Fixes#695
`where claude` on Windows returns both the npm-generated POSIX shell
wrapper (extensionless) and the Windows batch shim (`claude.cmd`)
side-by-side. The existing `path_from_successful_output` helpers in
`claude_cli.rs` and `codex_cli.rs` picked the first existing line,
which gave the POSIX wrapper - Windows tried to load it as a PE32,
`CreateProcessW` returned ERROR_BAD_EXE_FORMAT (193), and the AI chat
panel surfaced "%1 is not a valid Win32 application."
Mirror the fix that landed for Gemini in `gemini_discovery.rs`: when
`cfg!(windows)`, filter candidates through the existing shared utility
`cli_agent_runtime::has_windows_cli_extension` so .bat/.cmd/.com/.exe
win over extensionless POSIX scripts. Non-Windows behavior is unchanged.
Added a regression test in each file matching the
`windows_path_lookup_prefers_cmd_shim_over_extensionless_npm_script`
pattern from gemini_discovery.rs.
Tolaria's MCP server is pure ESM with only standard `node:fs|path|http|child_process`
imports and pure-JS dependencies, so Bun can execute it identically to Node.
Until now `find_node()` was the single entry point for spawning the WebSocket
bridge and writing external AI tool config — users with Bun but no Node would
hit "node not found in PATH or common install locations" and lose access to MCP
tools entirely.
Introduce `find_mcp_runtime()` which returns the first verifying runtime,
preferring Node 18+ when present and falling back to Bun 1+. The generic
PATH and login-shell lookup helpers (`find_on_path`, `find_in_user_shell`,
`lookup_command`, `lookup_paths`) are now parameterised by command name so
both runtimes share the same machinery. Bun candidates cover `~/.bun/bin/bun`,
mise/asdf/proto shims, Homebrew, and `%USERPROFILE%\.bun\bin\bun.exe` on
Windows. The Codex CLI and Windows .cmd-shim resolution paths keep using the
strict `find_node()` since they specifically need Node.
`spawn_ws_bridge_with_paths`, `mcp_config_snippet`, and `register_mcp` now
resolve through `find_mcp_runtime` so the runtime that gets registered into
`~/.claude.json`, `~/.gemini/settings.json`, `~/.cursor/mcp.json`, etc.
matches the one actually present on the machine.
Locale copy updated from "nodeRequirement" to "runtimeRequirement" across
all 15 catalogs to reflect "Node.js 18+ or Bun 1+".
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Port the useful dynamic-vault MCP direction from Domenico Lupinetti's proposal into the current mounted-workspace model.
Durable external MCP registration is now vault-neutral, Node MCP resolves active mounted vaults from explicit env or Tolaria vaults.json, and vault context includes root AGENTS.md instructions per active vault.
Based-on: https://github.com/refactoringhq/tolaria/pull/603
Co-authored-by: Domenico Lupinetti <domenico@translated.net>
Move common shell lookup logic (PATH check, login shell probe,
first_existing_path) into cli_agent_runtime as find_cli_binary()
and check_cli_availability(). Rewrite kiro_discovery to use these
shared utilities, reducing duplication from 100 lines to ~20.
Other discovery modules (pi, gemini, opencode) can adopt the same
pattern in a follow-up.
- Add kiro_discovery.rs: binary detection (PATH, login shell, known paths)
- Add kiro_cli.rs: streaming via stdin, ANSI stripping, MCP config merge
- Register Kiro in ai_agents.rs enum, status, and dispatcher
- Add 'kiro' to settings normalization
- Add Kiro to frontend agent definitions and helpers
- Update test for agent cycling
Kiro uses plain text streaming (not JSON) so it has its own stream
handler rather than using cli_agent_runtime::run_ai_agent_json_stream.
MCP config merges into existing .kiro/settings/mcp.json to preserve
user-configured servers. Prompt is passed via stdin to avoid E2BIG.
