v0.7.0

Added

• Chat. Talk to an agent that acts on Druks as your account. A conversation is
  a live agent session in a sandbox that your account leases, not a workflow. Send
  stays available during a reply, with Stop beside it, and a queued message waits
  its turn. After a restart, Druks runs an interrupted delivery again. The list has
  title search, pins that stay with your account, and day groups in your timezone.
  Chat runs on Claude in this release. Its harness, model, billing, and effort come
  from its row in Chat → Bots. The guide is docs/chat.md.
• WhatsApp is a Chat source. Connect a WAHA service under Connections →
  Services, then link a number from an app's Channels tab with a QR code. A person
  who writes to the number gets their own conversation, and Druks sends the reply
  through WAHA. The person who holds the number's phone is its admin. A run that
  waits asks that admin in the phone's chat with itself, and only the admin can
  answer. A message typed on the phone pauses the chat for four hours. An operator
  links their own number under Connections → Accounts. WAHA must run the NOWEB or
  GOWS engine and reach /_external/waha/events/. Set urls.webhook_host or
  urls.endpoint before you link a number.
• An app can declare a Bot. Bot(prompt, user_tools, admin_tools) from
  druks.agents answers the numbers linked to the app. A route tagged bot is a
  tool of the Bot, and it takes the person who writes as user: BotUser. A run
  that a bot tool starts remembers its conversation, so self.review() asks the
  number's admin. The Bot has a row in the app's Settings → Bots as <app>.bot.
• A YesNo gate. druks.workflows.YesNo parks with the actions yes and no
  and an optional note. A gate's own fields shape its review: the note box shows
  only when the gate declares note, and the values of action are the controls
  when the ask names none. The review shows the question in the ask's label.
• A page can put actions on more blocks. A Callout takes controls, a Form
  takes extra_actions, a Table takes select with row actions, a
  StatusValue takes a link, and a ControlsValue puts actions and links in a
  cell, a fact, or a list item. An operation that answers {"url": "https://..."}
  sends the browser to that URL.
• A skill collection installs from a private repository when the GitHub App is
  installed on it.

Changed

• Activity is compact and app-scoped. Rows take two lines under day headings,
  and a panel shows the details. Filters and the date range live in the URL. A type
  belongs to its app. Search matches the recorded work key or title. Rows show work
  titles, PR references, review summaries, and failure causes. A spent budget
  records as spend_limit, and the run fails without a quota wait. The live stream
  resumes from a transaction snapshot, so a late commit does not disappear.
• A subject's stable handle is key. get_key(), subject.key,
  SubjectSummary.key, and subjectKey on the wire replace the label. A migration
  renames the column. Rename an override of get_label() to get_key().
  get_summary() projects the id and the key; override it for a title. An
  announcement that names title, run, or kind raises WorkflowError.
• A Choices source returns dicts. Each choice is {"value", "label"} with an
  optional group. This replaces (value, label) pairs. The select is searchable
  and grouped. A stored value that the source does not list shows under
  Unavailable choices, and save and druks doctor reject it.
• Schedules is a full-width page. Rows group by app. Preset and pause changes
  save at once. Each row shows a next-run estimate and up to eight recorded
  invocations. Run now queues one invocation, and a paused schedule stays paused.
• Software Factory pages are easier to scan. Projects has collapsible headings,
  flat repository rows, and an inline repository profile. Overview has compact
  desktop rows. A retry shows as a continuation of its build, with the summed time
  and cost, and Retry works only on the latest run of a work item. The work item
  shows a finished call's artifact as a tab beside the transcript, and the
  evaluation is that artifact. An external review request shows as information
  with a link to its source.
• The evaluator is quieter. It files no follow-up tickets. Its findings stay in
  review_notes and in the GitHub review, and the operator decides what becomes
  work. After the first round, a review says only what changed. A rework push
  resolves the review threads it addressed. A review request keeps the note around
  the mention. A retry with no failed step forks from the last declared step. Jira
  gets no transition when the status already matches.
• The GitHub App needs Administration: write to create a repository from a
  template. An existing App must add the permission and accept it on the
  installation. The manifest routes now serve under /api/core/services/github/.
• urls.webhook_host is the HTTPS host of this installation's /mcp endpoint.
  When it is empty, urls.endpoint is the base. docs/deployment.md describes an
  installation behind an external proxy.
• Connections explain more. Removing an MCP server lists the accounts whose
  connections it revokes. MCP sign-in shows "Preparing sign-in" and gives discovery
  and client registration 30 seconds in total. A failed identity lookup records its
  reason on the connection.
• Codex reserve quota is separate from normal capacity. An empty reserve raises
  no exhaustion alert while the main quota has capacity. The Usage page keeps the
  reserve collapsed until the main quota is exhausted. Codex model variants are
  priced by their family, so a dated gpt-5-mini id gets the mini rates.
• A lock that a dead process holds frees in 30 seconds. Provider and OAuth token
  refresh, the sandbox rotation gate, and the browser writer lock renew one TTL. A
  run that loses its lock fails with LockLostError. Reading a saved browser
  session's status no longer decrypts its profile.
• The sandbox and browser images boot with the Drukbox base entrypoint. The
  SSH key, the caller environment, the secrets-proxy CA, and the gh git
  credential come from the base. Both images have build-essential. The browser
  image installs Google Chrome on arm64 too.
• druks create app writes hyphenated folder and distribution names. The app
  name and its package keep underscores. The command accepts both spellings.
• The execution-defaults error names the agent that a default change breaks.

Fixed

• Druks resumes every run after an upgrade. A run parked at a gate before an
  upgrade stayed pending, because any edit to any workflow changed the DBOS
  application version. The version is now one constant, so DBOS resumes parked and
  interrupted runs after every upgrade. A run that the new code cannot replay fails
  at the step that differs, with an error that names it.
• The 0.6.0 upgrade completes on an installation with older agent calls. Its
  billing-source check failed on calls from before provider-namespaced models. The
  check now applies to new calls, and older calls stay unbilled.
• A saved browser profile keeps its login cookies. The launcher writes the
  session-only cookies at shutdown and adds them back at launch.
• A templated sandbox boots its template image. Druks sent the installation
  image with the template, so Drukbox booted the plain image and lost the setup.
• Six open dashboard tabs no longer block a seventh. A hidden tab closes its
  live streams and opens them again when shown.
• Hosted deployments run the current Drukbox janitor, so expired sandboxes are
  removed.
• A page Link to an /api/ path loads from the server. Before, the shell
  handled the click and nothing happened.
• A bare repository name in a ticket matches only that name. % and _ no
  longer act as wildcards.
• Druks installs dbos below 3. DBOS 3 refuses the queues at import.
• The Software Factory timeline shows one running row for an active agent call,
  and a work-item row omits the PR cell until a PR exists.