cacli

coding agent CLI

Recovering Changes From Transcripts

cacli apply-transcript <transcript.jsonl> can recover reconstructable file edits from Claude Code or Codex JSON/JSONL transcripts.

Examples:

# Produce a reviewable patch without touching the working tree
cacli apply-transcript session.jsonl --mode generate-patch

# Apply the auto-applicable prefix directly to the current tree
cacli apply-transcript session.jsonl --mode auto-apply

# Approve each recoverable edit interactively
cacli apply-transcript session.jsonl --mode interactive

Behavior:

• The command scans the full transcript before doing anything else.
• It reconstructs Write, Edit, MultiEdit, and apply_patch style changes.
• It flags mutating shell commands that cannot be reconstructed, such as sed -i.
• Automatic replay stops at the first unreconstructable mutating shell command or edit that no longer applies cleanly.
• generate-patch writes a patch for the auto-applicable prefix without modifying the working tree.

Run an agent

A run selects a CLI harness, an inference provider, and an unqualified model ID:

cacli run "Review this diff" --harness claude --provider claude-api --model claude-opus-4-8
cacli run "Review this diff" --harness pi --provider workers-ai --model @cf/zai-org/glm-5.3
cacli run "Review this diff" --harness codex --provider azure --model gpt-5.3-codex

The CLI reads non-secret provider parameters from AWS_REGION, CLOUDFLARE_ACCOUNT_ID, or AZURE_OPENAI_ENDPOINT and checks required credential names or login files before execution. build_launch() returns the command, non-secret environment, environment variables to unset, and credential wire metadata without reading secret values. See the provider design.

Workflows (cacli flow)

Flows are sandboxed Python scripts that orchestrate coding-agent subprocesses across harnesses and inference providers. The orchestration script has no direct filesystem, network, environment, or subprocess access; its side effects go through the injected workflow API, while the agents it launches run with their normal permissions.

Create review_flow.py:

META = {"name": "cross-review", "description": "cross-vendor review"}

async def main():
    phase("Review")
    reviews = await parallel([
        lambda: agent("Find correctness bugs in the current diff", harness="codex"),
        lambda: agent("Find security issues in the current diff", harness="claude"),
    ])
    log(f"Received {len([r for r in reviews if r])} reviews")
    return {"reviews": reviews}

Then run it with a dollar cap:

cacli flow run review_flow.py --budget 5.00

Useful commands:

cacli flow run flow.py [--args JSON] [--budget 10.00] [--resume RUN_ID]
cacli flow runs
cacli flow show RUN_ID
cacli flow docs | less
cacli flow install-skill

cacli flow docs prints the complete DSL reference. Claude users can install the same reference as an agent skill with cacli flow install-skill.

Headline features:

• Harness selection for Claude, Codex, Devin, Cursor, OpenCode, and Pi, with explicit inference providers
• Structured output validated against JSON Schema
• Journal-based resume that replays unchanged calls
• Dollar budgets with resumable budget-exceeded runs
• Sandboxed orchestration scripts with a focused async Python DSL

Development

uv sync                     # installs pinned dev tools (ruff, ty, pytest, prek)
uv run prek install         # once: run the hooks on every commit
uv run prek run --all-files # lint, format, typecheck, and test the whole repo

Hooks come from .pre-commit-config.yaml and need prek; stock pre-commit cannot run it. Tool versions are pinned in pyproject.toml's dev group.