feat(prompt): slim squad.agent.md (83KB → ~40KB) by extracting Routing / Team Mode / Source-of-Truth / Response Mode / Spawn / Init Mode to satellite skills #1308

Symptom

The coordinator file squad.agent.md is now 83KB / 1,137 lines (the canonical .squad-templates/squad.agent.md on dev tip 89a1b96a). For every Copilot session against a Squad coordinator, that entire blob is loaded as the agent prompt before the first tool call. This is wasteful and only growing — the v0.10 stabilisation set added another +10.6KB / +114 lines.

Measured (today, on dev)

Top sections by byte count:

These ten sections alone account for 37.6KB / 45% of the file. Most of them describe machinery the coordinator only needs in specific situations (during init, during a spawn, when picking a response mode, etc.) — not on every prompt.

Proposal

Apply the same canonical-doc + load-on-demand-skill pattern we used for cross-squad (discovery in main file, full protocol in cross-squad-communication/SKILL.md):

1. Keep in squad.agent.md (under ~40KB target):

   • State backend handshake + HARD RULE (fix(prompt): coordinator must probe for squad_state/memory tools before mutating state (closes #1305) #1306)
   • Routing trigger table only (the action cells can be one-liners pointing at satellite skills)
   • Always-on roster identity (Scribe, Ralph, Rai, Fact Checker)
   • Hard rules (eager-execution exception, ask_user enforcement, spawn-a-squad)
   • Source-of-truth one-paragraph summary with → load coordinator-source-of-truth skill for full hierarchy

2. Extract to satellite skills (lazy-loaded via the skill tool when the routing table fires):

   New skill	Source section	~Bytes saved
   coordinator-routing (full algorithm + examples)	Routing	7,000
   coordinator-team-mode	Team Mode	5,000
   coordinator-source-of-truth	Source of Truth Hierarchy	4,000
   coordinator-response-mode	Response Mode Selection	3,500
   coordinator-spawn	How to Spawn an Agent + Role Emoji	4,500
   coordinator-init-mode	Init Mode Phase 1 + Phase 2 + Casting	5,000
   (existing cross-squad-communication)	unchanged	—

3. Routing-table action cells become: Load coordinator-{xxx} skill, then follow the {xxx} protocol. Same shape as the spawn-a-squad row (fix(coordinator): route "spawn a squad" / "another squad" prompts through the cross-squad skill #1307) and the squad-commands row (feat(skills): make /squad a real slash command (user-invocable: true + rename squad-commands → squad) #1303).

Target

• squad.agent.md: 83KB → ~40KB (51% reduction)
• New coordinator-* skills installed alongside the existing ones via MANIFEST_SKILL_NAMES
• Behaviour unchanged — same triggers fire the same actions; just the action knowledge is loaded on demand

Why now

• v0.10 stabilisation is done and shipped (@bradygaster/squad-cli@0.10.0-insider.1); good baseline to measure against.
• We just did the exact same pattern with cross-squad-communication (feat(sdk): bundle cross-squad-communication skill (#5 cross-squad protocols) #1295) and it worked well — the routing-table row + Hard trigger — keyword-to-skill match paragraph (fix(coordinator): route "spawn a squad" / "another squad" prompts through the cross-squad skill #1307) is the load mechanism we'd reuse.
• Every new feature added in v0.11+ will otherwise pile more bytes onto a file that's already at the ceiling.

Risks

• Regression risk on coordinator behaviour if a satellite skill isn't loaded when it should be. Mitigated by: (a) reusing the proven Hard trigger — keyword-to-skill match rule from fix(coordinator): route "spawn a squad" / "another squad" prompts through the cross-squad skill #1307, (b) regression tests in test/template-sync.test.ts like the ones we already added for cross-squad, (c) staged rollout (extract one section at a time, ship as separate PRs).
• Cold-start cost: each satellite skill load is an extra ~50ms file read. Acceptable given the prompt-size savings.
• Coordinator may overlook satellite skills if the routing row + hard-trigger paragraph isn't airtight. Same failure mode we fixed in fix(coordinator): route "spawn a squad" / "another squad" prompts through the cross-squad skill #1307; same defences apply.

Out of scope

• Tier-aware loading by session-frequency (e.g., "load coordinator-routing only on first prompt of a session, cache for the rest") — that's a runtime optimisation; this issue is purely about prompt size.
• Slimming the rich agent charters (Rai 4.5KB, Fact Checker 3KB) — those are per-agent prompts loaded only when the agent is spawned, not on every coordinator session.

Validation plan

Before merging the slimming PR:

1. Run the full test/ suite — must pass.
2. Smoke-test the same end-to-end "build a small Node CLI" scenario we used for v0.10.0-insider.1 — coordinator must dispatch + complete in similar time.
3. Smoke-test the routing for each extracted section (spawn-a-squad, init-mode, response-mode-selection, etc.) — each must fire its satellite skill before acting.

Filed by the v0.10 stabilisation effort. Tracking: tamirdresher/squad-stack feat/coordinator-md-slimming (to be created).