Don't let broken agent changes merge.

Your coding agent returns a plausible answer and the tests pass, but it now loops, skips verification, or rewrites a test to hide a bug. Maida checks an agent change before merge. It checks how the agent worked alongside your tests of the result.

Use Python 3.12–3.14 and an existing Git repository with Claude Code:

uv tool install "maida-ai==0.6.1"

cd my-repo
maida init

# Run one normal Claude Code task and exit the session.

maida check
# Then run the exact "View:" command printed by Maida.

Approve init's setup preview, then start a new agent session. A successful first report says 3 active checks passed, identifies your task, and prints the viewer command. Open it to see the task's execution timeline. You need no tutorial clone or agent-code changes. If Maida is already in the project's uv environment, prefix its commands with uv run; plain claude still works.

For example: maida view 83aa19e3. Use the command from your own report.

Runs on your machine or CI runner. No Maida cloud account required. Task evidence is not uploaded to Maida; your coding agent still uses its normal provider.

The first report checks completion, recorded loops, and guardrail events. Next, review the observed behavior and keep a baseline and policy so you can compare the next change. Instructions, skills, tools, model configuration, harness code, and application code can all change behavior. Maida gates the resulting change whether a human or the agent authored it.

Keep ordinary tests and evals: a Maida result covers observed behavior, not answer correctness. Once a local pass → safe failure → repair works, add the PR gate and repository protection.

Comparisons report PASS, FAIL, or INCONCLUSIVE. Exit 0 includes INCONCLUSIVE; read the verdict rather than treating process success as approval.

In the canonical storefront demo, a coding agent simplifies shipping, then changes the VIP test to approve $15 shipping instead of $0. All four application tests pass. Maida fails the agent change because the agent rewrote the protected regression test. The deterministic rehearsal uses released Maida v0.6.1 and runs offline after installation.

For a smaller canned example without a clone, run maida demo --regression: expect a FAIL verdict and PR-comment preview. The rehearsal exits 0 when that expected failure is reproduced.

• Maida: the engine, CLI, and public contracts. Start here to check your own agent change.
• Maida tutorials: the canonical runnable coding-agent project, walkthroughs, and examples.
• Maida Action: the GitHub PR boundary around the same engine and CLI.

• Product skills: instrument an agent, add the gate, and debug regressions.
• OpenCode plugin: record local OpenCode traces for Maida.
• TypeScript trace writer: a limited write-side mirror for TS/JS integrations; Python owns behavior and schema.

• Check your own task: setup, first report, and protection for the next change.
• Investigate a regression: see what the agent did.
• Add the PR gate: CI and repository protection requirements.
• Integrate another agent/framework: supported capture integrations; Python agent walkthrough.
• Product skills: help with setup, adding a gate, or debugging it.
• CLI and technical reference: commands, coverage, configuration, and data formats.

Website · Documentation · PyPI

The website source, this organization-profile repository, and the Action smoke fixture are public infrastructure maintained for Maida. They are outside the product and extension hierarchy.