dashboard@0.107.0

Minor Changes

• ae7f01b: The assistant detail panel is now fully configurable and observable. Overview settings (name, model, concurrency, warm TTL) are editable in place. The Sessions tab shows aggregate stats (sessions, messages, cost, tokens) over a selectable time range defaulting to the last 30 days, with per-session cost in the list. Triggers expand in place to show their recent traffic via the new triggers.listEvents endpoint, with each dispatched event linking to the conversation it was routed to.
• ca0f1c1: Encrypt platform OpenRouter API keys at rest. New keys are written with an
  AES-256-GCM encrypted copy alongside the plaintext column (dual-write during
  the expand phase), every read path prefers the encrypted copy and lazily
  records ciphertext for legacy plaintext rows, and the credits monitoring
  activity decrypts inside the activity boundary. A new platform-admin
  adminOpenRouterKeys service and dashboard page list every organization's
  keys with their credit limit, live usage, and encryption state, with actions
  to encrypt (verify the ciphertext, then clear the plaintext), enable, and
  disable a key. Enable and disable actions are audit logged against the owning
  organization; the encrypt action is internal storage hygiene and is not
  surfaced in customer-visible audit logs.

Patch Changes

• e3bb138: Render @tool and /skill references in the assistant composer as colored
  chips, and make the composer a contenteditable so they can be real inline
  elements rather than paint under a transparent textarea.

  A textarea holds one flat string, so a reference inside a draft could only be
  mirrored underneath the input — and anything painted there that occupied width
  slid the caret off the glyphs after it. The input is now a plaintext-only
  contenteditable whose chips are contentEditable={false} spans with real
  padding, so the browser places the caret around them. The draft still lives on
  the runtime as a string: the element reports edits as text and exposes
  value / selectionStart / selectionEnd / setSelectionRange, so the
  mention autocomplete, prompt recall, and dictation are unchanged.

  Skills move into the draft with them. Picking a skill writes its /name token
  into the text (and focuses the input, caret at the end) instead of only
  toggling hidden state, and the composer derives the attached-skill set back out
  of the draft — so deleting the token detaches the skill, and a message carrying
  nothing but a reference can be sent. Tool names containing hyphens now match at
  all; hyphenated source-slug names previously stopped at the first hyphen and
  resolved to nothing. The badge rows above the input are gone, since the tokens
  name themselves, and sent messages render the same chips in a bordered
  white bubble.

• c66958e: Agent sessions routed through LiteLLM keep their LiteLLM association even when the agent's own hook stream captures the transcript: they match the LiteLLM platform filter and display as " via LiteLLM" in the session list and detail views.

• d893bcb: Fix natural-language input in the dashboard time range picker. The picker's
  "type any date" parsing POSTs to /chat/completions, which requires both
  Gram-Session and Gram-Project headers, but most pages rendered the picker
  without a projectSlug, so the request 401ed and parsing silently did nothing.
  The DashboardTimeRangePicker wrapper now injects the request project slug
  via useProjectSlugForRequests() (callers can still override it), fixing the
  project home, MCP overview, security overview, watchdog, and risk overview
  pages in one place — and org-scoped pages like Billing through the same
  default-project fallback every other SDK request uses.