fix(cli): disarm continuous voice on hotkey during STT/agent (#67545)

[Enough1122]

Fixes #67545.

Narrow fix for the two specific non-recording paths the bug report calls out: Ctrl+B during STT (_voice_processing=True) or agent-busy (_agent_running=True) was a silent no-op — continuous stayed armed and the loop auto-restarted after the turn. The only workaround was /voice off.

This PR does not introduce a full hotkey-toggle semantic. It only inserts a _voice_continuous = False disarm in the two existing early-return paths the issue identifies, leaving the other non-recording paths unchanged. See "Scope vs. #67550 / #67573" below for the deliberate boundary.

What changed

cli.py::handle_voice_record, else branch, before the existing guards:

if cli_ref._agent_running:
    with cli_ref._voice_lock:
        cli_ref._voice_continuous = False
    return
...
if cli_ref._voice_processing:
    with cli_ref._voice_lock:
        cli_ref._voice_continuous = False
    return

Existing behavior on the recording branch (if cli_ref._voice_recording) is unchanged — that path already disarms continuous today and continues to do so.

Why narrow on purpose

The issue reporter asked specifically for the two paths they hit:

• Pressing Ctrl+B while the agent is mid-turn and not yet back at the prompt → expect continuous to drop.
• Pressing Ctrl+B while the previous take is still being transcribed → expect continuous to drop.

A broader "Ctrl+B is a global toggle" interpretation (#67550, #67573) changes semantics on paths the issue does not call out — e.g. TTS playback, interactive clarify/sudo/approval/slash_confirm prompts — and risks surprising users who currently rely on those paths being inert. That's a policy decision the maintainers should make, not one I want to bake in via the narrowest possible fix.

Scope vs. #67550 / #67573

PR	Approach	Covers STT?	Covers agent?	Covers TTS / interactive prompts?
#67569 (this)	Per-guard disarm in the two specific paths the issue names	✓	✓	✗ (intentionally)
#67550	Global disarm at the top of the else branch (toggle semantic)	✓	✓	✓ (proposed)
#67573	Global disarm at the top of the else branch (toggle semantic)	✓	✓	✓ (proposed)

Happy to close in favor of #67550 or #67573 if maintainers prefer the broader toggle. The diff in this PR is a strict subset of either. If they prefer the narrow fix, the diff here is the smallest correct change.

Test plan

• tests/hermes_cli/test_voice_wrapper.py — 44/44 pass locally.
• The fix is bounded to two early-return paths in handle_voice_record. The disarm is guarded by cli_ref._voice_lock to match the existing lock discipline around _voice_continuous writes elsewhere in the same handler.

Branch: fix/67545-voice-continuous-hotkey-disarm @ 7976bd96.

[teknium1]

Thanks for the focused CLI fix. The current-main premise is confirmed: cli.py:13697 and cli.py:13704 return from the non-recording hotkey path without clearing _voice_continuous, while cli.py:14937 restarts recording after a turn when that flag remains armed. The added locked assignments address that state transition without changing the recording branch.

Problems

• The PR changes only cli.py; it adds no regression coverage for the classic CLI binding. tests/hermes_cli/test_voice_wrapper.py:682 tests the separate hermes_cli.voice API, and tests/tools/test_voice_cli_integration.py:711 currently uses source/AST checks rather than executing handle_voice_record.

Suggested changes

• Add a runtime test for Ctrl+B/configured record-key handling with _voice_continuous=True in both _agent_running and _voice_processing states, asserting the flag is cleared and recording is not started.

Automated hermes-sweeper review.

[teknium1]

Please add a runtime regression test for both new disarm branches. The existing voice-wrapper test exercises hermes_cli.voice.stop_continuous(), not this classic CLI keybinding, so it would not catch a future regression here.

[alt-glitch]

This was generated by AI during triage.

Related to #67550 and #67573. The competing patches differ in the exact non-recording paths that disarm continuous mode.

[Enough1122]

cc @alt-glitch — re: the competing #67550 / #67573 note. Updated the PR body to make the boundary explicit. Summary:

#67569 (this) is intentionally the narrowest possible fix. It only touches the two early-return paths the issue reporter actually hit:

• _agent_running == True → disarm continuous, return
• _voice_processing == True → disarm continuous, return

No global hotkey-toggle semantic. #67550 and #67573 both propose putting a disarm check at the top of the else branch, which would also change behavior on:

• TTS playback
• _clarify_state / _sudo_state / _approval_state / _slash_confirm_state

Those paths are not in the scope of #67545, and changing them is a policy call (some users may rely on Ctrl+B being inert during a confirmation prompt so they don't accidentally drop continuous mode mid-confirmation). I didn't want to make that call as part of a "fix the reported bug" PR.

Recommendation:

1. If maintainers prefer the global toggle semantics → close this in favor of fix(cli): disarm continuous voice on hotkey press during STT/agent (#67545) #67550 or fix(cli): Ctrl+B voice hotkey disarms continuous during STT/agent #67573. The diff here is a strict subset of either; no merge conflicts expected.
2. If maintainers prefer the narrow fix → keep this one open. fix(cli): disarm continuous voice on hotkey press during STT/agent (#67545) #67550 and fix(cli): Ctrl+B voice hotkey disarms continuous during STT/agent #67573 should then close (or re-scope to just the issue's two paths, which is what this PR already does).

I've added a "Scope vs. #67550 / #67573" comparison table to the PR body so the choice is obvious from the description without needing to read the other two diffs.

Branch fix/67545-voice-continuous-hotkey-disarm is unchanged — only the PR body moved.

— written by Hermes Agent on behalf of @Enough1122

[Enough1122]

Review feedback addressed — a09b6e0a7 adds 3 runtime tests in tests/tools/test_voice_cli_integration.py:

• test_disarms_when_agent_running
• test_disarms_when_voice_processing
• test_does_not_disarm_when_no_guard_active (negative control)

The handler is extracted from cli.py AST and exec'd against a real CLI instance built via _make_voice_cli, so the assertions fail if either of the locked _voice_continuous = False assignments in #67569 is reverted. — written by Hermes Agent on behalf of @Enough1122

[Enough1122]

Closing this stale PR for now: the current patch still has unresolved review/test-scope gaps, and it is unlikely to be merged in its present form. Reopen or submit a focused follow-up if the issue remains relevant.