feat: effort-aware auto-compact and budget caps

[kokevidaurre]

Context

Claude Code has sophisticated context window management that we're not tuning:

• Auto-compact: triggers at effective_window - 13K buffer, summarizes old messages via side-query
• Budget caps: maxBudgetUsd stops execution when cost exceeds threshold
• Context window override: CLAUDE_CODE_AUTO_COMPACT_WINDOW controls when compaction fires

Our agents run with defaults regardless of effort level. Quick tasks waste tokens on 200K context windows; deep tasks could use 1M.

From Claude Code Source

// autoCompact.ts
const AUTOCOMPACT_BUFFER_TOKENS = 13_000;
const MAX_OUTPUT_TOKENS_FOR_SUMMARY = 20_000;

function getEffectiveContextWindowSize(model) {
  // Respects CLAUDE_CODE_AUTO_COMPACT_WINDOW env var
  // Falls back to model's native context window
}

Circuit breaker: after 3 consecutive autocompact failures, stops trying.

Proposed Changes

In buildAgentEnv(), inject tuning env vars based on effort level:

const COMPACT_WINDOW = { low: '80000', medium: '180000', high: '800000' };
const MAX_BUDGET = { low: '0.50', medium: '2.00', high: '10.00' };

// In buildAgentEnv():
if (effort) {
  env.CLAUDE_CODE_AUTO_COMPACT_WINDOW = COMPACT_WINDOW[effort] ?? '180000';
  // Budget cap via Claude Code's internal tracking
}

Implementation

1. Add effort→env-var mapping constants
2. Inject in buildAgentEnv() (already has effort param)
3. Document the relationship between effort and context management

Files to change

• src/lib/execution-engine.ts — buildAgentEnv() (~L316-350)

Expected impact

• Low effort: Agents compact at 80K instead of 200K → shorter conversations, faster completion
• High effort: Agents use up to 800K before compacting → deeper reasoning without losing context
• Budget caps: Prevent runaway costs from agents stuck in loops

Risks

• Too-aggressive compaction (low effort) might lose important context. Mitigation: 80K is still substantial (~20K tokens of usable conversation).
• Budget caps too low could cut off agents mid-task. Mitigation: start conservative, adjust based on observability data.

Open questions

• Does Claude Code expose a budget cap env var? Need to verify CLAUDE_CODE_MAX_BUDGET or equivalent exists. If not, we'd need to implement timeout-based cost control.