fix(gateway): allow settings slash commands during active sessions

[cometjc]

Summary

Slash commands like /model, /verbose, /fast, /reasoning, /yolo, /personality, /provider, /voice are silently eaten when sent on Telegram while an agent is actively running. Instead of invoking the command handler, the user gets the ⚡ Interrupting current task... ack and the command text is fed into the running agent as interrupt input.

Root Cause

gateway/platforms/base.py:1614-1634 checks should_bypass_active_session(cmd) before any per-command logic in gateway/run.py runs. Commands not in the bypass allowlist fall through to _busy_session_handler (gateway/run.py:1391), which calls running_agent.interrupt(event.text) and replies with the interrupt ack. The command handler is never reached.

The bypass allowlist ACTIVE_SESSION_BYPASS_COMMANDS in hermes_cli/commands.py previously only covered control/info commands (agents, approve, background, commands, deny, help, new, profile, queue, restart, status, stop, update). Settings commands were missing, so /model (and the other settings toggles) fell through to the busy handler.

Side effect: the /model guard at gateway/run.py:3022-3024 (return "Agent is running — wait or /stop first, then switch models.") became dead code, since the adapter short-circuits before _handle_message is ever called.

Fix

1. Extend ACTIVE_SESSION_BYPASS_COMMANDS with 8 settings commands: model, provider, verbose, fast, reasoning, yolo, personality, voice.
2. Replace the dead /model string-return in gateway/run.py:3022-3028 with an actual dispatch to _handle_model_command, kept as a defensive fallback for adapters that don't call the bypass check or if the allowlist is later narrowed.

Handler safety audit

Each of the 8 added handlers was audited — none call running_agent.interrupt/cancel or mutate _running_agents. They only read/write config:

Command	Handler	What it touches
/model	_handle_model_command	_session_model_overrides, _agent_cache eviction, cached entry switch_model() (public API)
/provider	_handle_provider_command	read-only
/verbose	_handle_verbose_command	display.platforms.<p>.tool_progress YAML
/fast	_handle_fast_command	_service_tier + YAML
/reasoning	_handle_reasoning_command	_reasoning_config, _show_reasoning + YAML
/yolo	_handle_yolo_command	session-yolo helpers in tools/approval
/personality	_handle_personality_command	_ephemeral_system_prompt + YAML
/voice	_handle_voice_command	_voice_mode, adapter voice state

All 8 effects apply to the next turn (or are read-only), so dispatching them mid-run is safe.

Tests

New: tests/gateway/test_active_session_bypass_commands.py (TDD — failed before fix, 23 green after):

• 8 parametrized tests for the added commands
• 13 parametrized regression guards for pre-existing bypass commands
• 2 sanity tests: known-but-excluded command (/skills) and unknown command both return False

Verified green locally:

tests/gateway/test_active_session_bypass_commands.py .... 23 passed
tests/hermes_cli/test_commands.py ........................ passed
tests/gateway/test_busy_session_ack.py ................... passed
tests/gateway/test_verbose_command.py .................... passed
145 passed in 2.12s

Test Plan

• Unit tests passing locally.
• Manual: send /model on a Telegram thread with an agent mid-run — should now open the model picker, not produce the interrupt ack.
• Manual: confirm /stop and /new still interrupt as expected (regression check).

Scope

Touches 3 files (hermes_cli/commands.py, gateway/run.py, new test). No behavioral changes outside the allowlist extension.

[cometjc]

Closing as duplicate — #12334 already landed the more comprehensive fix on main yesterday (commit 632a807). That PR takes the broader approach of treating any resolvable slash command as a bypass, covering /model and also /reasoning, /voice, /insights, /title, /resume, /retry, /undo, /compress, /usage, /provider, /reload-mcp, /sethome, /reset. My allowlist-extension approach here is strictly narrower and now redundant.

Thanks for the clean upstream fix.