CCA-SWE-Bench: Agent Harness for SWE-bench

This repository contains a minimal but production-usable harness around Confucius Code Agent (CCA) for SWE-bench style coding tasks.

Why this repo exists

The core design keeps the agent inside the target runtime instead of building a heavy host-to-container abstraction layer.

Benefits:

• simple execution path
• low latency tool calls
• fewer moving pieces when debugging
• easier extension of agent capabilities

Tradeoff:

• less platform abstraction compared with runtime-proxy frameworks

What is in this repository

• confucius/: CCA framework core, orchestrators, extensions, IO, model managers
• scripts/run_swebench.py: reference one-shot runner for SWE-bench style tasks
• scripts/cca.py: unified runner (TUI + one-shot + provider runtime)
• scripts/provider_runtime.py: provider auth/model discovery/validation runtime
• tests/: unit and e2e tests for launcher/runtime behavior

Execution Modes

Reference CCA workflow

• Entry points:
  • python -m scripts.run_swebench --prompt <file>
  • confucius code (interactive REPL)
• Focus:
  • SWE-bench style execution with CCA harness extensions

Unified cca workflow (recommended)

• Entry point:
  • cca
• Features:
  • one command for interactive TUI and one-shot execution
  • dynamic model discovery and alias resolution
  • live model ID validation
  • provider runtime setup from existing local OAuth sessions

For a full architecture and workflow reference, see:

• docs/cca-harness-workflow.md

Quick Start (Recommended)

1) Install dependencies

python -m venv .venv
.venv/bin/pip install -r requirements.txt

2) Install global cca command

python3 -m pip install --user -r requirements.txt
python3 -m pip install --user -e .

3) Run

cca

By default, cca enters interactive TUI mode when stdin is a TTY and no prompt is provided.

Provider Authentication

Codex provider

Use your existing Codex CLI login state:

codex login

cca reads:

• ~/.codex/auth.json
• ~/.codex/models_cache.json

It uses those credentials to call the Codex backend API via OpenAI SDK runtime settings.

OpenAI provider

export OPENAI_API_KEY=...

Gemini provider

Either:

• API key (GOOGLE_API_KEY or GEMINI_API_KEY)

Or OAuth flow with local creds:

• ~/.gemini/oauth_creds.json

Bedrock provider

export AWS_REGION=us-east-1
# plus normal AWS credentials/profile

Common cca Usage

Interactive TUI

cca

Or force TUI mode:

cca --tui

One-shot with file prompt

cca --provider codex --prompt /path/to/problem.txt

One-shot with inline text

cca --provider codex --prompt "Fix failing tests" --raw-prompt

One-shot with stdin

echo "Fix failing tests" | cca --provider codex --raw-prompt

Show resolved runtime only

cca --tui --dry-run

Select provider/model explicitly

cca --provider codex --model gpt-5.3-codex-spark --tui
cca --provider openai --model gpt-5.2 --prompt "Update docs" --raw-prompt
cca --provider bedrock --model claude-sonnet-4-5 --prompt /path/to/task.txt

Model discovery and validation

List discovered models and aliases

cca --provider codex --list-models
cca --provider gemini --list-models

Validate discovered model IDs with live API calls

cca --provider codex --validate-models-live
cca --provider gemini --validate-models-live

cca mode behavior

Mode selection is automatic:

• TTY + no prompt -> TUI
• prompt provided -> one-shot
• piped stdin + no prompt -> one-shot

Solo mode defaults:

• one-shot: on
• TUI: off

Override:

cca --solo-mode on
cca --solo-mode off

TUI commands

• /help
• /model
• /exit
• /quit

Reference run paths

One-shot SWE-bench style

.venv/bin/python -m scripts.run_swebench --prompt /path/to/problem.txt

REPL

.venv/bin/python -m confucius.cli.main code

Testing

Unit and behavior tests

.venv/bin/pytest tests/test_cca.py tests/test_tasks.py tests/test_provider_runtime.py tests/test_intent_guard.py -q

Real interactive TUI e2e

Requires valid Codex auth state.

.venv/bin/pytest tests/test_cca_tui_e2e.py -m e2e -q

SWE-Bench Pro (Reproducible)

Use the reproducible runner:

• scripts/run_swebench_pro_codex.py

Full step-by-step guide:

• docs/swebench-pro-repro.md

Docker workflow

This repo supports the packaged SWE-bench container workflow.

Build artifacts

conda activate confucius
conda install -c conda-forge conda-pack
conda-pack -n confucius -o cf_env.tar.gz

pex . \
  -r requirements.txt \
  -m scripts.run_swebench \
  -o app.pex \
  --python-shebang="/usr/bin/env python3"

Expected workspace layout

workspace/
|- app.pex
|- cf_env.tar.gz
|- solutions/
|- logs/
`- problem_statements/
   `- <task_id>.txt

Container run

docker run --rm \
  -e TASK_ID=<task_id> \
  -e AWS_BEARER_TOKEN_BEDROCK=<token> \
  -v <workspace>:/data \
  --network host \
  --userns=host \
  --entrypoint /data/run_sbp.sh

Troubleshooting

Failed to prepare runtime: stdin prompt is empty

You ran one-shot mode without prompt input. Use one of:

• --prompt <file-or-text>
• piped stdin
• --tui

Codex auth not found

Run:

codex login

Then verify files exist:

• ~/.codex/auth.json
• ~/.codex/models_cache.json

Gemini OAuth project errors

Set one explicitly:

export GEMINI_OAUTH_PROJECT=<project-id>
export GEMINI_OAUTH_LOCATION=us-central1

Model not discovered

Use explicit model ID:

cca --provider codex --model <model-id> ...

License

MIT. See LICENSE.