Extract cold-path sections from squad.agent.md into satellite files #98

Problem

squad.agent.md is ~27K tokens / ~108K chars / ~1,900 lines. It's loaded as the system prompt on every turn of every conversation. Users are experiencing "response was interrupted due to a server error. retrying" loops — even on simple chat turns with no agents spawned. The system prompt alone consumes too much of the context window.

Proposed Solution — Satellite File Extraction

Extract cold-path sections from squad.agent.md into satellite files that the coordinator reads on demand via view tool calls when specific features are triggered.

Hot path (~7-10K tokens, stays in squad.agent.md)

• Coordinator identity + refusal rules
• Init/Team mode gate
• Routing logic + trigger tables
• Response mode selection + model selection
• Spawn templates (standard + lightweight)
• Anti-patterns list
• One-line trigger summaries for every cold-path feature with read {file} before executing pointers

Cold path → satellite files in .github/agents/squad/

Implementation notes

• Satellite files live in .github/agents/squad/ alongside squad.agent.md
• index.js copy logic needs to recursively copy .github/agents/squad/ to consumer repos
• Platform loads ONE .agent.md file — satellites must be explicitly read via view tool (each costs one tool call)
• Each extracted feature keeps a 1-2 line trigger description in the hot prompt so the LLM never forgets a capability exists
• VS Code: same behavior, no platform-specific issues
• Estimated reduction: 70%+ (from ~27K to ~7-10K tokens)

Risks

• False negatives: If trigger descriptions are too terse, coordinator may skip loading a satellite when it should. Needs empirical testing.
• Latency: First turn of a cold-path feature adds one tool call. Acceptable tradeoff vs. server errors.

Why bundle with v0.5.0

The .ai-team/ → .squad/ rename (#69) already touches every path reference in squad.agent.md. Doing extraction simultaneously avoids rewriting paths twice.

Team consensus (from discussion 2026-02-19)

• Keaton (Lead): Target 8-10K hot path. Move whole sections out intact — don't compress. Ship casting tables first.
• Verbal (Prompt Engineer): Tiered prompting with trigger-gated lazy loading. Keep one-line summaries of every capability in hot path.
• Fenster (Core Dev): Mechanically straightforward. Flag index.js copy logic as real work. Prototype with 2-3 satellites first.