repo-ai

Know what a change actually did before you trust it.

repo-ai is a local-first CLI for AI-assisted (and ordinary) software work. It indexes your repository offline, tells you what a change touches and which tests can show a break, runs those tests sandboxed in a disposable copy against the base commit, and reports exactly what was observed — never more. Coding agents are optional: when you opt in, repo-ai can drive them under isolation, frozen acceptance tests and explicit apply/undo.

• Context, offline. Task-aware file shortlists, change impact, test links and "why" explanations from a local index. No agent, model, daemon or network needed.
• Evidence, not vibes. verify --run separates new failures from ones already at the base, flags stale evidence, and says when a verdict rests only on static links. coverage checks a requirement→test map you declare against the tests that actually ran. A PASS means the selected tests passed — never the task is correct.
• Agents, if and when you want them. repo-ai do runs a provider (Bob, Codex, Antigravity) in an isolated worktree with an explicit cost ceiling. Tournaments abstain instead of crowning an arbitrary winner; handoff capsules are integrity-checked; nothing is applied without your command, and undo restores the bytes.

Status: 0.9 is a beta. The verification core is the most tested part (below). Many commands are experimental and say so; the full matrix with evidence is in docs/STATUS.md. Not yet demonstrated: that repo-ai makes a coding agent succeed more often than the same agent alone.

Install

Python 3.10–3.14 on macOS, Linux or Windows. Runtime dependencies: pathspec (+ tomli on 3.10). The PyPI distribution is repo-ai-cli (repo-ai was taken by a similar name); the command and the Python package are repo-ai / repo_ai. If you installed a 0.9.0.devN pre-release wheel, remove it first: pipx uninstall repo-ai.

pipx install repo-ai-cli        # or: python -m pip install repo-ai-cli
repo-ai --version               # the command is repo-ai

Stable GitHub releases also carry the wheel, sdist, SHA256SUMS, a CycloneDX SBOM and build-provenance attestations.

Two-minute first run (no agent)

cd your-git-project
repo-ai quickstart                 # where you are and what to run next
repo-ai init --agent none          # build the local index under .repo-ai/ (+ a .gitignore block)
# ...edit some code...
repo-ai verify                     # read-only: changed units, linked tests, confidence, the command
repo-ai verify --run               # run them sandboxed in a disposable copy of the working tree
repo-ai evidence                   # does the last result still cover the working tree?

Real output (trimmed) of verify --run on a two-file project (calc.py, test_calc.py) after editing mean():

verifying 1 selected test file(s) in the seatbelt sandbox, in a disposable copy of the working tree…
  command: pytest  [inferred from the test files]
  pass    test_calc.py (0.3s)
selected 1 of 1 test files; confidence MEDIUM: some of the change rests on static evidence only (imports or file names), not on runtime evidence

PASSED: every selected test passed (the full suite was not run)
  1 passed
  source tree unchanged
evidence saved; `repo-ai evidence` shows whether it still covers the working tree.
This covers only the tests that ran; it is not proof the change is correct.

repo-ai uninstall --agent all removes everything repo-ai created (--dry-run shows it first).

With Claude Code

repo-ai init --agent claude        # CLAUDE.md block, skills, confidence-gated context hooks
claude                             # context arrives with each prompt when repo-ai has evidence

repo-ai never adds permission allow-rules or bypassPermissions. Details: docs/CLAUDE_CODE.md.

Missions with coding agents (opt-in)

repo-ai agents list                # offline: what is installed; no model call
repo-ai agents route               # which provider would be chosen and why; no model call
repo-ai do "make mean() return 0 for an empty list" --dry-run     # plan only (the default)
repo-ai do "make mean() return 0 for an empty list" --agent codex \
    --allow-write calc.py --run --allow-live-provider              # live, isolated, consented
repo-ai apply <mission-id>         # explicit; re-verifies after applying
repo-ai undo <mission-id>          # restores the original bytes

Execution needs a clean committed checkout, explicit write paths, pytest or node --test, and a sandbox backend (macOS or Linux; missions are not supported on Windows). --tournament codex,agy runs independent candidates against the same frozen tests and abstains unless exactly one distinct candidate passes; shortlist explains an abstention. Recipes: docs/roadmap/REALITY-ENGINE-USAGE.md.

"Live" means a model-generated change was independently re-verified through repo-ai on this project's development machine, with small n. Your account, quota and platform decide what works for you; agents route reports it without spending anything.

Platforms

What has and hasn't been measured

• Verify (held-out-7, frozen, independently labelled): 0 false assurances in 127 failing states across 5 pytest/node --test projects in containers; the working tree was unchanged in 234/234 runs. Slower than the native suite on 3 of 5 small suites. Not qualified: unittest, TypeScript, Jest/Vitest/Mocha.
• Agent benefit: four small preregistered studies: abc-v1 (Bob, 4 tasks) was A 2/4 vs B 3/4 — inconclusive at that n, with one repo-ai false green; abc-v2..v4 hit the ceiling (the bare provider solved every task). An earlier 150-run Claude study found no difference (92% plain vs 90%/88%). Comparative lift is NOT_YET_DEMONSTRATED.
• Retrieval: strong on symbol names, weak on prose (recall@5 0.32 on natural-language queries). See docs/RETRIEVAL.md.

Safety and privacy

Local only, no telemetry. Network use happens only on explicit commands (live missions, --online skill vetting, sandbox pull). Secrets are quarantined or shown as rule + path:line, never the value. Repository text is data, never instructions. Writes never leave the repository or follow symlinks. repo-ai privacy audit lists every store and whether it is committed. Threat model: docs/SECURITY.md, docs/VERIFY-THREAT-MODEL.md.

Documentation

• Getting started · CLI reference · Full guide
• Verify · Status matrix · Architecture · Changelog

Contributing, security, licence

CONTRIBUTING.md · SECURITY.md (report vulnerabilities privately) · CODE_OF_CONDUCT.md · SUPPORT.md · MIT licence (LICENSE); bundled third-party evaluation tasks keep their own licences (notices).