[Rust] v0.2.7

Package: crates.io agent-commander@0.2.7

Added

• Added enforceable read-only planning mode for supported tools and clear rejection for unsupported tools.

Fixed

• Removed JavaScript command execution's dependency on runtime CDN imports, avoiding Deno CI failures when esm.sh is unavailable.

Changed

• Synced model maps from hive-mind v1.57.2 (issue #22): Claude opus -> claude-opus-4-7; Codex default gpt-5 -> gpt-5.5 and added the gpt-5.5/5.4/5.3/5.2/5.1 family; Agent default minimax-m2.5-free -> nemotron-3-super-free with qwen3.6-plus-free retained as deprecated.

Added

• Added Rust package README metadata and language-specific GitHub Release tags.

Fixed

• Fixed Rust CLI pass-through for Claude fallback, session, prompt append, verbose, and replay options.

Fixed

• Pipe prompts for stdin-based tools through temporary prompt files during execution so large generated prompts are not embedded in nested shell commands.

Added

• Added --prompt-file / prompt_file support for callers that already have prompt content on disk.

Added

• Added native execution passthrough controls for Claude, Codex, OpenCode, and @link-assistant/agent, including custom executables, raw tool environment variables, raw tool arguments, and suppressible default safety bypass flags.

Added

• Expose normalized result metadata on AgentResult, including success classification, session IDs, usage-limit details, summaries, cost estimates, stream token usage, sub-agent calls, and execution error information for claude, codex, opencode, and agent.
• Add AgentResult::usage for Rust parity with the JavaScript result shape.

Fixed

• Fixed Rust release recovery when a GitHub release tag exists without a matching crates.io publication, and made Rust GitHub releases explicitly become the repository latest release.

Fixed

• Added Qwen and Gemini prompt-file stdin handling plus native executable, environment, and raw argument passthrough parity.

Added

• Map --read-only and --plan-only for the agent tool to its native --permission-mode (agent v0.24.0): --read-only → readonly and --plan-only → plan. The agent tool now supports enforceable read-only/planning mode instead of being rejected, and exposes typed permission_mode and permission (OpenCode-compatible JSON policy) passthrough options.

Added

• Add a uniform per-command approval relay ("ask" mode). The new --approve-each flag (alias --permission-mode ask, AgentOptions.approve_each) maps to each backend's native per-command approval mechanism and relays native permission prompts as normalized permission_request events (carrying an opaque id, tool, command/pattern, title, and a scope), forwarding the consumer's normalized decision (once | always | reject) back in the native wire format. Supported (relayable) for agent (--permission-mode ask + --input-format stream-json, scope session) and claude (stream-json can_use_tool, scope tool-input); rejected with a clear error for codex, qwen, gemini, and opencode. Adds a permissions module (normalize_permission_request, build_permission_response, PermissionRelay, permission_parity, supports_ask, ASK_DECISIONS, ask_scope).

Added

• Add real PTY-backed interactive agent capture with input, control-key, resize, unrolled transcripts, semantic events, and replay artifacts for Claude, Codex, OpenCode, Agent, Gemini, and Qwen.