Agentic Code Fixer (Gemini Flash)

A production-ready multi-agent AI system that reads your codebase, diagnoses bugs, and fixes them autonomously using an iterative agentic loop powered by Gemini Flash.

Features

• Multi-agent roles: Planner, Reviewer, Executor, Verifier
• Retry with exponential backoff: handles 429/rate limits gracefully
• Proper Python logging: structured, timestamped, verbose mode
• Dry-run mode: preview proposed edits before applying
• Session history: every run saved as JSON for audit/replay
• Git checkpoints: optional commit after each iteration
• Error context input: seed agent with stack traces/logs
• API health check: validate key/model/quota from CLI
• Test suite: unit tests for all core modules

Project Structure

AIAgents/
├── agentic_fix/
│   ├── __init__.py
│   ├── agent.py          # Main agentic control loop
│   ├── api_check.py      # API quota/access health check
│   ├── config.py         # Env + runtime settings
│   ├── context.py        # Repo scanning + context building
│   ├── gemini_client.py  # Gemini API with retry + error handling
│   ├── git_utils.py      # Git checkpoint utilities
│   ├── logger.py         # Structured Python logging setup
│   ├── prompts.py        # Role-specific prompts
│   ├── schema.py         # Data models for plans/edits/results
│   ├── session.py        # Run session history (JSON logs)
│   ├── verify.py         # Safe verification command runner
│   └── workspace.py      # Apply/preview edits
├── tests/                # Unit tests for agent modules
│   ├── test_context.py
│   ├── test_schema.py
│   ├── test_session.py
│   ├── test_verify.py
│   └── test_workspace.py
├── sample_buggy_repo/    # Simple demo target
├── sample_complex_repo/  # Multi-file buggy demo target
├── main.py               # CLI entrypoint
├── requirements.txt
└── .gitignore

Setup

1. Create and activate a virtual environment.
2. Install dependencies:

pip install -r requirements.txt

3. Add environment variables in .env:

GEMINI_API_KEY=your_api_key_here
GEMINI_MODEL=gemini-2.0-flash
MAX_AGENT_ITERATIONS=5
MAX_FILES_IN_CONTEXT=20
MAX_CHARS_PER_FILE=4000
ENABLE_REVIEWER=true

Check API Quota/Access

python main.py --check-api

Performs a tiny live call to validate your key, model, and quota.

Run

python main.py "Fix failing tests" --repo ./sample_complex_repo --verify "python -m pytest -q" --verbose

Run Against Any External Repo

python main.py "Fix billing module bugs" --repo "C:/path/to/repo" --verify "python -m pytest -q"

CLI Flags

Flag	Description
--repo PATH	Target repo folder (default: .)
--verify CMD	Verification command (e.g. pytest -q)
--verbose	Detailed step-by-step logging
--max-iterations N	Override max loop iterations
--no-reviewer	Disable reviewer agent for this run
--error-file PATH	Seed first iteration with error logs
--check-api	Validate API key/model/quota and exit
--dry-run	Preview edits without applying
--git-checkpoint	Git commit after each iteration

Demo With Complex Buggy Repo

python main.py "Fix checkout pricing, shipping threshold, and tax bugs" \
  --repo ./sample_complex_repo \
  --verify "python -m pytest -q" \
  --verbose \
  --error-file ./sample_complex_repo/error_logs/pytest_failure.txt

Run Tests

python -m pytest tests/ -v

Agentic Multi-Agent Architecture

1) Context Builder

• Scans repo files (extension filtering + ignored dirs).
• Produces a bounded snapshot of file contents.

2) Planner Agent (Gemini Flash)

• Receives task + repo snapshot + previous failure context.
• Returns strict JSON with edits, verify command, done flag.

3) Reviewer Agent (Gemini Flash)

• Reviews planner output for safety/minimality/quality.
• Can approve, return feedback, or provide revised plan.

4) Executor

• Applies edits (create / replace).
• Path-escape protection + permission error handling.

5) Verifier

• Runs safe allowlisted commands with timeout.
• Captures stdout/stderr.

6) Reflect + Iterate

• Failed verification output fed back to next planner iteration.
• Optional --error-file seeds first iteration context.
• Loop stops on: success, planner done, or max iterations.

7) Session Logging

• Every run writes a JSON session log to .agentic_fix_logs/.
• Contains all iterations, plans, reviews, edits, and verify results.

8) Git Checkpoints

• With --git-checkpoint, commits after each successful edit iteration.
• Easy rollback to any iteration.