Skip to content

Latest commit

 

History

History
63 lines (54 loc) · 3.57 KB

File metadata and controls

63 lines (54 loc) · 3.57 KB

Observed usage, plans and output

The SDK backend reports public host observations. Muse 1.1.1-R2514.1 and 1.2.1-R2847.1 are verified against local provider fixtures. Native durable compaction is a separate unavailable operation and does not gate these reports. Legacy exec retains its existing output path.

/status is an idle local command. It reports requested model/provider/effort/mode, last observed model/approval policy, root-session token totals, context pressure and plan size. Missing observations say unknown. The command starts no model turn and does not execute accompanying text or attachments. It does not inspect credentials, infer quotas or estimate missing values.

Clients may negotiate clientCapabilities._meta["muse/usage"] = 1. The adapter then announces version 1 and emits optional session_info_update metadata:

{
  "muse/usage": {
    "scope": "rootSession",
    "cumulative": { "promptTokens": 10, "outputTokens": 3, "totalTokens": 13 },
    "lastModelCall": { "turnId": "host-turn-id", "promptTokens": 10, "totalTokens": 13 }
  }
}

Cumulative counters are host-computed replacements, never increments for the client to sum. Optional raw last-call counters retain the host's cache convention; they are not reinterpreted as cumulative values. Native child usage is separate. Known context occupancy and window size use standard ACP usage_update. If the host supplies no size, the adapter does not invent a denominator: negotiated metadata and /status preserve the known facts and unknown fields. No cost or account quota is inferred. Baseline clients need no extension to use /status, standard context updates, plans or tool/summary output.

Todo events replace the entire ACP plan, including an empty list. Native inProgress maps to in_progress; cancelled entries are marked completed for ACP display with an explicit [Cancelled] label. Other unknown statuses retain a visible label and pending display state. Muse supplies no priority here, so entries use neutral medium presentation. Plans are bounded to 500 entries and 8 KiB per entry, with visible omission/truncation notices.

Only public reasoning.summary parts and their deltas become agent_thought_chunk. Completion-only summaries render once in part order; providers with no summaries remain supported. Raw reasoning text and encrypted provider state are never read for this feature.

Tool output deltas update their original call ID before completion. Final snapshots replace the tool content, and absent final text preserves already observed output. Public unknown item kinds display kind/status/fallback text; worker-specific lifecycle/control remains separate. Text is bounded to 64 KiB per surface with explicit truncation notices. Metadata-only media keeps an explicit unavailable explanation. Stored output references can be read through the negotiated output interface on verified Muse 1.2.1; other clients/hosts retain explicit fallback notices. The adapter neither fetches an invented output endpoint nor treats descriptive image text as bytes. Observed file-change reports keep their content and these notices.

The SDK owns live gap recovery. The adapter reads only a current fold and replaces observations after catch-up. Load/resume reconstruct the latest root facts through at most 20 backward public pages of 100 events; missing facts within that budget remain unknown, and obsolete client plans are cleared before new state is shown. Opaque cursors are relayed, never sorted or fabricated. This is read-only and starts no provider request. Repeated facts do not accumulate twice.