Repository navigation
2026.10.7: write each model's specs into the clients that list models, core v0.64.0 - #289
Merged
Merged
Conversation
Groundwork for writing context window, max output, reasoning and image input into the clients that list models in their own config. Each client module still writes only the ID; staleness now compares what a client can hold (Client::as_written). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Specs dialog gains two segmented choices next to context window and max output: Price table / Yes / No for reasoning and image input. The line under each says what the price table has (or that it has nothing). Saving with all four left to the price table removes the manual spec. The model row's "Manual specs" mark and its tooltip cover the two new specs. Generated types hand-edited to control API v40 (ModelRow reasoning / image_input with their sources, ModelSpecSave reasoning / image_input); to be regenerated from core. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Pi and omp get contextWindow, maxTokens, reasoning and input per model; opencode gets limit, reasoning, attachment and modalities in the v1 shape, and limit and capabilities.input in the native v2 shape. Only what the gateway reports is written, and every field is read back so a changed spec makes the client page offer an update. The YAML write-back check now reads nested numbers and booleans as strings, as the YAML reader does, so omp's model list verifies. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Qwen Code gets generationConfig.contextWindowSize and modalities.image. Grok Build gets context_window and supports_reasoning_effort. Unknown fields are left out so each client's own catalog and defaults apply. Max output is written to neither: Qwen's samplingParams.max_tokens and Grok's max_completion_tokens are sent verbatim on every request, not capability ceilings. Qwen has no reasoning field that does not change every request (capabilities.reasoning needs an effort menu and sends its default; reasoning: false turns /effort off), so reasoning is left out there; /effort already offers every tier for undeclared models. Grok has no image field. as_written drops what each client does not hold, and models_in reads the written specs back so a changed spec shows as stale. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…easoning and image input Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Taking over opencode, Pi, oh-my-pi, Qwen Code or Grok Build now writes each model's specs next to its name: the context window, and where the client uses them, the output limit, reasoning and image input. Before, only model IDs were written, so these clients ran on their defaults. In Pi, for example, every model got a 128K context and no reasoning levels (#288).
/v1/models(core v0.64.0:context_window,max_output_tokens,supports_reasoning,input_modalities). Whatever the gateway doesn't know is left out, so the client falls back to its own default.contextWindow,maxTokens,reasoning,input.limit,reasoning,modalities. Native v2:limit,capabilities.input.contextWindowSize,modalities. No output limit and no reasoning, because both change every request.context_window,supports_reasoning_effort. No output limit, because it is sent verbatim.Client::as_written), so a changed spec asks for an update.config.model_spec_nothing_set.release-notes/2026.10.7.md.Checked locally against the new core:
cargo fmt,clippy -D warnings,cargo test(including the control-plane handshake),pnpm typecheck,pnpm test(872) andpnpm build.🤖 Generated with Claude Code