v2026.7.30-2

New Features

• Safer main-model selection — Warn interactively when a main-session provider, model id, or display name selects
  Minimax or Qwen, including startup, /model, exact references, favorite rotation, and post-auth defaults. The
  prominent Korean/CJK warning remains scoped to the main session; subagent routing and safe model families are
  unchanged (#530).
• High-reasoning risk signal — Emit high_reasoning_warning through the session API, RPC JSONL, and interactive
  red warning box when GPT-5.x -sol models run at xhigh or max. Detection is intentionally narrower than
  capability support, so Claude, DeepSeek, ordinary GPT, and unrelated solar ids do not inherit the warning.
  Events are deduplicated by provider, model, and level
  (#517,
  #526).
• Broader OMO workflow tips — Add discoverable tips for init-deep, debugging, refactor,
  remove-ai-slops, and visual QA alongside the existing planning, ultrawork, research, and review guidance. The
  additions keep the existing task-extension availability gate
  (#521).

Breaking Changes

Added

• Add packages/coding-agent/docs/release-guide.md, the authoritative config-neutral build, plugin verification,
  local installation, Jobdori installation, troubleshooting, rollback, cleanup, and maintenance guide for
  2026.7.30-2.

Changed

• Compact proactively at agent idle when the persisted context is over threshold, rather than waiting for the next
  user prompt to discover the same required compaction. The idle path reuses existing admission, lifecycle,
  cancellation, and stale-revision safeguards; below-threshold and already-running sessions remain unchanged
  (#522).
• Update the bundled MCP runtime dependency from @modelcontextprotocol/sdk 1.29.0 to 1.30.0. Builtin registration
  order, the mcp-last invariant, user MCP configuration format, and the separate hooks plugin-manifest loader are
  unchanged.
• Override the bundled brace-expansion runtime from 5.0.7 to patched 5.0.8, closing
  GHSA-mh99-v99m-4gvg, an unbounded expansion-length
  denial-of-service reachable transitively through glob and minimatch. This is dependency-only: glob matching,
  extension discovery, configuration formats, and command behavior are unchanged.

Fixed

• Compare Senpi CalVer releases by the repository's YYYY.M.D-N contract instead of npm prerelease ordering when
  deciding whether an update is newer. The unsuffixed date is revision 1, so 2026.7.30-2 now correctly sorts after
  2026.7.30; clients on a same-day revision no longer accept the registry's older bare-date release as an update
  and downgrade themselves. Cross-day CalVer ordering and ordinary semver package versions retain their prior
  behavior.
• Continue a goal immediately when create_goal is the final tool action of the user’s turn. Freshly created goals
  are now marked goal-driven before the turn ends, so the first hidden continuation starts at once instead of being
  mistaken for a side question and waiting for the 60-second user-grace timer. User grace still applies to real
  messages sent while a goal was already active, and monitor delays, continuation caps, stale/repetition guards,
  single-flight delivery, and completion handling are unchanged.
• Reset the goal continuation progress cap after an observable tool action. Long-running goals can therefore continue
  after real progress instead of exhausting a lifetime counter and stopping despite new work, while no-progress
  repetition detection, stale-turn fencing, monitor waits, and single-flight continuation delivery remain bounded
  (#534).
• Restore user input that was queued while pre-prompt compaction failed, was rejected, or was aborted back into the
  interactive editor instead of allowing terminal recovery to consume or strand it. Failure metadata is scoped to
  the matching compaction attempt, terminal pre-prompt recovery is marked complete exactly once, and successful
  compaction keeps the existing queued-input admission path
  (#535 by
  @realsigridjin).
• Recover Kimi XTML response channels through the shared model-runtime boundary even when no tools are registered,
  and retry one terminal empty Kimi response once. Structural markers no longer leak into final output; successful
  recovery appears exactly once, and repeated emptiness fails visibly and finitely
  (#523,
  #537).
• Reuse the active turn’s runtime API key for automatic session-title generation before resolving provider headers
  and compatibility options. A CLI --api-key turn and its background title request therefore use the same
  credential instead of allowing the title request to fall back to a configured key and fail with 401
  (#519).
• Serialize apply-patch file mutations and disclose concrete parse, seek, file-operation, and partial-success failure
  reasons. Multi-file patches no longer race concurrent writes, and callers receive the successful source-backed
  patch details together with an actionable failure instead of a generic unsuccessful result
  (#524).

Removed