) → determines tier
2. Route to model tier via Cost-Aware Routing
3. Deepen plan if 3+ subtasks (Deepen-Plan Protocol)
4. Check pre-delegation policy (5-point checklist below)
5. Delegate using Compact Delegation Envelope
6. Handle output per Status Handling table
7. Log via observability-logging skill

For specialist agent registry, model assignments, see agent-registry.md.

Cost-Aware Model Routing

| Tier | Use For | |------|---------| | Premium | Team Lead orchestration, highest-stakes decisions | | Quality | Feature implementation, UI/frontend, security, architecture, complex reasoning | | Standard | Large-scale analysis, schema design, cost-efficient coding, repo exploration | | Fast | Terminal-heavy tasks, E2E tests, data pipelines, agentic workflows | | Economy | Docs, simple config, formatting, boilerplate |

Selection: Default to agent's assigned tier. Downgrade pure docs/config → Economy. Upgrade security/architecture ambiguity → Quality/Premium. Never Premium/Quality for boilerplate. 3+ parallel agents → prefer Economy/Fast/Standard.

Complexity-Based Task Scoring

| Factor | Low → High | |--------|------------| | Files touched | 1–2 → 3–5 → 6+ / cross-library | | Reasoning depth | Boilerplate → pattern matching → architecture/security/tradeoffs | | Ambiguity | Clear spec → some judgment → multiple valid approaches | | Risk | Reversible → moderate impact → DB/auth/breaking | | Dependencies | None → 1–2 upstream → complex chain |

| Score | Tier | Examples | |-------|------|----------| | 1–2 | Economy/Fast | Docs update, config tweak, rename, simple test | | 3–5 | Standard/Quality | Component, CMS query, API route, migration | | 8 | Quality | Architecture decision, security audit, complex refactor | | 13 | Quality + Panel | DB migration with data transform, auth flow redesign |

Overrides: Blocker (blocking 2+ downstream) → upgrade one tier. Security-touching → Quality+. Pure docs → Economy. Registry default takes precedence unless complexity clearly warrants change.

Deepen-Plan Protocol

| Plan Complexity | Action | |----------------|--------| | 1–2 subtasks, familiar | Skip — delegate directly | | 3–5 subtasks, mixed | Quick deepen — single Researcher sub-agent | | 6+ subtasks, unfamiliar | Full deepen — parallel Researcher sub-agents |

Quick deepen: Fire one Researcher for exact file paths & line numbers, patterns to follow (file:line examples), relevant lessons from LESSONS-LEARNED.md, risks/blockers per subtask. Full deepen: Split by domain into parallel Researchers. See agent-registry.md for scope examples.

| Field | Before Deepen | After Deepen | |-------|--------------|------------------| | Files | "some component" | Exact path + line range | | Pattern | "follow existing style" | Specific file:line reference | | Risks | unknown | Known issues identified | | Lessons | unchecked | Relevant lessons applied | | Dependencies | assumed | Verified with exact imports |

Agent Output Status Handling

| Status | Action | |--------|--------| | Complete | Fast review | | Complete with concerns | Resolve before review | | Needs context | Provide info; re-dispatch | | Blocked | Upgrade model/escalate; never re-dispatch unchanged |

Pre-Delegation Policy Checks

1. Tracker issue exists for this task
2. File partition clean (no overlap with parallel agents)
3. All dependency tasks Done
4. Delegation prompt has file paths + acceptance criteria
5. Self-improvement reminder included (Read LESSONS-LEARNED.md first)

Feature work adds: (6) Known issues reviewed, (7) Architecture docs read, (8) Existing code searched. High-risk work adds: (9) Panel review planned, (10) Rollback path identified.

Compact Delegation Envelope

{
  "tracker": "TAS-XX",
  "agent": "Agent Name",
  "objective": "One sentence: what to do and why.",
  "files": ["path/to/file.ts", "path/to/other.ts"],
  "acceptance_criteria": ["AC 1", "AC 2"],
  "constraints": "Only modify listed files. Read LESSONS-LEARNED.md first.",
  "output_contract": "Return: files changed, lint/type/test pass/fail, discovered issues."
}

tracker required; acceptance_criteria verbatim from tracker; files = exact resolved paths (not globs); output_contract = agent's Base Output Contract. Maintain running delegation log in session checkpoint (see session-checkpoints skill).

Context Source Tagging

Prefix each agent's output summary ### [Agent Name] TAS-XX Description. Never merge outputs from different agents. Cite source agent when referencing prior output. Include Agent column in checkpoint "Completed Work" tables.

Dead Letter Queue Format

Log to .opencastle/AGENT-FAILURES.md when agent fails 2+ attempts, background output fails all gates, or unrecoverable error occurs. Panel 3x BLOCK → create dispute instead.

Entry (DLQ-XXX: Short description): Date, Agent, Tracker Issue, Failure Type (verification-fail / tool-error / panel-block / timeout / scope-creep), Attempts, Task, Failure Details, Resolution. Scan DLQ for pending retries at session start.

Error Recovery

For common failure modes, recovery procedures, load orchestration-protocols skill.

Dispute Protocol

Triggers: Panel 3× BLOCK, agent-reviewer disagreement, criteria contradictions, no convergence, needs human input.

Create in .opencastle/DISPUTES.md:

1. Number (DSP-XXX), set priority, document both perspectives with file references
2. Build attempt history; present ≥2 options with rationale/risk
3. Link panel reports, DLQ entries, changed files
4. Log with observability-logging dispute command

After resolution: resolved → re-delegate with decision as constraint. deferred → follow-up issue.