mirror of
https://github.com/Donchitos/Claude-Code-Game-Studios.git
synced 2026-06-27 13:01:50 +00:00
f36494e70c
Model tier assignment: - model: haiku → help, sprint-status, story-readiness, scope-check, project-stage-detect, changelog, patch-notes, onboard (read-only/format) - model: opus → review-all-gdds, architecture-review, gate-check (multi-doc synthesis, high-stakes verdicts) PostCompact hook: - New .claude/hooks/post-compact.sh — fires after compaction, reminds Claude to re-read production/session-state/active.md to restore context - Registered in settings.json between PreCompact and Stop Parallel Task spawning: - review-all-gdds: Phase 2 (consistency) and Phase 3 (design theory) now explicitly instructed to spawn as parallel Task agents simultaneously Error Recovery Protocol: - Standard BLOCKED-handling section added to: review-all-gdds, architecture-review, dev-story, team-combat, team-qa, team-narrative, team-level, team-ui, team-audio, team-release, team-polish - Pattern: surface blocker → assess dependencies → offer 3 options via AskUserQuestion → always produce partial report Coordination rules: - Added Model Tier Assignment table with routing rationale - Added Subagents vs Agent Teams section (experimental agent teams docs) - Added Parallel Task Protocol (when/how to spawn parallel agents) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
74 lines
3.7 KiB
Markdown
74 lines
3.7 KiB
Markdown
# Agent Coordination Rules
|
|
|
|
1. **Vertical Delegation**: Leadership agents delegate to department leads, who
|
|
delegate to specialists. Never skip a tier for complex decisions.
|
|
2. **Horizontal Consultation**: Agents at the same tier may consult each other
|
|
but must not make binding decisions outside their domain.
|
|
3. **Conflict Resolution**: When two agents disagree, escalate to the shared
|
|
parent. If no shared parent, escalate to `creative-director` for design
|
|
conflicts or `technical-director` for technical conflicts.
|
|
4. **Change Propagation**: When a design change affects multiple domains, the
|
|
`producer` agent coordinates the propagation.
|
|
5. **No Unilateral Cross-Domain Changes**: An agent must never modify files
|
|
outside its designated directories without explicit delegation.
|
|
|
|
## Model Tier Assignment
|
|
|
|
Skills and agents are assigned to model tiers based on task complexity:
|
|
|
|
| Tier | Model | When to use |
|
|
|------|-------|-------------|
|
|
| **Haiku** | `claude-haiku-4-5-20251001` | Read-only status checks, formatting, simple lookups — no creative judgment needed |
|
|
| **Sonnet** | `claude-sonnet-4-6` | Implementation, design authoring, analysis of individual systems — default for most work |
|
|
| **Opus** | `claude-opus-4-6` | Multi-document synthesis, high-stakes phase gate verdicts, cross-system holistic review |
|
|
|
|
Skills with `model: haiku`: `/help`, `/sprint-status`, `/story-readiness`, `/scope-check`,
|
|
`/project-stage-detect`, `/changelog`, `/patch-notes`, `/onboard`
|
|
|
|
Skills with `model: opus`: `/review-all-gdds`, `/architecture-review`, `/gate-check`
|
|
|
|
All other skills default to Sonnet. When creating new skills, assign Haiku if the
|
|
skill only reads and formats; assign Opus if it must synthesize 5+ documents with
|
|
high-stakes output; otherwise leave unset (Sonnet).
|
|
|
|
## Subagents vs Agent Teams
|
|
|
|
This project uses two distinct multi-agent patterns:
|
|
|
|
### Subagents (current, always active)
|
|
Spawned via `Task` within a single Claude Code session. Used by all `team-*` skills
|
|
and orchestration skills. Subagents share the session's permission context, run
|
|
sequentially or in parallel within the session, and return results to the parent.
|
|
|
|
**When to spawn in parallel**: If two subagents' inputs are independent (neither
|
|
needs the other's output to begin), spawn both Task calls simultaneously rather
|
|
than waiting. Example: `/review-all-gdds` Phase 1 (consistency) and Phase 2
|
|
(design theory) are independent — spawn both at the same time.
|
|
|
|
### Agent Teams (experimental — opt-in)
|
|
Multiple independent Claude Code *sessions* running simultaneously, coordinated
|
|
via a shared task list. Each session has its own context window and token budget.
|
|
Requires `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1` environment variable.
|
|
|
|
**Use agent teams when**:
|
|
- Work spans multiple subsystems that will not touch the same files
|
|
- Each workstream would take >30 minutes and benefits from true parallelism
|
|
- A senior agent (technical-director, producer) needs to coordinate 3+ specialist
|
|
sessions working on different epics simultaneously
|
|
|
|
**Do not use agent teams when**:
|
|
- One session's output is required as input for another (use sequential subagents)
|
|
- The task fits in a single session's context (use subagents instead)
|
|
- Cost is a concern — each team member burns tokens independently
|
|
|
|
**Current status**: Not yet used in this project. Document usage here when first adopted.
|
|
|
|
## Parallel Task Protocol
|
|
|
|
When an orchestration skill spawns multiple independent agents:
|
|
|
|
1. Issue all independent Task calls before waiting for any result
|
|
2. Collect all results before proceeding to dependent phases
|
|
3. If any agent is BLOCKED, surface it immediately — do not silently skip
|
|
4. Always produce a partial report if some agents complete and others block
|