Evidence-first AI bugfix bot

AI fixes with proof, policy gates, and reviewer-ready PRs.

efix turns a failing reproduction command into a pull request with a minimal patch, test updates, before/after evidence, and explicit rollback notes. It is designed for teams that want auditability, not just plausible diffs.

Get Started See Pricing

Mandatory test policy with documented exceptions
Bounded diff enforcement and per-file justification
Command allowlist + default network-deny policy
Artifacts reviewers can inspect, rerun, and audit

Current status: beta and dogfooding. Honest limitations and hardening notes are shown below.

MacOS terminal style visual showing a failing test, patch diff, and passing verification run for an evidence-first bugfix workflow — A reviewer-facing PR with reproduction, tests, evidence, and rollback sections.

What teams care about

Ship faster without lowering review standards.

Trust
PR body includes proof, not just patch text.
Control
Policies are enforced in code, not left to prompt compliance.
Adoption
Starts as a GitHub Action, so teams can pilot without a platform migration.

Positioning

Why efix is not another generic coding agent

Most coding agents optimize for patch generation. efix optimizes for reviewer confidence: reproducible failures, enforced test expectations, bounded changes, and an evidence artifact trail.

Evidence-first by design

Every run records commands, exit codes, and before/after output in `artifacts/evidence.json` and `artifacts/evidence.md`.

Policy-enforced fixes

Mandatory tests, diff limits, and command allowlists are checked by the orchestrator. The model does not get to skip guardrails.

Action-first rollout

Start with `workflow_dispatch` or `/efix` comment triggers. No new hosted control plane is required for a pilot.

How it works

Reproduce -> Localize -> Fix -> Prevent -> Prove

Reproduce Run the failing command and capture baseline stdout/stderr + exit code.
Localize Use the failure signal and repo file list to form a root cause hypothesis and candidate files.
Fix Generate and apply the smallest practical unified diff patch.
Prevent Add or update tests, or document a valid exception with an alternative prevention note.
Prove Run verification commands and embed before/after evidence in the PR body and artifacts.

MVP note: semantic verification is still limited; current `verify()` summarizes command pass/fail while the policy layer enforces test presence and command execution evidence.

Precision workspace infographic with evidence cards and macOS terminal tiles showing the efix protocol and policy gates — Protocol and policy gates are visible to operators and reviewers.

Precision workspace trust-model visual showing collaborator checks, fork pull request rejection, and scoped git staging behavior — Trust model: who can trigger, where it can run, and what it can stage.

Operator trust model

Designed for guarded automation, not open-ended autonomy

Trusted requesters only: `/efix` comment trigger is intended for trusted collaborators (`OWNER`, `MEMBER`, `COLLABORATOR`).
Same-repo PRs only: fork PRs are rejected in the current workflow for safety and secret handling reasons.
Command policy: repro and verification commands must match `commands.allow`; network tools are blocked when `security.network: deny`.
Scoped staging: workflow commits only expected paths (`artifacts/`, `src/`, `test/`) plus tracked deletions, not arbitrary workspace files.
Evidence trail: failures produce comments/artifacts rather than silent retries or hidden mutations.

Documentation discoverability focus (current snapshot)

The project already added core architecture/getting-started docs. The next docs pass should add dedicated, easier-to-find coverage for authorization rules, fork rejection, “repro command must fail” behavior, scoped staging edge cases, and an operator/security trust model page.

Read hardening process notes

Troubleshooting fast path

Common operator errors to document and surface

These are the high-friction cases worth surfacing on the website and in future docs so teams can self-serve.

Unauthorized `/efix` requester

If the comment author is not a trusted collaborator association, the workflow should decline the request and post a clear explanation instead of attempting a run.

What to tell operators

Add the user as a collaborator/member or use `workflow_dispatch` for a supervised run.

Fork PR rejected

Current workflow restricts comment-triggered runs to same-repo PR heads. This prevents secret exposure and reduces the blast radius of untrusted branches.

What to tell operators

Reproduce the issue on a same-repo branch or run a manual workflow with explicit oversight.

Repro command must fail

efix expects a failing signal. If the command already passes, the run should stop and explain that it lacks a baseline failure to fix and prove.

What to tell operators

Provide a narrower failing test command or a minimal repro script that exits non-zero.

Scoped staging surprises

The workflow stages only expected code and artifact paths. Files created elsewhere during repro/verify will not be committed.

What to tell operators

Move intended generated outputs into approved paths or extend staging rules deliberately.

Plans

Choose a rollout plan that matches your team stage

Start with a low-risk solo trial, move into a supervised team pilot, and upgrade to a procurement-ready enterprise plan when you need identity, audit, and contracting support. Pricing is seat-based with pooled run quotas and explicit overage caps.

Solo entry

Solo / Indie

$9/mo

14-day free trial, then $9 monthly or $90 annually

1 billable operator seat
1 private repository
20 bugfix runs per month included
Evidence artifacts, structured PR body, and policy defaults
Email support

Overage is off by default. Opt in for +25 runs at $10 and set a monthly spend cap.

Start 14-day trial

Recommended

Team Pilot

$24/dev/mo

$30 month-to-month or $24 annual (per billable seat)

GitHub Action + `/efix` comment triggers
Evidence artifacts and structured PR body
Mandatory-test and bounded-diff policies
Operator controls and redaction defaults
60 bugfix runs per billable seat per month (pooled across the workspace)
Email/Slack support during pilot

Overage is opt-in only: +50 runs for $15, billed only until your configured monthly cap.

Start Pilot

Contact sales

Business / Enterprise

Custom

Annual contracts with custom seats, quotas, and rollout support

SSO / SAML and seat governance workflows
Audit logs and evidence retention guidance
SLA options and priority support
Invoicing and procurement onboarding
Security review support (questionnaires + architecture walkthroughs)
Custom overage controls and approval rules

Best for multi-repo rollouts, procurement review, and teams that need contracting support.

Contact Sales

Billing mechanics (defined for launch)

A billable seat is a named human operator who can trigger `workflow_dispatch` or `/efix` in a paid workspace during the billing cycle.
Reviewers and approvers without trigger access are not billable seats.
Included run quotas pool at the workspace level and reset monthly during beta (unused runs do not roll over).
Overage is disabled by default. Admins can enable priced run packs and set a monthly spend cap; runs pause at quota/cap if overage is off.
Monthly plans cancel at the next renewal with no partial refunds after the billing period starts. Annual plans are billed upfront; added seats are prorated and seat reductions apply at renewal.
Seat reassignment is allowed at any time by an admin. Replacements do not create double charges, but seats cannot be shared concurrently.

For benchmark comparisons and price anchors, see the pricing FAQ below.

How to get started

Pilot efix on one repo in under 30 minutes

efix works best as a supervised pilot first. Start with one repository, a small set of trusted operators, and a narrow `commands.allow` policy.

1. Install workflow + source

mkdir -p /path/to/repo/.github/workflows
cp /path/to/veri-pr-agent/.github/workflows/efix.yml \
   /path/to/repo/.github/workflows/efix.yml

# Copy efix source into the repo (MVP colocated layout)

The current MVP expects `src/cli.ts` and the rest of the efix source to exist in the same repository.

2. Add config + secret

cp /path/to/veri-pr-agent/.efix.yml.example /path/to/repo/.efix.yml

Then add OPENAI_API_KEY in GitHub Actions secrets and review commands.allow, diff.max_files, and test.required_suites for your repo.

3. First run via workflow_dispatch

Open the Actions tab.
Select efix.
Run workflow with an exact failing repro command (must currently fail).
Review the uploaded efix-artifacts artifact.

If the repro command already passes, efix has no baseline failure to prove against.

4. Enable trusted `/efix` comment runs

/efix npm test -- --testNamePattern "auth rejects invalid token"

Trusted collaborators only
Same-repo PR heads only (fork PRs not supported)
Command must pass policy allowlist

Docs to read next

Getting Started - setup, configuration, triggers, troubleshooting
Architecture - protocol phases, components, data flow, limitations
Analysis - engineering review and market-readiness constraints
Process Log - hardening sprint decisions and changes

FAQ

Honest limitations (today)

Trust increases when constraints are explicit. These are product truths worth saying up front.

Is this production-ready for unattended use?

Not yet. The current positioning is beta/dogfood. The value is strongest in supervised team workflows where reviewers inspect evidence artifacts and PRs before merge.

Does efix validate semantic correctness beyond test pass/fail?

Partially. MVP verification is command-based and policy-based. Deeper semantic verification is a v1 area.

How does efix pricing compare to other developer tools?

Launch pricing is positioned below premium automated bugfix/review tools while staying above general coding assistant baselines. The benchmark table is kept here for transparency, not as the primary customer-facing plan presentation.

Product	Observed price	Why it matters
GitHub Copilot Business	$19/user/mo	General coding assistant baseline in many teams
CodeRabbit Pro	$24/user/mo annual or $30 monthly	Closest AI PR/review automation anchor
Greptile Enterprise	$30/developer/mo (minimum 15)	Higher-trust codebase reasoning / enterprise posture
Graphite	$20 Starter, $40 Team per seat/mo	PR workflow tooling with team-level review workflows
Cursor Bugbot add-on	$40/user/mo	Premium automated bug review/fix positioning anchor

Benchmark snapshot (2026-02-25). Verify current vendor pricing before customer-facing quotes or procurement decisions.

Why not price higher if it writes code changes?

Early pricing should optimize for adoption, pilots, and case-study generation. Once success-rate evidence and cost telemetry are mature, pricing can move upmarket.