v2026.7.3

Added

Changed

• Tightened the default dynamic system prompt: merged overlapping intent-gate rules, collapsed the redundant execution-stance bullets, removed the no-op "no trigger tools" line and decorative freedom rhetoric, and de-duplicated directives already covered by Policies. Same behavioral contract at about a third fewer tokens in the shared sections, across every model preset and the fallback prompt.
• Rewrote the Claude Opus 4.5–4.8 system prompt preset tuning against Anthropic's Opus prompting guidance: removed lines restating native model behavior, extended tools-over-reasoning guidance to Opus 4.7, compensated literal instruction following with evident-intent scoping, overrode the default frontend house style on 4.7/4.8, reduced post-user-turn re-reasoning on 4.8, and told all Opus presets not to wrap up early since senpi auto-compacts context.
• Rewrote the GPT-5.5 system prompt preset as a full outcome-first core (per the GPT-5.5 prompting guide) via a new corePrompt override on the dynamic prompt builder; behavior contracts (routing line, todo discipline, verification tiers, hard limits, file-operations routing) are preserved at roughly half the static prompt tokens. Other model presets are unchanged.

Fixed

• Fixed inherited startup model selection to skip unauthenticated saved defaults so configured local custom models can be selected instead.
• Fixed inherited Escape aborts to clear runs stuck in extension context hooks that ignore abort signals.
• Fixed the inherited question extension example to run question tool calls sequentially so multiple questions in one assistant turn remain answerable.