(0.5.7, passthrough, models, ux)

[evalstate]

Release Notes (Draft)

Added

• Tool-result passthrough flow so tool outputs can be returned directly without an extra post-tool LLM
  synthesis step.
• Local filesystem runtime for shell-enabled agents with ACP-compatible read_text_file/write_text_file,
  plus local apply_patch support.
• Native apply_patch tool definition (grammar-backed for Responses custom tools), preview rendering, and
  improved apply-patch streaming/output formatting.
• New model runtime control: /model fast (service tier), including ACP slash-command support.
• New prompt shortcuts and controls:
  • Shift+Tab service-tier cycle
  • F6 reasoning cycle
  • F7 verbosity cycle
  • F8 web search toggle
  • F9 web fetch toggle
• New “quiet hash-send” mode via ##agent message (delegated send without normal chat/tool transcript
  rendering).

Changed

• OpenAI Responses integration now supports:
  • custom tool calls (custom_tool_call) including apply_patch payloads
  • assistant message phase propagation (commentary / final_answer)
  • service tier wiring (fast/flex, with model/provider validation)
• Shell/file tooling behavior is model-aware:
  • configurable write_text_file_mode (auto|on|off|apply_patch)
  • auto prefers apply_patch for GPT-5/Codex-style models
  • optional enable_read_text_file control
• Interactive UI toolbar/model chip was upgraded with service-tier, reasoning, verbosity, web-search, and
  web-fetch indicators.
• Model catalog/default aliases were refreshed (including new GPT-5.4 and updated codex defaults), and
  model picker now supports generic/local custom model entry flow.
• Skill prompt generation now uses explicit read-tool naming (read_text_file vs read_skill) based on
  runtime capability.

Fixed / Improved

• Improved cancellation/interrupt handling across interactive, CLI runner, ACP prompt handling, and tool-
  loop history reconciliation.
• Better post-cancel recovery: stale progress row cleanup, clearer cancelled-turn messaging, and safer
  task cancellation state handling.
• Improved read_text_file and shell result display behavior (cleaner summaries, truncation behavior,
  display suppression integration).
• Refined MCP connection/reconnect handling and stream transport tracking hooks.
• Result export reliability improvements for one-shot message runs (including transient result history
  capture when needed).

Internal / Refactor / Tests

• Shared formatting and marketplace-argument parsing helpers for command handlers.
• Unified A3 header rendering utility used across CLI/command surfaces.
• Docs layout switched to a docs git submodule pointer.
• Broad test expansion for passthrough flow, filesystem runtime, apply_patch display/preview, quiet hash-
  send, display suppression, model controls, and ACP response-mode handling.