| name | specialists |
| description | Sub-loader for domain specialists (agent infra, terminal, harness, perf, security, speed). Each entry captures the WHY and decision tree only — not implementation details. Trigger via `/specialists` or whenever a decision in one of these domains is on the table. |
Specialists
Index of domain specialists. Each file answers why to pick an approach and when — never how. Implementation details (APIs, constants, syntax) stay in code and library docs (context7).
When to consult
- Before a non-trivial design decision in the listed domain.
- During review when a choice feels arbitrary — load the specialist to check the why.
- When unsure which approach fits — read the decision tree first, then commit.
Index
Agent / AI infra
| Specialist | When to load |
|---|
| mcp-design | Building MCP servers, tool granularity, MCP vs in-process |
| prompt-caching | Structuring prompts for reuse, cache breakpoints, invalidation reasons |
| agent-memory | What to persist across sessions, retrieval triggers, memory shape |
| tool-routing | Eager vs deferred tool loading, large catalogs, dispatch accuracy |
Terminal / TUI (no Tauri)
| Specialist | When to load |
|---|
| ratatui-arch | Structuring a ratatui app, state ownership, blocking work |
| ansi-pty | Escape sequences, capability detection, PTY semantics |
| terminal-input | Kitty protocol, mouse, bracketed paste, raw mode safety |
Agent harness
| Specialist | When to load |
|---|
| agent-sandbox | Running model-generated commands, fs/net scoping, blast radius |
| harness-hooks | Pre/Post tool hooks, SessionStart, enforcing invariants |
| parallel-subagents | Fan-out decisions, context isolation, when sequential wins |
| context-mgmt | Compaction, deferred tools, summarization triggers |
Performance
| Specialist | When to load |
|---|
| profiling | Picking a profiler, baselines, measurement discipline |
| cache-locality | Data layout in hot loops, SoA/AoS, false sharing |
| allocator | Arena/slab/system choice, lifetime-based reasoning |
| contention | Lock-free vs mutex vs sharding, RwLock pitfalls |
Security
Speed (latency-first vs perf throughput-first)
Format contract
Every specialist file follows:
- When (trigger)
- Why care (consequence of getting it wrong)
- Decision tree (if X → Y, because reason)
- Tradeoffs (A vs B, what each pays)
- Anti-patterns (why each is bad, not how to write the good version)
No code, no constants, no syntax. If a file accumulates implementation details, refactor — that content belongs in the relevant library doc (use context7) or the source itself.
Maintenance
- When the user gives feedback that contradicts a specialist, update the file (not memory).
- Adding a new specialist: drop a file here, add a table row.
