omind 10.3.1

omind 10.3.1 — Grok recalls count as consults, and the commit freshness gate now runs in every harness, not just Claude.

10.3.1

Fixed

• Grok's Claude-hook fast path counts a use_tool recall as a consult
  (#472). omi-guard.sh
  reads .tool_name only, and v10.3.0's Grok translation runs in the Python
  adapter. On the bash fast path a recall arrives as use_tool with
  omi__recall-note nested inside, so it is not a consult, and the gate
  denies the recall that would have cleared it. The script now accepts
  camelCase, unwraps use_tool / CallMcpTool, and maps Grok's built-in
  names (run_terminal_command, read_file, …) onto the names it already
  classifies. A snake_case-only use_tool or built-in name on
  --harness claude is translated the same way, so PostToolUse records the
  consult. Listing tools still do not clear the gate, and vault writes are
  still not consults. A plain Claude event is unchanged. A dispatcher's
  arguments sent as a JSON string is parsed, an empty first payload no
  longer hides the real arguments, and a dispatcher with no inner tool name or
  an event with no tool is left alone instead of erroring jq. The Python
  translator applies the same rules.
• The commit freshness gate runs in every harness, not just Claude
  (#473). Its checks matched the tool name Bash only. Antigravity
  (run_command), Gemini (run_shell_command), Hermes (terminal), OpenCode
  (bash) and Poolside (shell) send their own names, so a commit with no
  fetch went through, and the repo-work and global-config checks were skipped
  too. The adapter now maps each harness's shell tool onto Bash
  (guard.SHELL_TOOL_NAMES), and guard.decide does the same for a caller
  that builds the action itself. The command is read from command,
  CommandLine or cmd, as a string or an argv list (Codex local_shell),
  by both the PreToolUse path and the PostToolUse retraction, so a failed
  fetch under any of these names is retracted.
• The no-jq, no-core last resort in omi-guard.sh fails closed for every
  shell tool name and for any use_tool / CallMcpTool dispatcher. It
  matched Bash and run_terminal_command only. A test pins its list to
  guard.SHELL_TOOL_NAMES.
• A dispatcher whose first name field is empty is still unwrapped. jq's
  // kept "", so {"tool_name":"","name":"run_terminal_command",…} stayed
  use_tool and its command never reached the Bash path. jq now takes the
  first non-empty name, as the Python translator does.
• Antigravity resolves the repo from run_command's own Cwd. It wins
  over a top-level cwd and the workspace root, so a git fetch or
  git commit run in a repo outside the root is judged against that repo. A
  JSON-quoted Cwd is decoded, and a relative one is taken from the
  workspace root.

Tests

• A fetch, then a commit with a multi-line -m trailer block, run from a
  different repo's cwd, is fresh in every harness that runs the guard
  (Claude, Codex, Grok, Antigravity, Gemini, Hermes, OpenCode, Poolside;
  MiniMax models run inside these). omind 10.0.2 split that commit at the
  newline inside the quotes, lost its git -C path, and judged it against
  the cwd's repo, so a commit right after a clean fetch was blocked. Later
  releases fixed that as a side effect; these tests pin it.