Repository navigation
fix: restore native Opus limits and thinking continuity - #92
Merged
Merged
Conversation
pavelhov
changed the title
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Opus 5.5 now receives its native output/context capacity through Codex and retains thinking state across tool calls. A post-0.1.24 failure was conclusively an ordinary turn: the omitted output limit still selected the legacy 40,192-token total, while the earlier fix applied only to compaction.
Ordinary native Opus 5.5 requests now default to 128,000 total output tokens. The catalog advertises the documented 1M context, with a 900K default compaction threshold, including offline discovery and supported native URL forms. Empty signed/redacted thinking and separate signed blocks survive JSON/SSE tool continuation; signature fragments are retained within the translator budget. Explicit limits, requested ordinary effort, smaller live context limits, custom endpoints, and the 64K/medium compaction policy keep their semantics.
Existing Codex processes must reload the refreshed catalog to use the larger context window. This restores proxy-controlled model capabilities; it does not replace Codex’s agent runtime with Claude Code’s.
Validation:
bun run test:parallelpassed all 633 files. Typecheck and privacy checks passed. GUI lint, i18n checks, 888 tests and build passed; the docs build produced 44 pages. Source:38b11ccf9. Independent review findings were repaired and rechecked, including native Models URL aliases and terminal-guard thinking continuation. Synthetic regressions cover JSON/SSE, incomplete responses at the former 40,192 boundary, explicit caps, context precedence and signature lifecycle. No paid inference or conversation replay was used.Security review: native metadata matching does not decide credential or routing destinations; existing authentication and transport ownership remain unchanged. Signature collectors enforce retained/transient byte accounting and cleanup. Request bodies, credentials and account identifiers are not logged. Dependency, workflow and release-distribution code is unchanged. Release promotion requires the existing Ed25519 anchor, authenticated historical inventory, fresh draft/public download verification, and six isolated installed-app Sparkle scenarios; no production installation or restart is performed by this task.