| name | request-immunity |
| description | Request immunity patterns for user-facing agents — tier selection (T1/T2/T3), Phase 0 injection templates, and RV gate definitions. Use when creating user-facing agents, injecting request validation, or checking immunity gate compliance. |
Request Immunity
Standard for preventing user-facing agents from acting on inconsistent, incomplete, ambiguous, or misleading requests. Every user-facing agent MUST include a Phase 0 that validates requests before execution. Validation intensity scales with the agent's exposure to freeform input.
When to Use This Skill
- Creating a user-facing agent that needs request validation
- Selecting the appropriate immunity tier (T1/T2/T3) for an agent
- Injecting Phase 0 validation patterns into an agent's methodology
- Checking RV-1 through RV-4 quality gates on existing agents
- Reviewing whether an agent's request immunity matches its exposure level
Methodology
Phase 1: Tier Selection
| Agent Characteristic | Tier | Approach | Required Assets |
|---|
| Handles freeform strategy/design/analysis requests with open scope | T1: Full | request-evaluation skill (full methodology) → mode-interactive for critical gaps | request-evaluation + mode-interactive skill refs |
| Executes specific tasks but user-invokable with potentially vague input | T2: Input | request-evaluation skill inline → mode-interactive if gaps | request-evaluation + mode-interactive skill refs |
| Narrow scope, well-defined target, low ambiguity surface | T3: Scope | Inline target identification check → ask if target unclear | None — self-contained |
Decision rules:
- Primary input is freeform natural language with open scope → T1
- Clear action verb but scope/target may be vague → T2
- Well-defined narrow scope that just needs a target → T3
- When in doubt → choose one tier higher
Phase 2: Inject Pattern
Select the injection pattern matching the chosen tier.
T1: Full Validation
Inject as Phase 0: Request Validation in the agent's methodology:
1. **Complexity check** — Single clear action with obvious scope?
- YES → bridge minor assumptions, note them, proceed
- NO → continue to step 2
2. **Apply `request-evaluation` skill** (full methodology) — Context Decomposition, Deliverable Analysis, Gap Detection, Challenge & Bridge
3. **Process results**:
- No critical gaps → proceed with bridged assumptions documented
- Critical gaps → apply `mode-interactive` skill to present gaps as questions
4. **Proceed** with validated, gap-free request
Agent constraints: add request-evaluation + mode-interactive skill references.
T2: Input Validation
Inject as Phase 0: Input Validation in the agent's methodology:
1. **Sufficiency check** — Apply `request-evaluation` skill (Context Decomposition only):
- Target identifiable? (file, module, feature)
- Action clear? (what to do)
- Scope inferable? (how much)
2. **Bridge** — If 1-2 gaps resolvable from project conventions → bridge, note assumptions
3. **Escalate** — If target OR action undetermined → `mode-interactive` with 1-2 focused questions
4. **Proceed** with validated input
Agent constraints: add request-evaluation + mode-interactive skill references.
T3: Scope Validation
Inject as Phase 0: Scope Validation in the agent's methodology:
1. **Target identification** — Can I determine the specific target?
- Path, module, error context, or plan reference available? → proceed
- Multiple candidates? → ask: "Which {target type} should I focus on?" (list candidates)
- No target? → ask: "What would you like me to {action verb}?"
2. **Proceed** with identified target
No additional assets required — pattern is self-contained.
Phase 3: Verify RV Gates
After injection, verify all 4 request immunity gates pass:
| Gate | Check | Fail Condition |
|---|
| RV-1 | Phase 0 exists | Methodology doesn't start with request/input/scope validation |
| RV-2 | Tier appropriate | Validation tier doesn't match agent's freeform input exposure |
| RV-3 | Assets wired | Required skills (request-evaluation, mode-interactive) missing from frontmatter/constraints |
| RV-4 | Escalation path | Agent can't surface unresolvable gaps to user |
Anti-Patterns
- ❌ Skipping Phase 0 — Agent jumps straight to execution without validating request → wasted work on wrong target
- ❌ Over-tiering — T1 validation on a narrow-scope agent → unnecessary overhead, slows simple tasks
- ❌ Under-tiering — T3 on an open-scope agent → misses ambiguity, acts on assumptions
- ❌ Missing asset wiring — Selecting T1/T2 but not adding skill references → Phase 0 can't execute
- ✅ Right-sized validation — Match tier to actual freeform input exposure, wire required assets, verify RV gates
