update-models

<user_request> $ARGUMENTS </user_request>

  You MUST:
  - Use Task tool to delegate ALL data manipulation to agents
  - Use AskUserQuestion for user approval gate
  - Use Bash for running sync script and git commands
  - Use Read for file verification only
  - Use Glob/Grep for finding files

  You MUST NOT:
  - Use Write or Edit tools directly
  - Perform filtering, categorization, or file updates yourself
  - Manipulate data structures
  - Only coordinate workflow and make decisions about delegation

  **Rationale:** Orchestrators coordinate but never manipulate data. All data operations
  must be delegated to specialized agents with appropriate tool permissions.
</orchestrator_role>

<todowrite_requirement>
  You MUST use the TodoWrite tool to create and maintain a todo list throughout your orchestration workflow.

  **Before starting**, create a todo list with all 5 phases:
  1. PHASE 0: Initialization
  2. PHASE 1: Scrape and Filter Models
  3. PHASE 2: User Approval
  4. PHASE 3: Update Shared File
  5. PHASE 4: Sync to Plugins

  **Update continuously**:
  - Mark tasks as "in_progress" when starting
  - Mark tasks as "completed" immediately after finishing
  - Keep only ONE task as "in_progress" at a time
</todowrite_requirement>

<delegation_contract mandatory="true">
  **Strict Delegation Rules:**

  - ALL scraping → model-scraper agent
  - ALL filtering/deduplication/categorization → model-scraper agent
  - ALL file updates → model-scraper agent
  - Orchestrator only: coordinates, validates, approves, recovers from errors

  **Never violate this contract.** If you find yourself about to use Write/Edit,
  STOP and delegate to model-scraper instead.
</delegation_contract>

<user_approval_gate mandatory="true">
  You MUST present filtered model list to user before updating:

  - Show: model IDs, providers, categories, pricing
  - Allow user to: approve, modify, or reject
  - Do NOT proceed without approval
  - Support structured modification input (see knowledge section)
</user_approval_gate>

</critical_constraints>

<core_principles> - Orchestrator coordinates but never manipulates data - All data operations delegated to specialized agents - Clear separation of concerns: orchestration vs implementation - Tool restrictions enforced: NO Write/Edit in orchestrator

<principle name="Context-Aware Filtering" priority="high">
  Provide these filtering rules to model-scraper agent:

  1. **Anthropic Filter**: Exclude all Anthropic models (Claude available natively)
  2. **Provider Deduplication**: Max 1 per provider (keep top-ranked)
  3. **Category Balance**: Min 2 models per category (coding, reasoning, vision, budget)
  4. **Target Count**: 9-12 models
  5. **Diversity**: ≥5 different providers

  **Category Balance Override:**
  - If category has <2 models after deduplication, allow 2nd model from that provider
  - Priority: Category diversity > Provider deduplication
</principle>

<principle name="Safe Updates" priority="high">
  - Always read existing file first for verification
  - Create backup before sync
  - Rollback on any failure
  - Preserve existing file if update fails
  - User approval required before modifications
</principle>

<principle name="Error Recovery with Retry Limits" priority="high">
  - Handle scraping failures gracefully (max 3 attempts)
  - Rollback on sync script errors
  - Partial sync recovery (continue with successful, retry failures)
  - Log all errors for debugging
  - User decision gates for ambiguous failures
</principle>

</core_principles>