v0.2.1

Documentation fix. No behaviour change, no API change.

0.2.0 renamed nine MCP tools into a single ratchet_<verb>_<object> scheme, but the bridge package's own README was missed — so the published tarball documented ratchet_check_effect and ratchet_usage, two names the server rejects. Anyone reading the npm package page and following its tool table would have called a name that does not exist.

Fixed, and the naming tests now read every surface a user takes a tool name from rather than only src/mcp/tools.ts, so a rename cannot be complete in the source and wrong in the shipped artifact again.

The 15 tools themselves are unchanged from 0.2.0.

v0.2.0 — per-destination ceilings

Ratchet is an effect gate for AI agents: an agent asks before it does anything it
cannot take back, and gets a durable decision, so the same real-world action is
attempted at most once across crashes and retries.

This is the first tagged release. Everything below is enforced by a test.

Counting a destination without ever seeing it

A ceiling could be scoped to a workspace, an API key or an effect type. None of
those is a destination, so twenty distinct $500 refunds to one bank account
passed every check. A destination lives in the payload, and Ratchet never stores
payloads.

Counting does not require reading. A caller declares dimensions on begin;
only HMAC(secret, workspace | name | value) is kept. Ratchet cannot say who the
counterparty is and cannot reverse the value — and because the workspace id is
inside the MAC, the same account in two workspaces yields two unrelated
identifiers. It can still count.

PUT /v1/policies/payment.refund
{ "required_dimensions": ["counterparty"],
  "dimension_limits": { "counterparty": { "daily_micros": 200000000,
                                          "daily_count": 20 } } }

A declaration tightens and never loosens: it adds a limit, never removes the
workspace, key or type limits; omitting a required one is refused rather than
allowed; and a retry cannot move an effect into a fresh bucket.

Three analyses over the blinded ledger

• GET /v1/analysis/structuring — amounts crowding a threshold, by comparing two
  equal bands beneath it. Point it at a line Ratchet does not enforce with
  structuring_threshold_micros.
• GET /v1/analysis/fan — fan-out measured by novelty, not width (payroll
  reaches five hundred people every month and is fine), and fan-in, which no
  per-agent ceiling can see.
• GET /v1/agents/{id}/reliability — reporting rate, decision mix from receipts,
  idempotency-key hygiene, declared-versus-actual cost, lease hold.

None of them is a verdict, and each says so: a cap produces bunching honestly, and
a first payroll run is 100% new counterparties.

Also in this release

• Enterprise plan, sold rather than bought — selfServe: false makes checkout
  refuse it structurally.
• Auto-recharge for overages, and card saving for non-subscribers.
• Signup returns a runnable next step, and blocked now names the call that
  clears it.
• Mail survives a spent sending quota instead of being discarded within the
  hour.
• Roughly 12× lower p95 on the gate after removing five sequential round
  trips.

Breaking

MCP tool names are now ratchet_<verb>_<object> throughout. Seven were
renamed off a second noun_noun pattern; heartbeat_effect became
extend_lease and check_effect became get_effect. MCP clients discover tools
at connect time, and the npm bin bridges to the live service, so this needs no
client change.

Full contract: https://ratchetgate.com/openapi.json