* Improve compaction token budgeting
Use MessageWithMetadata.metrics input/output token counts when estimating message size for compaction, falling back to the existing chars/3 heuristic only when metrics are unavailable. This makes trigger decisions and post-compaction accounting use provider-reported token usage instead of relying only on serialized character estimates.
When an explicit compaction maxInputTokens budget is configured and the model exposes maxTokens, reserve half of the model output budget before triggering compaction. This gives the next provider request room for completion tokens and reduces edge cases where the local context estimate passes but the provider rejects the prompt as exceeding its limit.
Keep explicit reserveTokens and thresholdRatio overrides intact, and add regression coverage for metric-based token estimation, fallback estimation, and output-token-aware trigger budgeting.
* new target tokens and trigger tokens value
* fixes
* use imports and add unit test
* resolveMaxInputTokens
---------
Co-authored-by: Saoud Rizwan <7799382+saoudrizwan@users.noreply.github.com>
The standalone AgentRuntime({ providerId, modelId }) constructor built the gateway model in resolveRuntimeConfig and returned { ...rest, model } without deriving messageModelInfo. The prebuilt-model path preserves it, and core sessions populate it via buildMessageModelInfo, but standalone SDK callers lost it -- so assistant-message modelInfo and model-tagged telemetry (reasoning tokens, action-follow-through) emitted without provider/model dimensions.
Derive messageModelInfo as { id: modelId, provider: providerId } in that path (family omitted; it is optional and unavailable here). An explicit caller-provided messageModelInfo still wins. Adds provider-form tests covering both the derived and explicit-override cases.
The bun run version catalog regen dropped the xiaomi mimo-v2-omni,
mimo-v2-pro, and mimo-v2-flash models from the live data. mimo-v2-omni
is the xiaomi provider's defaultModelId in builtins.ts, so shipping the
refreshed catalog would point the default at a missing model and broke
the provider-ids test. Revert the catalog to the pre-release state and
ship v0.0.53 as a pure SDK code release; the catalog will refresh in a
later release once upstream data is stable.
* refactor(vscode): use string type for api provider in proto
Replace the ApiProvider proto enum with plain string fields across\nmodels.proto and state.proto, and drop the enum<->string conversion\nmappings in api-configuration-conversion.ts. Updates ApiOptions and\nOpenAICompatible settings components accordingly.
* fix custom provider render
* format
* id
* remove extension providers file
* uses includes
* fix(ci): harden ext-vscode stable release workflow
- Resolve previous tag to the latest vX.Y.Z ancestor instead of the most
recent reachable tag. The nightly workflow now pushes a nightly-main-*
annotated tag on every main commit, so git describe was resolving the
release notes' Full Changelog compare link to a nightly tag rather than
the prior release tag.
- Extract the changelog section by exact version heading (and fail if it
is missing) instead of always taking the first ## [ block, so a stale
top entry can no longer ship as the release notes for a different
version.
- Add a pre-publish Verify Changelog Entry gate so a release cannot be
published unless CHANGELOG.md leads with the version being released.
- Add a Verify Marketplace Tokens gate so a missing VSCE_PAT/OVSX_PAT
fails fast before packaging rather than mid-publish.
- In existing-tag mode, require the tag to point at the tested SHA so the
published artifact always matches what CI verified.
- Add a concurrency group keyed on the tag to prevent duplicate
concurrent publishes of the same release.
* fix(ci): validate stable release metadata before publish
The --thinking description added in #11656 is long enough that at 120
columns commander wraps it, splitting "omitted leaves provider default"
across two lines. The TUI e2e assertion uses a contiguous getByText, so it
failed on the ubuntu-only TUI test leg, blocking the SDK publish gate.
Widen the help terminal to 200 columns so long descriptions render on a
single line.
The Skills note pointed users to "Settings → Features → Enable Skills,"
but that toggle no longer exists — the Features settings section has no
Skills entry and skills are loaded by default. Point users to the actual
Skills menu (scale icon → Skills tab), consistent with the access path
already documented later in the same page.
Fixes#11740
Co-authored-by: Minhkunn <minh.12072k6@gmail.com>
