Skip to content

v2026.7.16

Choose a tag to compare

View all tags
@github-actions github-actions released this 16 Jul 18:09
v2026.7.16
bab8046

Breaking Changes

  • Changed AuthStorage.list() from a synchronous provider-ID list to asynchronous non-secret credential metadata matching the pi-ai CredentialStore contract.

Added

  • Added inherited Kimi K3 model availability and deferred tool loading for compatible providers.
  • Added ModelRuntime as the canonical async SDK and internal model/auth facade while retaining the fork's AuthStorage, ModelRegistry, and corresponding CreateAgentSessionOptions compatibility APIs.

Changed

  • Tuned the gpt-5.6 system prompt preset against the oh-my-opencode Hephaestus GPT-5.6 prompts: verification wording now names validators senpi can actually run (type check/lint instead of a nonexistent diagnostics tool), parallel tool calls are the stated default with serial as the exception and no ;/&& chaining of unrelated shell steps, todo items are named by deliverable and reconciled at turn end, and bracketed 【F:...】-style citations are banned from output.
  • Reframed the skill-loading trigger in every system prompt to load skills on loose description match, stating the cost asymmetry (an irrelevant load costs little; a missed relevant skill degrades the work).
  • Added an instruction-file precedence rule to the dynamic system prompt's Project Context section: instruction files bind files under their directory, deeper files win on conflict, and explicit user instructions override.

Fixed

  • Fixed automatic compaction to preserve user, OMO steer, and goal follow-up messages appended while compacted context is rebuilt.
  • Fixed SDK-created sessions to inherit the configured agent stream idle timeout.
  • Fixed inherited Cloudflare endpoint placeholder resolution and lazy provider final-message forwarding.
  • Fixed cloning or forking a session before its first assistant response to explain that the session must be saved first.

Removed

Assets 11
Loading