Your AI writes the PR.
Reflow proves it works.

List every flow Reflow tracks for a repo: slug, path, branches seen, and the latest run per branch. Call this first when working in a repo — it tells you what coverage already exists before you write anything. An empty result means no flows yet: author one in .reflow/flows/ (see get_flow on any other repo for the format) and run it with create_run. Args: repo (owner/name), branch? — defaults to all branches.

Fetch one flow in full: the markdown source (intent, plain-language steps, RFL blocks, expectations), recent runs, and the current visual baselines per device. Use before editing a flow or diagnosing a failure — the prose intent tells you what the flow is protecting, and the RFL blocks show how each step currently compiles. Args: repo, slug | path.

Execute one or more flows against a target URL in real browsers. Returns run ids immediately; poll get_run for results. Provide the flow content you want executed (it is upserted content-addressed — safe to send repeatedly) plus repo, branch, commit_sha, and target_url — a PR preview, staging, or your local dev server. Each declared device in a flow's frontmatter runs separately. Idempotent per idempotency_key: reuse owner/repo:pr:sha:slug to make retries safe. Args: repo, ref, commit_sha, target_url, flows[], pr_number?, idempotency_key.

create_run(flows: [signup], devices from frontmatter)
→ { runs: [
    { id: "run_188", flow: "signup", device: "iphone-15", status: "queued" },
    { id: "run_189", flow: "signup", device: "desktop",   status: "queued" }
  ] }

Status and evidence for a run: state (queued|running|succeeded|visual_diff|failed), the step-by-step timeline with which lines ran as RFL vs free text, per-step screenshots, token usage against budget, and proposed_changes if the run healed anything. Poll until terminal. On "failed", the failing step names the expectation that could not be met — treat that as a product bug to report, not a flow to patch. On "visual_diff", fetch get_run_proposed_changes and show the human the comparison. Args: run_id.

get_run(run_188) → status: "running", events:
  ✓ compiled      4 steps from cached RFL
  ✓ step 1/4      fill the signup form          rfl · 1.8s
  ✓ step 2/4      submit                        rfl · 0.6s
  ◐ step 3/4      read email, follow magic link rfl · waiting on inbox
  · step 4/4      welcome screen greets by name
  · validate      end state satisfies intent

Presigned URL for the screenshot at any step of a run — what the browser actually saw, per device. Use when explaining a failure to a human, or when you need to look at the page state yourself before proposing a fix. Args: run_id, step_index, device?.

The heal as a unified diff against the flow file: RFL rewrites (new anchor chains prepended, old kept as fallbacks) and plain-language expectation edits. Review it, then commit it to the PR branch with your own git tooling — Reflow never writes to your repo. If the diff changes an expectation, surface that line to the human: it is a product-behavior decision, not a mechanical fix. Args: run_id. Returns: path, before, after, unified_diff.

Worst-case cost in USD for a batch of flows (runner minutes × devices + storage headroom) against the team's current balance — before anything is spent. This is the only tool that talks money: runs hold their worst case up front and settle to actual seconds on completion, so what you see here is the most a batch can cost. Call before create_run when the balance is low or the batch is large; if the estimate exceeds the balance, tell the human to top up rather than letting create_run fail with insufficient_credit. Args: flows[], devices?.

estimate_run_cost(flows: [signup, checkout], devices: [iphone-15, desktop])
→ { est_max_usd: 0.09, current_balance_usd: 24.61 }

Stop an in-flight run. The browser stops, billing settles to actual seconds used, and the hold releases. Use when a newer commit supersedes the one being tested. Args: run_id.

The model keys this team has registered: label, provider, last used. Never returns key material. Use to pick a provider_key label for a flow's frontmatter, or to tell the human they need to add a key in the dashboard before runs can execute (keys cannot be created over MCP by design).