feat: make global checklists work with Hermes Agent (and every polyhook caller)

[tupe12334]

Summary

Makes steplock, including global checklists, work the same under Hermes Agent as under Claude Code, and fixes response formatting for every non-Claude caller.

• Answer in the caller's own format (fixes Hook responses are always serialized in legacy Claude Code format — caller context discarded by parsing stdin instead of reading it #196, supersedes fix(cli): preserve caller context so responses use the agent's wire format #197). run_app parsed the stdin bytes directly with parse::parse_event. That skipped polyhook::read_from, which records the detected caller for polyhook::respond, so every response fell back to the legacy Claude Code shape. It now reads through read_from:
  • Hermes gets {"action": "block", "message": ...}.
  • Claude Code gets a PreToolUse permissionDecision: "deny". The old top-level decision: "block" ended the whole Claude session instead of denying one tool call, which made "retry your original command" impossible.
  • Other callers get their own format instead of one they may not parse. fix(cli): preserve caller context so responses use the agent's wire format #197 made the same change on the pre-clap main.rs. This PR is rebased on the current main, so fix(cli): preserve caller context so responses use the agent's wire format #197 can be closed.
• Same global directory under Hermes. In containers and with TERMINAL_HOME_MODE=profile, Hermes gives hooks a per-profile HOME and exports the real one as HERMES_REAL_HOME (hermes_constants.apply_subprocess_home_env). dirs::home_dir() reads HOME first, so steplock would have picked a different "global" directory under Hermes. It now prefers an absolute $HERMES_REAL_HOME. STEPLOCK_GLOBAL_DIR and XDG_CONFIG_HOME still win over it.
• Docs: Hermes Agent setup in Installation.md, plus a README note:
  • pre_tool_call with matcher: "terminal", and on_session_end for cleanup.
  • The consent and hooks_auto_accept requirements.
  • hermes hooks test for verification.
  • Hermes runs hooks in its own start directory, not the terminal tool's workdir. Global checklists don't care; a project .steplock/ is found only when Hermes was started in that project.
• Lint ratchet: clippy::map_flatten is now deny, placed away from fix(ci): make MSRV real — raise to Rust 1.88 and actually test on it #212's line to avoid a merge conflict.

Nothing else needed changing for Hermes. polyhook 0.1.11 already maps pre_tool_call to tool:before, terminal to bash, tool_input, session_id, and on_session_end to session:stop. Existing checklists (on_tool = "bash", input.command...) match Hermes shell commands unchanged.

Test plan

• cargo test passes: 115 lib, 35 bin, 18 cli, 11 integration.
• New tests: Claude Code deny shape; Hermes block shape with ack.sh; Hermes approve is {}; Hermes session_id scopes state; Hermes on_session_end cleans the global session; binary run with a sandboxed HOME plus HERMES_REAL_HOME still finds the shared global checklist; HERMES_REAL_HOME resolution order.
• With the old main.rs, the two response-shape tests fail (33 passed, 2 failed), so they catch the bug.
• cargo clippy --all-targets -- -D warnings, cargo fmt --check, and cargo doc with -D warnings all pass.
• Live through real Hermes v0.21.4 (hermes hooks test, isolated HERMES_HOME, empty project, global dir from steplock init --global):
  • git push: parsed (Hermes wire shape): {"action": "block", "message": "[example-gate: 1/2] ..."}.
  • After ack.sh: step 2/2 blocks. After the second ack.sh: approved ({}).
  • on_session_end: steplock: cleaned up session ...; the global sessions/ directory is empty.
  • Nothing was written to the project directory.
• Not reproduced live: the sandboxed HOME case. On this macOS host Hermes keeps HOME as the real home even with TERMINAL_HOME_MODE=profile (both HOME and HERMES_REAL_HOME point at the user home), so the rewrite only happens in containers or profile-home setups. That path is covered by the unit and binary tests.

🤖 Generated with Claude Code