Skip to content

v2026.7.28

Choose a tag to compare

@github-actions github-actions released this 28 Jul 14:27

New Features

  • Claude Agent SDK provider — Sign in through native OAuth, import existing Claude credentials, pin accounts, fail over across multiple accounts with session affinity, and use in-process MCP/custom tools through the built-in provider (#409).
  • Self-teaching interactive UI — Discover commands, keybindings, favorites, settings, and tool workflows through startup tips, working tips, /help, /keybindings, the shortcut overlay, and expanded feature coverage (#405).
  • Cache-aware long-running tools — Foreground bash calls respect prompt-cache budgets, auto-detach at the safe deadline, and direct users to background sessions plus monitor subscriptions rather than fixed waits (#422).
  • Durable goal execution — Goal completion is gated on finished todos, blocked/completed audits are decisive, todo continuity survives follow-up turns, and monitor-aware continuation avoids needless wakeups (#427, #428, #430).
  • Billing-aware model failover — Billing and quota failures can permanently pin the configured fallback as the active session model instead of repeatedly returning to an exhausted provider (#431, #434).

Breaking Changes

Added

  • Add a built-in Claude Agent SDK provider with OAuth login/import, account listing and pinning, HRW session affinity, automatic account failover, custom-tool mapping, and in-process MCP execution (#409).
  • Add rotating startup and working tips, /help, /keybindings, a keyboard-shortcut overlay, external-editor support, favorite-model guidance, and settings to control tip presentation (#405).
  • Add cache-aware foreground bash timeout budgets and auto-detach behavior with completion notifications (#422).
  • Discover project configuration from the nearest parent directory instead of requiring the working directory to be the project root (#423).
  • Add a cancellable session_before_reload extension event so extensions can veto reload before session state is replaced (#418).
  • Show active monitor count in the interactive footer and show a pirate (🏴‍☠️ OmO Native) badge when the native omo-senpi plus senpi-task stack is installed (#417, #438).
  • Add OpenAI -fast priority-tier model variants to the default catalog (#420).
  • Add a Kimi K3 ambiguity preset that reflects before asking a focused clarification question (#390).
  • Render Kitty graphics through tmux passthrough with split-safe placement (#389 by @minpeter).
  • Add tool_schema() and self-correcting eval tool-schema feedback (#407).

Changed

  • Permanently switch the session model after configured billing-class fallback, including unconditional pinning for billing errors (#431, #434).
  • Keep emergency compaction pruning engaged with hysteresis so repeated threshold crossings do not destroy the prompt cache (#425).
  • Bound compaction and summarization stream acquisition with wall-clock budgets, then degrade eligible transient failures without issuing a duplicate request (#419, #436).
  • Support OpenAI Responses remote compaction v2 and provider-native compaction capability routing (#403).
  • Smooth streamed response pacing while keeping the final response immediate (#396 by @changeroa).
  • Pin footer anchors, prioritize the provider prefix, restore immediate provider counts, and remove the cache-total segment so narrow layouts remain useful (#406).
  • Lazily activate searchable MCP/eval tools only for the surface that requested them (#408).
  • Route long waits to background sessions and monitor subscriptions, removing stale tmux and bash_output wait guidance (#413, #435).
  • Hide internal terminal wake notifications while preserving the event that resumes the agent loop (#429).

Fixed

  • Do not warn when optional retry.fallbackChains is absent; explicit malformed values still produce configuration warnings.
  • Resolve provider-qualified fallback selectors within the selected provider rather than treating them as global model globs.
  • Retry pre-output provider stream failures and Cloudflare 522 responses through the bounded retry/fallback policy (#404, #421).
  • Omit unsupported Codex reasoning-summary payloads and preserve session affinity through every compaction and summarization path (#412 by @DevNewbie1826, #416).
  • Avoid unsupported minimal reasoning during compaction (#398).
  • Degrade transient blocking-compaction failures cleanly instead of aborting the active session (#391).
  • Keep the startup tip visible when extensions install a custom header (#410).
  • Initialize themes before starting the multi-session RPC host (#414).
  • Stop config reload loops caused by the global default extension shim, watch only loadable extension entries, and ignore unrelated runtime-state subtrees (#397, #400).
  • Prevent zombie processes from being counted as live during process-tree teardown verification (#402).
  • Keep monitor tools available with native bash and flatten the monitor schema so Anthropic payloads retain every field (#393, #401).
  • Retry transient session-title generation failures and sanitize runtime-error presentation (#394).
  • Stop reload from auto-starting a stopped agent and block an active goal when the user aborts outside an agent run (#395).
  • Preserve todo continuity, enforce open-todo completion gates, and clean up monitor-aware goal continuation after reload (#427, #428, #430).
  • Preserve paste-marker expansion across editor transfers and prevent repeated render loops in large sessions (#411 by @minpeter, #424 by @sigridjineth).

Removed