Context window optimization: decisions.md pruning, spawn template dedup, Init Mode compression #37

Problem

decisions.md has ballooned to 298KB (~80K tokens) — that's 40% of a 200K context window consumed on every agent spawn. Combined with squad.agent.md growth (29K tokens) and agent histories (~47K tokens), each agent spawn now costs 82-93K tokens (41-46% of context). This leaves agents working with barely half a context window.

Root Causes

1. decisions.md bloat — 266 blocks, 131 of which are sub-section fragments from old analysis docs. No pruning/archival mechanism exists (unlike history.md which has Scribe summarization). Most blocks are stale sprint artifacts from v0.3.0 planning.

2. squad.agent.md spawn template duplication — Three near-identical spawn templates (background, sync, generic) at lines 592-809, each ~1,800 tokens. They differ only in mode/description examples and placeholder names.

3. Init Mode dead weight — Init Mode (lines 28-112, ~1,471 tokens) is loaded on every message but only used once per repo lifetime (when .ai-team/team.md doesn't exist).

Solution

Immediate (this issue — v0.4.0)

• Prune decisions.md: Manual pass — keep ~30-40 active decisions, archive rest to decisions-archive.md. Target: 298KB → ~50KB (~13K tokens)
• Dedup spawn templates: Collapse 3 templates into 1 generic template with mode as a variable. Save ~3,600 tokens
• Compress Init Mode: Reduce redundancy, compress post-setup sources. Target: ~1,471 → ~800 tokens
• Update README: Context Window Budget table with post-optimization numbers

Future (v0.5.0 — automated)

• Add Scribe-driven decisions.md pruning: identify blocks older than N weeks, summarize, archive originals to decisions-archive.md
• Similar to existing history.md summarization (Scribe step 6) but for the shared decision ledger
• Add an evergreen tag for decisions that should never be pruned (e.g., "zero dependencies", "cost-first model selection")

Impact

Labels

type:chore, priority:p0, release:v0.4.0