Skip to main content
Approval presentation ordering.
New approval holds resolve their audience before scheduling the first card. An independent fallback becomes eligible after five seconds if resolution has not released presentation. Existing configured channel routing and warnings apply when no candidate is known; later added candidates receive only their targeted presentation. The interval is a scheduling window, not a delivery deadline. Rollback controls now refuse older source that lacks the compatible handoff and receipt protocol.
Console focus rings are easier to see.
  • When you move through the Console with the keyboard, the focus ring on buttons, inputs and other controls is now drawn in full-strength steel instead of a faded tint.
  • The ring now measures at least 6:1 against the page background in light mode and nearly 8:1 in dark mode. The faded version measured below the 3:1 minimum for focus indicators.
Tasks across projects over the API.
  • GET /api/v2/projects/admin/tasks lists tasks newest first (projects:read), across projects or of one (projectId), with the Console’s status, taskKind and limit filters. It answers only ordinary project tasks whose project the key may read, never a factory or project-less task; a Personal key reads as its owner, narrowed to the key’s scopes. A project the key may not read answers 404. See Projects and staffing.
Listing analyzed runs needs workflows:read.
  • GET /api/v2/runs/listAnalyses and the MCP tool forge_list_run_analyses need workflows:read at the platform scope, the key the Console’s list of analyzed runs uses, instead of platform:manage. Your scopes must still reach runs:read, and you still see only the runs you can read.
  • Who loses access: a credential or custom role that holds platform:manage without workflows:read is now refused, with a 403 naming workflows:read. That means a Personal key or External App scoped to platform:manage and runs:read without workflows:read, or a custom role that grants platform:manage alone. No shipped role bundle is affected: the only one that holds platform:manage, platform_admin, also holds workflows:read.
  • Starting a run’s analysis still needs platform:manage.
Reading a run's analysis needs workflows:read.
  • GET /api/v2/runs/getAnalysis and GET /api/v2/runs/listAnalysisMessages, and the MCP tools forge_get_run_analysis and forge_list_run_analysis_messages, need workflows:read at the platform scope, the key the Console’s analysis reads use, instead of platform:manage. Your scopes must still reach runs:read, and a run you cannot read is still a 404.
  • Who loses access: a credential or custom role holding platform:manage without workflows:read loses them (403 naming workflows:read): a Personal key or External App scoped to platform:manage and runs:read without workflows:read, or a custom role granting platform:manage alone. No shipped role bundle is affected: platform_admin, the only one with platform:manage, also holds workflows:read.
  • Starting an analysis with POST /api/v2/runs/startAnalysis still needs platform:manage.
Run canvas: readable blocks, folded to what ran.
  • The run page draws the workflow as blocks, top to bottom, at 80% zoom or closer. Parallel branches sit between rails that say how they join, and routes that leave the workflow are listed beside their step.
  • An ended run opens folded to what ran: routes a step did not take collapse into a N not run box. Show full definition shows every step with what never ran dimmed in place.
  • Follow keeps a live run’s newest running step in view, and a restarted run has an Attempt picker for its earlier attempts.
  • A failed step’s card shows its error as a sentence, and the Acting for org unit now appears in the step inspector’s header. Hover peek cards are gone: every card is readable at the canvas’s zoom.
The skills granting a sandbox command over the API.
  • GET /api/v2/tools/admin/commands/grants lists the skills that grant a sandbox command (agents:read), named by the command’s commandId, newest grant first. Each grant answers createdAt and its skill’s skillId, name, description, source and status; no row id is answered. See Tools.
Cost summaries wait for sign-in.
  • The Console’s cost alerts, the Run History cost card and the analytics token summaries now ask for their totals only once your sign-in session is ready.
  • The token-events summary refuses a signed-out caller instead of answering nothing. A signed-in caller without cost:read where it asks is still answered nothing.
Workforce repair previews over the API.
  • GET /api/v2/people/admin/integrity/findings now answers each finding’s findingId. The id names a finding within its run and never carries the person it concerns. It changes when the deployment’s signing key rotates; list the findings again.
  • POST /api/v2/people/admin/integrity/repairs/preview (workforce:manage) previews a repair from an audit run’s findings, named by findingId, and answers each action by its finding’s id and kind, with before, after and effects carrying public ids only (null for a record with none). Apply it with POST /api/v2/people/admin/integrity/repairs/apply. An id the run does not hold answers 422; an audit run, or a finding whose people are out of the key’s reach, answers 404. See People and members.