Skip to content

CI verification: Sonnet 5.5 and thinking-binding controls - #271

Merged
ualtinok merged 4 commits into
mainfrom
ci/sonnet-5-5-verify-20260928
Sep 30, 2026
Merged

ualtinok merged 4 commits into
mainfrom
ci/sonnet-5-5-verify-20260928

Conversation

@antauth-alfonso

@antauth-alfonso antauth-alfonso Bot commented Sep 30, 2026 •

Copy link
Copy Markdown

Purpose

CI-only verification of Sonnet 5.5 support and shared Opus 5.5 thinking-binding controls. This draft does not authorize merge, tagging, package publication, or a live-host restart. v2 remains held for the common-auth migration.

Changes

  • Register claude-sonnet-5-5 across Core, OpenCode and Pi with official pricing, 1M context and 128K output.
  • Normalize adaptive summarized thinking; OpenCode maps explicit opt-outs to bare between_tools at supported effort and removes unsupported forced tool choice while preserving structured-output validation.
  • Extend configured thinking-prefix controls to adaptive Sonnet 5.5 and Opus 5.5 replay. Account-default behavior is unchanged.
  • Advertise Sonnet 5.5's supported Pi effort levels and reject unsupported minimal locally.

Verification completed

  • Synthetic live OAuth checks accepted adaptive and between_tools Sonnet 5.5 requests using the existing Claude Code 2.1.280 persona.
  • A signed-thinking continuation after a synthetic prefix edit returned 200 with explicit drop_block; Anthropic reported thinking_dropped / prefix_binding_mismatch.
  • Core: 309 passing tests. Pi: 149 passing tests. Focused host/request suites and real Pi tool round-trip passed.
  • Workspace typechecks, lint, formatting, emitted Pi import guard, and packed CLI/TUI smoke passed.
  • Real isolated OpenCode process coverage verifies structured-output tool success and plain-text refusal for Opus 5.5 and Sonnet 5.5.

Checks still required

The full local OpenCode unit gate was red: 1,702 passed, one unchanged lock stress test exceeded its 30-second deadline, including when run alone. Full local process E2E was 43/49: five project-readiness timeouts and one fallback prompt timeout. Host CPU/swap measurements showed heavy contention, but those runs are not counted as passing and deadlines were not raised. Clean CI on this exact candidate must resolve the remaining verification gate.

Independent audit identified the Pi effort issue corrected in the follow-up commit. Its evidence package lacked the requested official-document excerpts and parent diff, so it is not blanket release clearance.

Official references:


View with [code]smith Autofix with [code]smith
Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.


Summary by cubic

Adds Claude Sonnet 5.5 support across Core, OpenCode, and Pi, and extends the configured thinking-binding prefix controls to adaptive Sonnet 5.5 and Opus 5.5 continuations. This branch is for CI verification only and is not authorized for merge, tagging, package publication, or a live-host restart; v2 remains held for the common-auth migration.

  • Registers claude-sonnet-5-5 with official pricing, 1M context, and 128K output.
  • OpenCode normalizes adaptive Sonnet 5.5 thinking, maps explicit opt-outs to bare between_tools at supported effort, and removes unsupported forced tool choice while keeping structured-output validation.
  • Pi exposes Sonnet 5.5's supported low–max effort levels and rejects unsupported minimal locally; docs clarify that between_tools requires OpenCode's raw-body path.
  • Thinking-prefix controls (error/drop_block) now apply to replayed Sonnet 5.5 and Opus 5.5 adaptive thinking; the control is never sent with between_tools, which rejects the field. Account-default behavior is unchanged.
  • Test floors ratcheted to clean-runner counts (Core 309, OpenCode 1703, Pi 149).

Written for commit 77e4c90. Summary will update on new commits.

Review in cubic

@ualtinok ualtinok closed this Sep 30, 2026
@ualtinok
ualtinok merged commit 77e4c90 into main Sep 30, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant