STRATO Fix All The Things

A multi-agent system for automatically fixing GitHub issues using Claude Code.

Architecture

+-------------------------------------------------------------------+
|                      MULTI-AGENT PIPELINE                         |
+-------------------------------------------------------------------+
|                                                                   |
|   +----------+     +----------+     +-------------------+         |
|   |  TRIAGE  | --> | RESEARCH | --> |   FIX <--> REVIEW |         |
|   +----------+     +----------+     |   (up to 3 loops) |         |
|        |                |           +-------------------+         |
|        v                v                    |                    |
|    Classify         Explore              Implement                |
|    issue            codebase             & iterate                |
|                                                                   |
+-------------------------------------------------------------------+

Pipeline Stages

Agent	Purpose	Key Outputs
Triage	Classify if issue is AI-fixable	Classification, confidence, complexity
Research	Deep codebase exploration	Root cause, files to modify, patterns
Fix	Implement minimal changes	File changes, confidence assessment
Review	Self-review the fix	APPROVE / REQUEST_CHANGES / BLOCK

Fix-Review Loop

If the review agent returns REQUEST_CHANGES, the pipeline loops back to revision:

1. Fix agent receives review feedback (concerns, suggestions)
2. Fix agent revises the implementation
3. Review agent re-evaluates
4. Repeat up to 3 times, then block if still not approved

Agent Details

1. Triage Agent

Classifies issues into categories:

• FIXABLE_CODE - Clear bug fixable with code changes
• FIXABLE_CONFIG - Configuration/environment changes
• NEEDS_CLARIFICATION - Issue too vague
• NEEDS_HUMAN - Requires human judgment
• ALREADY_DONE - Issue appears resolved
• OUT_OF_SCOPE - Not suitable for AI fixing

Only FIXABLE_CODE and FIXABLE_CONFIG proceed to the next stage.

2. Research Agent

Explores the codebase WITHOUT making changes:

• Locates all relevant files
• Maps architecture and data flow
• Identifies root cause
• Documents patterns to follow
• Notes risks and testing recommendations

3. Fix Agent

Implements changes based on research:

• Executes focused changes
• Follows identified patterns
• Does NOT commit (pipeline handles commits after review)
• Reports confidence scores

4. Review Agent

Self-reviews the fix before PR creation:

• Checks correctness, completeness, safety
• Validates style and scope
• Can block problematic fixes
• Provides concerns and suggestions for revision

Quick Start

1. Setup environment:

cp .env.sample .env
# Edit .env with your settings

2. Install prerequisites:

npm install -g @anthropic-ai/claude-code
# Ensure gh and git are available

3. Run on issues:

./run.py 5960              # Single issue
./run.py 5960 5961 5962    # Multiple issues

4. Run from an issue comment invocation (optional):

./run.py 6241 --issue-comment-id 123456789

This mode validates that the issue comment contains a supported @ invocation, posts an acknowledgement comment, and then runs the normal pipeline for the issue.

Configuration

Environment variables in .env:

Variable	Description	Default
GITHUB_REPO	Repository (owner/repo)	blockapps/strato-platform
PROJECT_DIR	Path to local repo clone	Required
BASE_BRANCH	Base branch for PRs	develop
TRIAGE_TIMEOUT	Triage agent timeout (seconds)	120
RESEARCH_TIMEOUT	Research agent timeout (seconds)	300
FIX_TIMEOUT	Fix agent timeout (seconds)	300
REVIEW_TIMEOUT	Review agent timeout (seconds)	180
MODEL_PROVIDER	Default LLM provider (claude or codex)	claude
TRIAGE_MODEL_PROVIDER	Optional provider override for triage	unset
RESEARCH_MODEL_PROVIDER	Optional provider override for research	unset
FIX_MODEL_PROVIDER	Optional provider override for fix	unset
REVIEW_MODEL_PROVIDER	Optional provider override for review	unset
CLAUDE_CLI_COMMAND	Claude CLI executable	claude
CODEX_CLI_COMMAND	Codex CLI executable	codex

Run Output

Each run creates detailed logs in runs/:

runs/2026-01-15_12-00-00-issue-5960/
├── issue.json              # Original issue data
├── triage.prompt.md        # Prompt sent to triage agent
├── triage.log              # Full triage output
├── triage.state.json       # Classification result
├── research.prompt.md      # Prompt sent to research agent
├── research.log            # Full research output
├── research.state.json     # Research findings
├── fix.prompt.md           # Prompt sent to fix agent
├── fix.log                 # Full fix output
├── fix.state.json          # Fix result
├── review.prompt.md        # Prompt sent to review agent
├── review.log              # Full review output
├── review.state.json       # Review verdict
├── fix-revision-2.prompt.md  # (if revision needed)
├── fix-revision-2.log        # (if revision needed)
├── fix-revision-2.state.json # (if revision needed)
└── pipeline.state.json     # Aggregate pipeline results

Confidence Scoring

Each agent reports confidence (0.0-1.0). The pipeline computes an aggregate:

Aggregate = Triage(0.15) + Research(0.20) + Fix(0.35) + Review(0.30)

File Structure

strato-fix-all-the-things/
├── run.py                 # Main entry point
├── src/
│   ├── agents/
│   │   ├── base.py        # Base agent class
│   │   ├── triage.py      # Triage agent
│   │   ├── research.py    # Research agent
│   │   ├── fix.py         # Fix agent
│   │   └── review.py      # Review agent
│   ├── claude_runner.py   # Claude Code CLI wrapper
│   ├── config.py          # Configuration loader
│   ├── github_client.py   # GitHub API (via gh CLI)
│   ├── git_ops.py         # Git operations
│   ├── models.py          # Data models
│   └── pipeline.py        # Pipeline orchestrator
├── prompts/
│   ├── triage.md          # Triage prompt template
│   ├── research.md        # Research prompt template
│   ├── fix.md             # Fix prompt template
│   ├── fix-revision.md    # Fix revision prompt template
│   └── review.md          # Review prompt template
├── runs/                  # Run logs (gitignored)
├── .env                   # Environment (gitignored)
├── .env.sample            # Environment template
└── README.md

How It Works

1. Fetch issue from GitHub
2. Triage classifies if it's AI-fixable
3. Research explores codebase, identifies root cause
4. Fix implements changes based on research
5. Review self-checks the fix
   • If APPROVE: proceed to PR
   • If REQUEST_CHANGES: loop back to fix (up to 3 times)
   • If BLOCK: stop and comment on issue
6. Commit changes with detailed message
7. Create PR as draft with ai-fixes-experimental label
8. Comment on issue linking to PR with details

If any stage fails or blocks, the pipeline stops and comments on the issue explaining why.

Safety Features

• Creates PRs as drafts (not ready for review)
• Review agent can block problematic fixes
• Fix-review loop allows self-correction (up to 3 attempts)
• .env files automatically excluded from commits
• Force-syncs to latest origin/{base_branch} before each issue
• Cleans up existing branches/PRs to avoid conflicts
• Comments on issues even when blocked/failed (for visibility)

Requirements

• Python 3.11+
• Claude Code CLI (npm install -g @anthropic-ai/claude-code)
• GitHub CLI (gh) - authenticated
• git

Roadmap

• Phase 2 model-provider and benchmark work is tracked in docs/model-benchmark-backlog.md.

Benchmark Reporting

Generate benchmark output from historical runs:

python3 scripts/benchmark_report.py --runs-dir runs --json-out runs/benchmark.json --csv-out runs/benchmark.csv

License

Apache-2.0 — see LICENSE