Skip to content

dashboard@0.84.0

Choose a tag to compare

@gram-bot gram-bot released this 10 Jul 05:47
· 4029 commits to main since this release
c8b1af0

Minor Changes

  • 317d86e: Hook browser login now delivers the minted API key to the local listener as a form POST instead of appending it to the callback URL, keeping the key out of browser history and request logs, and the sign-in tab closes itself once authentication completes. Older dashboards that still redirect with query parameters keep working.

Patch Changes

  • e3760a0: add confirmation dialog for deleting risk policies

  • 11da690: feat: show which users are running the device agent. The org Device Agent page gains an admin-only "Active Users" tab listing who has synced, attributed by the email each agent reports on its ~60s agent.getPlugins poll, with Page.Toolbar search (name/email) and an Active/Stale status filter. A best-effort per-(org, email) last-seen record (throttled to ≤1 write/min) backs a session-secured, org-admin-gated agent.listSyncedUsers endpoint.

  • 832f997: feat(observe): show device agent status on the Employee Enrollment table. A new admin-only "Device Agent" column surfaces whether each member's Speakeasy device agent has checked in — Active (synced recently), Stale (enrolled but not seen lately), or Not Enrolled (no agent activity) — attributed by email via the org-scoped agent.listSyncedUsers endpoint. The column only appears when the gram-device-agent feature is enabled, self-refreshes on a 30s tick, and is sortable so admins can surface who is (or isn't) running the agent alongside their telemetry enrollment.

    The standalone "Active Users" tab on the org-level Device Agent page is removed in favor of this column, so per-user agent activity lives in one place next to telemetry enrollment. The Device Agent page now shows only setup/installation instructions.

  • 74dbfed: feat: add a token usage breakdown to the billing page's Tokens Under Management section (DNO-404). A billing-cycle picker scopes the TUM usage card and a new "Token usage" panel to any contracted cycle; the panel renders a stacked bar chart of org-wide tokens for that cycle, sliced via a grouped, searchable breakdown picker — total, by token type (input / output / cache read / cache write), by risk involvement (tokens from sessions with at least one active risk finding, via the new org-scoped telemetry.queryRiskTokens endpoint), or by analytics dimensions — with daily/weekly/monthly granularity and a cumulative view. Beneath the chart, a usage details table lists per-metric cycle totals with sparklines: token types, agent sessions, tool calls, and message-level stats (tokens in messages with risk findings and tokens from tool-call messages, read from Postgres per-message token counts). The table's measures arrive in a single telemetry.queryTumDetails request, and its totals and time-based overage attribution are normalized to match the billed tokens-under-management numbers exactly, with finalized cycles served from the durable billing snapshots. The section also supports drill-down: clicking a chart bar (or dragging across bars) narrows the whole view to that range (re-bucketing daily), and a time-range picker beside the cycle selector accepts any custom period — typed in natural language or picked from a calendar — with billed normalization and overage reserved for full organization cycles; the usage card is labeled with the billing cycle its totals describe. Cycles are named by month ("June Billing Cycle"), table sections collapse individually or all at once, and a Reset button restores the initial view.

  • 93c2cee: The billing page now reports the billed population exactly. The chart's headline total plots the billed per-day series behind the usage card (matching it to the token), and every breakdown — model, user, division, role, source, token type — reads the new dimensioned billing aggregate with the same qualification and registry-driven source scoping as the billed totals, instead of org-wide analytics aggregates dominated by unbilled agent-fleet telemetry (cache reads included) that overstated usage by orders of magnitude. Assistants usage is tagged but excluded from the billed scope until BYOK. Dimensions billed completions don't carry (provider, account type, skill, MCP server/tool, cache token types) leave the billing page; they live on the costs/insights pages.

  • b0b0ef9: fix(insights): count employee "Agent Sessions" from telemetry instead of the Postgres chat list. The card previously counted chats matched by an email substring search, which only reflected sessions mirrored into Postgres via session_capture and under-reported real activity. It now uses summary.totalChats (distinct chat_ids in telemetry, keyed by the same user), so the count is consistent with every other metric on the employee detail page.

  • 02ac329: Issuer discovery now parses RFC 8414 service_documentation, op_policy_uri, and op_tos_uri and persists them on remote_session_issuers across the project, organization, and global admin surfaces.

  • 0517e60: Restrict the Observe dashboard section (Costs, MCP & Tools Insights, Employee Enrollment, Agent Sessions, Tool Logs) to org admins. The Observe nav stays visible (like the Secure section), but each Observe page is gated on org:admin, so basic members see an "Access restricted" notice. Basic members also no longer receive environment:read by default.

  • dfee73b: feat: surface the AI account email on agent sessions. chat.listChats and chat.load now return account_email from the linked AI account, and the dashboard shows the personal account's email (e.g. a gmail on Claude Max) on session list rows, the transcript's user messages, and the session details popover — instead of only the attributed employee's work email.

  • 4fa3e51: Split the org-admin organizationRemoteSessionIssuers service into three per-resource services mirroring the project-scoped layer: organizationRemoteSessionIssuers, organizationRemoteSessionClients, and organizationRemoteSessions. Pure refactor with no behavior or RBAC change, but breaking for the management API and SDK: every method drops its redundant resource suffix, so the RPC paths and SDK method names change (e.g. organizationRemoteSessionIssuers.createClient becomes organizationRemoteSessionClients.create).

  • c7edd5a: Link entities mentioned in Project Assistant replies to their dashboard pages, styled in Moonshine blue with an external-link icon

  • 3f15c7c: fix: apply the Tool Logs http.response.status_code filter at the trace level so status-less rows no longer leak 200/success traces into "Non-2xx responses", and add a first-class Error/Success/Blocked/Pending Status filter to the Tool Logs page.

  • Updated dependencies [608940e]

  • Updated dependencies [c7edd5a]

    • @gram-ai/elements@1.42.0