Skip to content

Automatic skill selection with Clef

In Desktop 1.8.0, Kora can use Clef or Clef Flash to select and load relevant, complementary skills before a turn. This requires verified Premium access and Clef enabled for the account. An older installed Desktop build does not gain this behavior simply because the decision service is available.

Open Settings → Smart Token Engine → Clef skill selection to read the automatic loading rules. The section remains visible when the token usage dashboard is disabled; its presence does not mean Clef is enabled for your account.

The Kora administrator manages the decision setting: No, Clef Flash, or Clef. This is separate from the chat model and the STE profile. Free accounts continue to use standard skill detection.

  1. Start a session with skills installed for the requested work. Do not select a skill manually for this test.
  2. Choose an STE profile in Settings or override it for the session in the composer.
  3. Send a concrete request covering complementary aspects of the work.
  4. Read the fork-icon event beside the message that started the turn. It identifies the decision model and the skills actually loaded.
  5. Check the yellow skill badges in the response. Their names identify loaded instructions, rather than merely suggested candidates.

Purple references represent manually selected skills. Yellow badges use the same presentation for automatically loaded instructions and actual skill-tool loads; repeated successful loads of the same skill are not shown twice. A failed or pending tool load retains its status.

Clef loading applies to that turn only. It does not add persistent session skills. Skills activated through the normal skill tool retain their per-session state.

STE profile Maximum automatic skills Maximum estimated instruction tokens
Eco 2 2,000
Balanced 4 6,000
Quality 8 12,000
Auto Uses the resolved Eco, Balanced or Quality limits Uses the resolved profile’s limit

These are ceilings, not quotas. Kora also limits automatic instructions to 5% of the model context window and 25% of the remaining context after the system prompt and conversation. The smallest token limit applies. A session profile override takes precedence over the global profile. Auto resolves from the model’s context-window capacity: Eco up to 128k, Balanced up to 512k, Quality above 512k.

Kora skips oversized instructions rather than cutting a skill body. Duplicate names or identical instruction bodies are not loaded again. Less available context can mean fewer skills; the event reports when the STE budget limits loading. With no available headroom, no decision request is sent.

See Smart Token Engine for profile and context controls. Automatic skill limits still apply when generation optimization layers are disabled.

Kora retrieves up to 64 eligible candidates from skill names, descriptions and triggers, then asks Clef to evaluate them together in one decision request. Complementary guidance can produce multiple selections; overlapping alternatives should not add duplicate context. Shared catalogs can contain skills used by other applications without requiring you to remove them.

The event reports the evaluated candidate count and the eligible catalog size. A no-skill decision refers to that evaluated subset; it does not prove that no relevant skill exists anywhere in the catalog.

Manual selections or skill context already supplied for the turn bypass Clef. The local bare-prompt override also bypasses it. Clef does not change the chat model, execute tools or waive approvals. Stop prevents a late decision from being applied.

Unavailable or uncertain decisions fall back to standard routing; automatic fallback loading in this Premium path respects the same instruction budget. A valid no-skill decision continues without adding Clef instructions.

Use Balanced, a fresh session without manually selected skills, and a React project with accessibility and React-form skills installed:

Review this project’s React registration form for accessibility (keyboard navigation, focus, labels, errors and contrast) and React form handling (state, validation and events). Use relevant complementary skills, avoid redundant guidance, and propose concrete corrections without modifying files.

Compare the names in the Clef event with the yellow response badges. Two skills are plausible, not guaranteed: the result depends on the installed candidates, the decision and the available instruction budget.

The decision service receives the current text request and candidate names and descriptions. It does not receive skill instruction bodies, images, full conversation history or full persistent memory as part of the decision request. Kora reads the selected skill bodies locally and adds them to the chat model’s context for the turn.

  • No Clef event → check that you are using Desktop 1.8.0 with verified Premium access and an enabled decision setting; try a fresh session without manual skills or bare-prompt mode.
  • Only one skill or none is loaded → the candidate set may not contain complementary guidance, or the context budget may be too small; inspect the event and installed skill descriptions instead of forcing a skill count.
  • Decision unavailable / standard routing → the event distinguishes uncertainty, connection errors and timeout; the turn can continue with standard routing. After correcting the issue, send a new message; Kora does not automatically retry an uncertain inference.
  • A skill disappeared on the next turn → Clef loads are turn-scoped; select the skill manually or use normal session activation when you need continuing guidance.