Collector: Gemini CLI

[willwashburn]

Target

Google Gemini CLI (~/.gemini/) — official Google agent CLI.

Data

• Path: ~/.gemini/ session JSONL files
• Format: JSONL, SessionEnd hook written into settings.json
• Auth: Passive read + optional SessionEnd hook. Since we control spawning (relay/workforce), passive reader is probably sufficient.

References

• TokenTracker: src/lib/rollout.js:299 (parseGeminiIncremental), src/lib/gemini-config.js (hook install)
• TokenArena: cli/src/parsers/gemini-cli.ts
• TokenTracker's usage-limits.js:319-416: Gemini per-model remaining-token budget endpoint (for burn limits: quota-window tracking across providers #5 quota tracking)

Coverage

• Usage: per-session totals
• Cache: yes (cache_read only; no separate cache_creation tier like Claude)
• Tool calls: yes
• Reasoning: depends on Gemini model (Gemini 2.5 Pro has reasoning; burn's Usage.reasoning covers it)

Gotchas

• GEMINI_HOME env var can override ~/.gemini/ — respect it.
• Gemini's model IDs include variants (gemini-2.5-pro, gemini-2.5-flash, etc.) — verify models.dev coverage. Lazyagent's hardcoded pricing only has gemini-2.5-pro (see their pricing.go); we need broader coverage.

Work

• Implement packages/reader/src/gemini.ts emitting TurnRecord with source: 'gemini-cli'.
• Fixtures under tests/fixtures/gemini-cli/.
• Wire into ingest walker.
• Decide on burn gemini spawn wrapper — check for session-id flag support.
• Add source: 'gemini-cli' to the TurnRecord.source union in packages/reader/src/types.ts.

Priority

Medium-high. Google's ecosystem matters for cross-harness comparisons, and it's a realistic next target after Codex. #5 (burn limits) also wants Gemini quota data.