mirror of
https://github.com/google-gemini/gemini-cli.git
synced 2026-05-15 14:23:02 -07:00
gemini-cli[bot]
1a8c7ec226
This PR implements several measures to reduce the cost of GitHub Actions usage, focusing on the highest-impact areas identified through real per-workflow minutes consumption analysis.
### Summary of Changes
1. **CI Matrix Optimization**: Reduced the `test_mac` matrix in `Testing: CI` to only run on Node.js 20.x.
- **Reason**: macOS runners (especially `macos-latest-large`) are significantly more expensive than Linux runners. Node.js 22.x and 24.x are still covered by the `test_linux` matrix, ensuring core compatibility. OS-specific issues are likely to be caught on the recommended Node.js version (20.x).
- **Impact**: Expected to reduce Mac runner usage by approximately 66% in the CI pipeline.
2. **Pulse Workflow Optimization**:
- Added a check to skip `npm ci` and subsequent steps if no reflex scripts are present in `tools/gemini-cli-bot/reflexes/scripts`.
- Reduced `fetch-depth` from 0 (full clone) to 1 (shallow clone).
- **Reason**: The Pulse workflow runs every 30 minutes. Installing dependencies when there is nothing to run is a waste of resources.
- **Impact**: Eliminates unnecessary dependency installation and reduces clone time for the Pulse workflow.
3. **Brain Workflow Optimization**:
- Reduced `fetch-depth` from 0 to 1.
- **Reason**: The Brain workflow does not require full repository history for its reasoning or metrics collection phases.
- **Impact**: Reduces clone time for the daily Brain workflow runs.
4. **Metrics Reporting Fix**:
- Implemented pagination in `tools/gemini-cli-bot/metrics/scripts/actions_spend.ts` to ensure a full 7-day window is captured.
- **Reason**: The previous 1000-run limit was causing significant under-reporting for constant-rate workflows like Pulse during CI surges, as the sample was saturated by bursty CI activity.
- **Impact**: Provides more accurate cost metrics, revealing the true scale of constant "heartbeat" costs.
### Data-Driven Justification
Analysis of the last 7 days of metrics (`actions_spend_minutes`) showed:
- **Testing: CI**: 4074 minutes (approx. 64% of total spend).
- **macOS Runners**: The primary driver of CI cost due to high per-minute rates on large runners.
- **Pulse Workflow**: Previously under-reported due to sampling limits, but now identified as a consistent baseline cost that can be significantly optimized.
These changes prioritize high-impact reductions in expensive runner minutes while maintaining robust cross-platform testing on the primary supported Node.js version.
Gemini CLI Bot (Cognitive Repository)
This directory contains the foundational architecture for the gemini-cli-bot,
transforming the repository into a proactive, evolutionary system.
It implements a dual-layer approach to balance immediate responsiveness with long-term strategic optimization.
Layered Execution Model
1. System 1: The Pulse (Reflex Layer)
- Purpose: High-frequency, deterministic maintenance.
- Frequency: 30-minute cron (
.github/workflows/gemini-cli-bot-pulse.yml). - Implementation: Pure TypeScript/JavaScript scripts.
- Classification: Optionally utilizes Gemini CLI for high-confidence semantic classification (e.g., triage, labeling, sentiment) while preferring deterministic logic for equivalent tasks.
- Phases:
- Reflex Execution: Runs triage, routing, and automated maintenance
scripts in
reflexes/scripts/.
- Reflex Execution: Runs triage, routing, and automated maintenance
scripts in
- Output: Real-time action execution.
2. System 2: The Brain (Reasoning Layer)
- Purpose: Strategic investigation, policy refinement, and proactive self-optimization.
- Frequency: 24-hour cron (
.github/workflows/gemini-cli-bot-brain.yml). - Implementation: Agentic Gemini CLI phases.
- Phases:
- Metrics Collection: Executes scripts in
metrics/scripts/to track repository health (Open issues, PR latency, throughput, etc.). - Phase 1: Reasoning (Metrics & Root-Cause Analysis): Analyzes time-series metric trends and repository state to identify bottlenecks or productivity gaps, tests hypotheses, and proposes script or configuration changes to improve repository health and maintainability.
- Phase 2: Critique: A technical and logical validation layer that reviews proposed changes for robustness, actor-awareness, and anti-spam protocols.
- Phase 3: Publish: Automatically promotes approved changes to Pull Requests, handles branch management, and responds to maintainer feedback.
- Metrics Collection: Executes scripts in
Directory Structure
metrics/: Deterministic runner (index.ts) and scripts for tracking repository metrics via GitHub CLI.reflexes/scripts/: Deterministic triage and routing scripts executed by the Pulse.brain/: Prompt templates and logic for strategic root-cause analysis (Phase 1:metrics.md) and technical validation (Phase 2:critique.md).history/: Persistent storage for time-series metrics artifacts.lessons-learned.md: The bot's structured memory, containing the Task Ledger, Hypothesis Ledger, and Decision Log.
Usage
Local Metrics Collection
To manually collect repository metrics locally, run the following command from the workspace root:
npx tsx tools/gemini-cli-bot/metrics/index.ts
This will execute all scripts within metrics/scripts/ and output the results
to tools/gemini-cli-bot/history/metrics-before.csv.
Development
When modifying the bot's logic:
- Reflexes: Add or update scripts in
reflexes/scripts/. - Reasoning: Update the prompts in
brain/to refine how the bot identifies bottlenecks. - Critique: Update the prompts in
critique/to strengthen the validation of proposed changes.