longhang2004/antigravity-reliability-skills/session-router

Purpose

The session-router skill acts as the entry-point traffic controller for Gemini sessions in Antigravity. It inspects context, categorizes the incoming task, calculates the risk level, and dynamically selects required reliability skills without polluting context with unnecessary prompt rules.

When to use

• At the start of any new session or workflow.
• When a user prompt introduces a significant shift in task scope or requirements.

When not to use

• During mid-stream atomic tool executions within an already classified task phase.

Inputs

• User prompt text.
• Workspace file tree state (list_dir, git status).
• Active open editor files and cursor position metadata.

Workflow

1. Context Inspection: Survey directory structure and Git status cleanly.
2. Task Categorization: Classify task into coding, debugging, research, review, long-running, or mixed.
3. Risk Tiering:

• low: Read-only queries, research, documentation lookup.
• medium: Single-file bug fix, localized unit test creation.
• high: Multi-file refactor, dependency upgrade, schema change.
• critical: Mass file deletion, security patch, production deployment.

1. Skill Selection: Select minimal set of active P0 skills (grounding-first, plan-before-change, bounded-edit-executor, verify-before-answer).
2. Decision Record Emission: Emit a concise Decision Record summary.

Decision Points

• Is the request read-only research? -> Activate grounding-first. Deactivate write executors.
• Does the request modify code across multiple files? -> Activate plan-before-change, bounded-edit-executor, and verify-before-answer.
• Is the task high risk? -> Enforce explicit plan approval before executing any file edits.

Safety Constraints

• MUST NOT ask the user to clarify information already present in the workspace context.
• MUST NOT default to loading all skills for simple read-only queries.

Expected Outputs

{
  "task_type": "coding",
  "risk_level": "medium",
  "required_skills": ["plan-before-change", "bounded-edit-executor", "verify-before-answer"],
  "allowed_paths": ["src/", "tests/"],
  "verification_requirements": ["run_unit_tests"]
}

Failure Handling

• If workspace context is ambiguous, inspect .git or README.md to infer project type.
• Default to grounding-first and plan-before-change if risk level cannot be computed.

Verification

• Validate Decision Record schema against catalog/skills-index.json.

Examples

Positive Example 1: Read-Only Research

User Prompt: "How does authentication work in this repo?" Agent Action:

{
  "task_type": "research",
  "risk_level": "low",
  "required_skills": ["grounding-first"],
  "allowed_paths": [],
  "verification_requirements": ["cite_source_lines"]
}

Positive Example 2: Multi-File Feature Request

User Prompt: "Add JWT validation to API endpoints in src/api/" Agent Action:

{
  "task_type": "coding",
  "risk_level": "high",
  "required_skills": ["plan-before-change", "bounded-edit-executor", "verify-before-answer"],
  "allowed_paths": ["src/api/", "tests/api/"],
  "verification_requirements": ["pytest tests/api/"]
}

Negative / Adversarial Example 1: Loading All Skills Unnecessarily

User Prompt: "What line is main() on in app.py?" Bad Agent Behavior: Loading full multi-file planning and write-executor pipelines, wasting context tokens. Correct Behavior: Classify as research (risk low), activate only grounding-first.

Negative / Adversarial Example 2: Asking Repeated Questions

User Prompt: "Fix the typo in README.md" Bad Agent Behavior: Asking user "Which repository is this? What framework are you using?" Correct Behavior: Inspect workspace directly using tools, emit Decision Record, execute fix.

Self-Checklist

• [ ] Task type correctly identified?
• [ ] Risk level assigned accurately?
• [ ] Decision record emitted before action?