agent-retrieval-and-contextlisted

# CIEL ADAPTATION: Agent Retrieval (Progressive Context) This skill solves the "Context Blindness" problem in multi-agent workflows. it formalizes the loop of searching, evaluating, and refining context before execution. ## The Iterative Loop (Max 3 Cycles) 1. **Dispatch (Broad)**: Start with high-level keywords and domain patterns (e.g., `src/auth/**`). 2. **Evaluate (Score)**: Assess candidate files for relevance (0.0 - 1.0). Identify specific gaps (e.g., "Missing JWT util types"). 3. **Refine (Narrow)**: Update search query with discovered terminology and exclude irrelevant paths. 4. **Finalize**: Proceed ONLY when 3+ high-relevance (>= 0.7) files are found or gaps are identified as "Unextractable." ## Context Optimization - **Token Budget**: Target 3000-5000 tokens for the initial context set. - **Exclusion Mandate**: Prohibit sending test files, lockfiles, or build artifacts unless explicitly required for a fix. - **Grep Over Read**: Use `grep_search` to identify relevant blocks before performing a full `read_file`. ## Subagent Handover - **Context Injection**: Pass only the "High Relevance" files and the "Problem Statement" to subagents. - **Instructions**: Subagents MUST be told: "Your context is limited; perform one refinement cycle if you lack library definitions." ## Anti-Patterns - **The Context Dump**: Sending 20+ files to a subagent without relevance scoring. - **Amnesiac Search**: Searching for the same keywords in every cycle instead of evolving them b