Large system prompt Injection on every turn

[shribr]

I asked Github Copilot to review and optimize all of the my custom agents and it called out Squad in particular saying it had "Massive system prompt injected every turn" -its words.

It noted that the body size was 1,158 lines / 73 KB.

Two structural facts:

1. Every message resends the full system prompt + accumulated history — that's inherent to the stateless API. Big agent bodies get re-billed on every turn.
2. squad.agent.md alone (~18K tokens) costs roughly $0.27 in input per turn on Opus — before any conversation history.

[shribr]

Optimize Squad agent to reduce token overhead

Problem

.github/agents/squad.agent.md is ~73 KB (≈18K tokens, 1,158 lines) and is re-billed as system-prompt input on every turn of a Squad session — before any conversation history. This is the single biggest token-cost driver among our custom agents.

Proposed changes (priority order)

1. Externalize mode-specific sections to on-demand files (biggest win)

Squad already uses the right pattern in two spots (.squad/templates/casting-reference.md, "read on-demand") — it just isn't applied broadly. The sections below are only needed in specific modes, not every turn, yet they're all inline:

Section	~Lines	Only needed when
Ralph — Work Monitor	150	squad watch / issue monitoring
GitHub Issues Mode	25	issues workflow active
PRD Mode	15	PRD workflow active
MCP Integration	35	MCP actually in use
Client Compatibility	45	non-CLI client
Per-Agent Model Selection tables	90	spawning an agent

Action: Move each to .squad/templates/*.md and leave a 2-line "read on-demand" pointer, matching the existing casting-reference pattern.

Impact: ~1,158 → ~550 lines (~50% cut), saving ~9K input tokens per Squad turn.

2. Fix stale model names (correctness + cost)

The model-selection tables and fallback chains reference models that no longer exist:

• claude-sonnet-4.5 → claude-sonnet-4.6
• claude-opus-4.5 → claude-opus-4.8
• gpt-5.2-codex → gpt-5.3-codex
• gemini-3-pro-preview → gemini-3.1-pro-preview

Action: Update all model references. Stale names risk silent fallback to a wrong/pricier model.

3. Pin Squad's own model

Squad has no model: in frontmatter → risks defaulting to Opus. It's an orchestrator (mostly delegates + writes spawn prompts).

Action: Add model: claude-sonnet-4.6 to frontmatter — quality for prompt-writing without Opus pricing for coordination.

4. Preserve what already works

• Keep the "cost-first, Haiku by default unless writing code" routing logic — only update the model names within it.
• Keep the drop-box / sub-agent fan-out architecture — workers get fresh context, so their history never bloats the orchestrator thread.

Expected outcome

• ~50% smaller always-loaded system prompt
• No stale-model fallback risk
• Pinned orchestrator model

For a heavy Squad user, this is roughly the difference between ~$0.27 and ~$0.13 of system-prompt overhead per turn.

Acceptance criteria

• Mode-specific sections moved to .squad/templates/*.md with on-demand pointers
• squad.agent.md reduced to ~550 lines
• All model names updated to current lineup
• model: claude-sonnet-4.6 pinned in frontmatter
• Routing logic and drop-box architecture preserved