Changelog
What has shipped in Minctrl, dated and honest — company-owned processes with the 45 flows as importable examples, the governed run engine, park/resume, user-configurable connectors with shadow-until-signed-off, BPMN 2.0 import/export, and this docs site.
This is a real, dated record of what has shipped. The API is v0.1 and path-unversioned (see Versioning below and the versioning reference). No version numbers are claimed here beyond v0.1 — the entries below describe the actual work.
2026-11-02 — A blank LLM budget means the platform default
Changed. A company that leaves the monthly LLM budget blank is now bounded by the platform
default (MINCTRL_LLM_MONTHLY_BUDGET_USD, $50) instead of having no monthly limit; Settings → AI
models shows the default in force next to each blank limit, and GET /llm-settings returns
platform_defaults.
2026-11-01 — Review fixes: embedding, model proposals, PHI, budgets, webhooks
Fixed.
- Embedding: a person mints an embed token only for their own email (
embed.signalso needsgate.approve); a host key mintsembed.signonly for a member who allows embedded signing (their own switch on Settings → Embedding, or the company's allow-list withsettings.governance); changing the allowed origins needssettings.governance. - LLM steps: a tool bound live without a declared reversibility (
reversible: true, or a catalog action declared reversible) is only proposed by the model, never run on its word. - Model proposals: the sign tier counts an LLM step's
llm.toolsand the run's proposals; a parked proposal is never signable from Slack or Teams, needs the gate id typed in the embed widget, on the phone and in the console, and every card shows the proposed tool and its arguments. - PHI mode: the retention purge also removes the model content of finished runs' traces (LLM outputs, tool-call inputs/results, proposal inputs); the audit chain receives field names and counts instead of form, extracted and decided values.
- Concurrent signatures: the sign door claims the park before it drives the run, so of two signatures at once only one calls the model or runs a reversible tool.
- LLM budgets: every call reserves its estimate against the monthly budget before it is made
(concurrent runs overrun it by at most one call); a blank per-run cap is the server default
(
MINCTRL_LLM_RUN_CAP_USD, $5). Migrationmigrate_llm_reservations.sql. - Host webhooks: the delivery lease is renewed before every send, so a second worker never sends an event a slow receiver is still being sent; deliveries connect to the address the SSRF guard checked at connect time (no DNS-rebinding window).
- Embed exchange: the claimed framing origin is checked against the request's
OriginandSec-Fetch-Site(403 when they contradict it). - Release bundles: the bundle routes cap the body (
BUNDLE_MAX_BYTES, 4 MiB; 413) and bound its nesting (64 levels, 422) before parsing — a hostile bundle is no longer a 500. - Embed widgets: a sign-off token covers one process unless it sets
all_processes: true; the widget names people instead of showing their email addresses. - Roles: who may claim a ticket's
required_roleis decided by permissions, not by the role's name; the built-in role seed has one authoritative definition, run last (migrate_builtin_roles_last.sql), with one seeding trigger. GET /embed/frame-policyanswers for the one framing origin asked about (origin, read by the UI from the widget request'sReferer) and no longer lists a company's allowed origins.
2026-11-01 — Document extraction
Added. Runs carry documents — PDF, PNG, JPEG, TIFF, WebP and
email (its PDF/image attachments become documents of their own) — attached at the start (Runs
page, API, a webhook's _documents) or while the run waits at a gate. The type is read from the
file's first bytes, never its name; limits on size, pages, count and a run's total; a PDF is
opened only to count its pages; files are encrypted at rest with a key of their own and served
back sandboxed. An agent step's extract block sends them to Claude as native PDF/image input
and fills a declared schema: each field with a confidence and a page citation, checked here
(schema, real page counts; one retry, then the step fails), flagged under its threshold. A gate
with review: "<step>" shows the documents beside the pre-filled fields, the uncertain ones
highlighted; what the reviewer signs replaces the model's values for the steps after it, and an
extraction_reviewed entry names each correction and who made it. Budgets, llm_call entries
and test-suite fixtures (extraction) as for LLM steps; the PHI purge deletes a run's documents
with a content-free entry. New permissions documents.upload and documents.read (every
built-in role). Migration migrate_documents.sql.
2026-11-01 — Embedding, white-label and host webhooks
Added. Embedding and white-label: a host product
mints short-lived embed tokens (≤ 15 minutes, used once, stored hashed) with an embed:issue
API key for one of its users and one resource, and mounts the sign-off queue or one run in an
iframe with the new SDK @minctrl/embed (Minctrl.mount(el, {token, view}), events
minctrl:ready, minctrl:resize, minctrl:signed, minctrl:error). The widget pages are
served with frame-ancestors from the company's allowed origins (every other page is
X-Frame-Options: DENY), take the token from the URL fragment, keep the session in memory (no
cookie), and sign through the one sign door as the token's member — roles, separation of
duties, gate forms and the typed confirmation of a live irreversible gate, checked by the
server on the new embed channel. Settings → Branding: display name, primary colour
(contrast-checked), support email, email sender name and a raster logo (SVG refused) brand the
widgets, the gate digest email and the sign-in page of a verified SSO domain. Settings →
Webhooks: per-company outbound webhooks for run.started, run.parked, gate.signed,
run.completed and run.failed, signed like triggers (Minctrl-Signature), recorded in the
run's own transaction, delivered in order with a cursor that moves only after a 2xx, retried
with backoff, with a delivery log and "Send test".
2026-11-01 — Environments: signed release bundles between installations
Added. Promoting between environments: a
published version is exported as a minctrl.bundle/v1 file signed with the company's Ed25519
key (graph, forms, SLA, decision tables, connectors by tool-id and template — never a
credential), imported on another installation only when its key is trusted (untrusted,
tampered or hash mismatch: refused, nothing created), lands in review with the importer as
author, and publishes only after its connectors are mapped to local connections and another
owner or admin publishes it. Promotion log and drift check per process; key rotation keeps old
bundles verifiable. Settings → Environments; scope processes:promote; permissions
process.promote, settings.environments; SDK methods in Python and TypeScript.
2026-11-01 — Process mining
Added. Process mining: import an event log — a CSV with its
columns mapped from a preview, an XES file (standard extensions, read as a stream, DOCTYPE and
ENTITY declarations refused), or your own run history — as a company dataset with a retention,
and see the directly-follows map with frequencies, median and p90 times, start and end
activities, bottlenecks, variants and a case explorer, filtered by date, variant and activity or
path frequency. Conformance aligns each case with a published version (token replay for very
large logs) and lists skipped steps, unexpected activities and wrong order, with the share of
conforming cases. A map becomes a process draft through the save door. A refused row is reported
with its line; a partial import is stored only when asked and is marked partial. New permissions
mining.read, mining.manage, mining.draft; audit kinds mining_dataset_imported,
mining_dataset_changed, mining_dataset_deleted. Migration infra/migrate_process_mining.sql.
2026-11-01 — LLM agent steps (Claude)
Added. An agent step may declare an llm block — model (Claude Sonnet 5.5 by default, Opus
5.5 or Haiku 5.5), system prompt, instructions, output JSON Schema, connector tools — and
Claude does the step: the answer is checked against the schema
(one retry with the error, then the step fails), written to run state as llm.<step>.<field>
and may route the next decision. Reversible tool calls run through the connectors; an
irreversible one only becomes a proposal the step parks on for a person's signature. Every call
is in the trace and an llm_call audit entry (model, prompt hash, tokens, cost, latency,
output hash, stop reason); prompt/output text is kept with the usage record per the company's
model-text switch (never in PHI mode), for its retention. Monthly budget and per-run cap (Settings → AI models, audited): reached → the step
fails with "LLM budget reached", never a heuristic stand-in. Company Anthropic key, write-only
and sealed. A resume never re-calls the model. Test suites replay recorded llm_output. New run
status failed.
2026-11-01 — Approvals in Microsoft Teams and on your phone
Added. Teams and mobile: each company connects its own Azure
Bot (App ID, client secret stored encrypted, tenant) on Settings → Notifications and uploads the
generated Teams app package. People who open the bot get an Adaptive Card when a gate waits;
a plain gate (the Slack rule: no form, one signature, nothing live and irreversible, not a
shadow run) is approved or refused on the card — a refusal with its required reason — through
the one sign door, audited with channel: "teams"; other signers' cards are refreshed with who
signed. Incoming activities are verified against the Bot Framework's signing keys (issuer,
audience = App ID, expiry, serviceUrl, tenant) and accepted once. Minctrl is installable as a
PWA, sends a Web Push (VAPID, RFC 8291 encryption) that says only how many gates wait, removes a
device the push service reports gone, and has a phone-first sign-off screen at /m/signoff with
forms and the typed confirmation of live irreversible gates. The service worker never caches
sign-off data. Migration infra/migrate_teams_mobile.sql.
2026-11-01 — Custom roles and named permissions
Added. Roles and permissions: every check now asks
for a named permission (gate.approve, gate.approve.high_tier, process.publish,
roles.manage, connectors.bind_live, …) instead of a role name. Companies create custom
roles on Settings → Roles from a permission picker grouped by area (or based on another
role), assign them on the Team page, map IdP groups and SCIM to them, and name them on gates.
Nobody grants a permission they do not hold; the owner role stays unique and fixed; a role in
use is not deleted. Role changes are audited (role_created, role_changed, role_deleted)
with the permission sets before and after. The five built-in roles hold exactly what they did
before; the migration seeds them for every company.
2026-11-01 — Self-hosting: Helm chart, Hetzner Terraform, data region
Added. A Helm chart for Kubernetes (API, UI, a migration hook
gated on every file, bundled or external Postgres, Redis, Ingress or Cloudflare Tunnel,
NetworkPolicy, non-root containers, and pg_dump → S3 backups with a weekly restore test) and
Terraform for the Hetzner single-server shape (server bootstrapped by cloud-init, SSH-only
firewall, kept IPs, optional volume and bucket). A deployment setting MINCTRL_DATA_REGION
(eu-central, eu-north, us-east) shows on /health and Settings → Install with the declared
backup location, is recorded in evidence packs, and picks the backup endpoint and server
location; an unknown id stops the API at startup. Several API replicas run the startup
cleanup under an advisory lock (MINCTRL_STARTUP_CLEANUP_LOCK). Tested by lint, render,
schema validation and mocked-provider plans — applying them is the operator's.
2026-11-01 — Decision tables (DMN 1.3 subset)
Added. A process version carries decision tables: inputs
read from run state, outputs, rules in FEEL-lite (literals, < 500, [100..500], lists,
not(...), -), hit policies UNIQUE, FIRST, PRIORITY, COLLECT (+ SUM/MIN/MAX/COUNT)
and an optional no-match default. A new decision step evaluates a table and writes
decision.<step>.<output>; an exclusive gateway's Branch on routes on it. No match without a
default, two matches on a UNIQUE table or a wrong type fail the run at that step (status
failed) with the reason. Tables are in the save door, the diff, the pinned hash, JSON
export/import and test suites; the trace and run_started / gate_signed name the matched
rules. The Processes screen has a table editor with live cell checks and a test panel; DMN 1.3
export and import (unsupported constructs refused with a list, 422); BPMN businessRuleTask
↔ decision step. No migration.
2026-11-01 — Compliance readiness (SOC 2, HIPAA)
Added. Settings → Compliance shows every SOC 2 (TSC 2017:
Security, Availability, Confidentiality) and HIPAA Security Rule control with its live state computed
from the product — SSO required, separation of duties, chain verification and the last external
timestamp, SIEM delivery, backups (from MINCTRL_BACKUP_REPORT_URL, else "not read"), API key age,
inactive members, connectors live vs shadow — and "not read" (never green) when a signal cannot be
read. The evidence pack gains a compliance/ folder: the matrix filled with live states, the access
reviews, the publish/rollback log and incident-relevant events of the period. PHI mode per
company: model text never stored, run inputs purged after N days (phi_inputs_purged, without the
data), every opening of a run recorded (phi_run_viewed). A quarterly access review reminds
owners and records their confirmation (access_review_opened, access_review_confirmed). Matrices,
policy templates, a subprocessor list and a BAA checklist are in docs/compliance/. This makes a
company audit-ready; a SOC 2 report or HIPAA attestation still needs an external auditor.
2026-10-31 — Process test suites and version comparison
Added. A process has test cases: inputs, the answer each
gate gets (outcome, form values, approvers on an N-of-M gate) and the path, final status and
step outputs the run must show. A suite runs against any version, never live — every connector
shadowed, gates answered from the case, agent steps from recorded fixtures unless a case says
llm: "live" — and is not a process run (not in Runs, analytics, queues or SLAs); one
process_suite_run entry records it. Each case passes or fails with its first divergence;
two versions are compared side by side. A case can be made from a run, and live step outputs
recorded as fixtures. The company switch publish_requires_green_suite (default off) makes
publish and rollback need a green suite on that exact version. New API-key scope
processes:test; both SDKs run a suite and read its result.
2026-10-31 — Escalation, delegation, reassignment, N-of-M approvals
Added. A gate may need several signatures: approvals: {required, distinct_people, order, roles}. Each approval is recorded (gate_approval_recorded) and the run continues only at
quorum; a denial by anyone ends the gate. The card shows "1 of 2 signed" with who; Slack offers
no buttons for such a gate. Members set an out-of-office delegate for a period
(/delegations, the Out of office page): their notices go to the delegate, who signs with
their own role, and the audit names both (on_behalf_of). A manager assigns a waiting gate
to a member who may sign it (PUT /process-runs/{id}/assignment, audited gate_reassigned,
notified); the company switch only_assignee_signs (default off) keeps others off an assigned
gate. SLA levels gain level_actions: notify, reassign_to_role, assign_to_user — they
never sign or reroute. See Escalation, delegation and N-of-M approvals.
2026-10-31 — Publishing: draft, review, published, rollback
Changed. A saved version is a draft and does not run. Its author submits it for review;
an owner or admin reads the diff (steps, connections, gate forms, deadlines, connectors)
and publishes it — with the company's separation of duties on (the default), one who did
not author it; with it off, the author may publish their own version and the entry records
"sod": false. Runs — manual, API key, trigger — start only from
the published version; a draft runs only as its author's shadow test run ("shadow": true,
every connector shadowed). Rollback publishes a retired version again under the same rule.
Every move is in the audit chain with the published version before and after. Processes
saved before this change: the latest version of each is published, older ones retired. See
Publishing processes. A trigger starts only the
published version too: with none, its webhook answers 409 and its schedule records the
reason as last_error and keeps its slot.
2026-10-31 — Triggers: webhooks and schedules start runs
Added. Triggers start runs of a saved process without a person:
a webhook (POST /hooks/t/<token> — the URL is shown once, optional timestamped
HMAC-SHA256 signature Minctrl-Signature: t=…,v1=… refused after 5 minutes, JSON body or a JSONPath mapping as inputs, Idempotency-Key returns the first
run for 24 h) or a schedule (5-field cron + IANA time zone; one run per slot with any
number of replicas, missed slots fire once). Runs record created_by: "trigger:<name>"; a
trigger never signs a gate. Managed by owners, admins or a key with triggers:manage, every
change in the audit chain; a Triggers tab on each saved process; both SDKs gain
triggers.
2026-10-31 — Notifications on park; Slack, Teams and email; sign from Slack
Added. When a gate parks, the people who may sign it are told on their channels — email,
a Slack DM from your company's own Slack app, or a Microsoft Teams card — with personal
preferences and quiet hours; SLA escalations go through the same fan-out. A plain gate (no
form, one signature, no live irreversible action) can be signed from the Slack message through
the console's own checks, recorded with "channel": "slack"; anything else links to the
sign-off card. Every delivery and its failure is on the run. See
Notifications, Slack and Teams.
2026-10-31 — Enterprise identity: SSO, require SSO, group roles, SCIM
Added. Sign in with your company's identity provider — SAML 2.0 or OpenID Connect (Okta, Entra ID, Google Workspace, Auth0) — after verifying your email domain; a test sign-in before it goes live; Require SSO for your domains, with an audited owner break-glass; IdP group → role mapping; and SCIM 2.0 provisioning, where deactivating a user suspends their access at once. See Single sign-on and SCIM.
2026-10-31 — Connector catalog and OAuth2
Added. A connector catalog of ten templates — FHIR R4,
Salesforce, ServiceNow, Jira, Slack, Google Sheets, HubSpot, Zendesk, Stripe, generic REST —
with OAuth2 client credentials and authorization code (state + PKCE), credentials sealed at
rest, and every action declaring reversible; an irreversible one stays shadowed until an
owner binds it live.
2026-10-31 — The audit chain goes out: SIEM sinks, ranged export, evidence pack, RFC 3161 timestamps
Added. Stream the audit chain to Splunk HEC, Datadog Logs or an HMAC-signed webhook — in order,
with a per-sink cursor that moves only after a 2xx, so an outage never skips an entry; export any
range as JSONL or CSV with each row's hash and a verification manifest; download an evidence pack
for an auditor; and stamp the chain head with an RFC 3161 Time-Stamp Authority, which
/audit-chain/verify now re-checks. Settings → Audit & SIEM. See
Audit export, SIEM sinks and evidence.
2026-10-31 — Forms in the gate
Added. A gate can declare a form — text, textarea, number, date, select and checkbox
fields, up to 20 — that the signer fills in on the Sign-off card. POST /process-runs/{id}/resume
takes the values as form and checks them server-side (422 naming the field, nothing
signed); they are recorded in the run's decisions and in the gate_signed audit entry, and
the steps after the gate read them as gate_form.<step_id>.<field_id>. Forms are edited in
the node inspector, kept in JSON export and BPMN (a minctrl:form extension element), and
replay with the pinned graph. See Forms in the gate.
2026-10-31 — API keys for scripts and integrations
Added. Owners and admins create company API keys (mk_live_…,
shown once) with scopes runs:start, runs:read, processes:read, audit:read; a key is
refused on every other route and never signs a gate (403 "A machine cannot sign a gate.").
Runs a key starts record created_by: "key:<name>"; create and revoke are in the audit
chain; each key has its own rate limit. Both SDKs accept api_key / apiKey.
2026-10-28 to 2026-10-30 — Processes belong to your company; the 45 flows become examples
Changed. Minctrl no longer ships built-in processes into a company. A new company starts
with no processes and adds its own: build one on the canvas, import a BPMN 2.0 model or a
minctrl.process/v1 JSON file, or describe it in text and review the draft. The admin's
Processes and Runs pages show only your company's processes.
- The 45 flows are examples. Each page on minctrl.com/flows
links a
minctrl.process/v1file athttps://minctrl.com/examples/<slug>.process.json, whichPOST /process-templates/importaccepts unchanged. See Example processes. - Runs name your process. Start a run with
{"vertical": "<slug>", "draft": true}; it is pinned to one version (draft_version, latest by default) and its hash for its whole life. - JSON export / import.
GET /process-templates/drafts/{slug}/exportandPOST /process-templates/importmove a process between environments; a file whose hash no longer matches its graph is refused (422).GET /process-templates/drafts/{slug}reads a saved process back as a graph. - Migrating. Calls that ran a built-in by name (
{"vertical": "aml"}withoutdraft) are still accepted by the API, but nothing in the admin uses them. Import the example and run it with"draft": trueinstead.
v0.1 — 2026-07-02
The first coherent release of the governed-orchestration platform: an engine that runs a process to a human gate, parks, and resumes on a signed verdict — with every irreversible side effect held behind a connector you opt in to.
Governed run engine + park/resume
- Durable process runs. A run executes a process's flow until it reaches a human gate,
then parks (
status: awaiting_human) and persists. It resumes when aresumesignal arrives carrying the gate's step-id and the human'sdecision, which is merged and the flow re-driven idempotently. See Governance model and Run a governed process. - Exclusive / parallel gateways in flows route on the recorded verdict, so a single gate
decision (e.g.
approved,SAR signed,sanctions hit confirmed) steers the run down the correct downstream branch.
User-configurable HTTP connectors, shadowed until signed off
- Bring-your-own connectors. Steps map to HTTP connectors you configure per step, so a step can call your real systems of record.
- Shadow-until-signed-off. Irreversible steps (
reversible: false) run shadowed (dry-run) by default and only fire for real once (a) the connector is deliberately opted in and (b) a human gate upstream has collected a signed verdict. Governance is the default, not an add-on. See Connect a step and Configure gates and autonomy.
BPMN 2.0 import/export
- Round-trip BPMN 2.0. Import an existing BPMN 2.0 process and export Minctrl flows back out, so the governed flow lives alongside the notation your organization already uses. See Import / export BPMN.
45 governed flows across 11 domains
- 45 process flows, spanning 11 domains — fintech, finance, healthcare, HR, insurance, IT, legal, logistics, security, supply-chain, and telecom. Each is an end-to-end flow with primary-source steps, exclusive/parallel gateways, human gates, and an irreversible-execute step held behind a gate. Changed 2026-10-28: these were built-in templates at the time; they are now examples you import (see above).
Docs site
- This documentation site launched — quickstart, tutorials, guides, cookbook recipes, concepts, an API reference generated from the OpenAPI schema, and the two SDKs generated against that same schema.
Versioning
The API is v0.1 and path-unversioned — there is no /v1 prefix in the URL; you call
/process-runs, /auth/*, and so on directly. While the API is at v0.1 it should be treated
as pre-1.0: the surface can still change as endpoints are hardened.
How breaking changes will be communicated. Until a stable v1.0 is cut:
- Every breaking change is recorded here in the changelog, dated, with a note on what moved and how to migrate.
- Additive changes (new endpoints, new optional fields) are not treated as breaking and may land at any time — the generated SDKs surface them automatically.
- When the API stabilizes, a versioning scheme and deprecation policy will be published in the versioning reference before anything is removed.
Pin the SDK version you build against and read this changelog before upgrading. During v0.1, assume any endpoint not listed in Rate limits as hardened may still change shape.