fix(discord): scope auto_thread/reactions/allow_mentions/reply_to_mode per profile

[nftpoetrist]

What & why

_apply_yaml_config()'s YAML→env bridge for auto_thread, reactions, allow_mentions.*, and reply_to_mode writes to process-global os.environ unconditionally (first-writer-wins, no _skip_env_bridge check) — unlike the auth-gate keys (allow_from, allowed_roles, allow_all_users, free_response_channels, ignored_channels, allowed_channels, no_thread_channels) already fixed for this exact bug class under issue #72348.

Under gateway.multiplex_profiles, a secondary profile's own config.yaml settings leak into shared env for every other Discord adapter in the process to inherit. Worst case: allow_mentions.everyone: true on one profile silently lets every profile's bot ping @everyone/@here on its own server — Discord bots default to parsing those pings, and this adapter's own docstring explains the safe-default exists specifically to stop LLM output/echoed content from doing that. reply_to_mode is the reverse-direction case: a secondary profile's own setting leaks into the (unscoped) default profile's read.

The read sides (_reactions_enabled, the inline auto_thread check in _handle_message, _build_allowed_mentions) also read raw os.getenv/os.environ directly with no per-profile fallback at all.

Fix

Reuses the established per-profile machinery from issue #72348 instead of inventing a new mechanism:

• _apply_yaml_config: gates the auto_thread/reactions/allow_mentions/reply_to_mode env writes with the existing _skip_env_bridge check (_profile_scoped_config_load()), and seeds auto_thread/reactions/allow_mentions into PlatformConfig.extra (they weren't seeded at all before, so a scoped profile's own YAML value would otherwise have nowhere to go).
• _reactions_enabled / new _auto_thread_enabled: use the existing self._gate_raw() per-profile accessor (the same one _get_allowed_channels etc. already use) instead of raw os.getenv, and add both env vars to _GATE_ENV_KEYS so they're captured in the connect()-time per-profile snapshot.
• _build_allowed_mentions: accepts an optional extra param (the call site now passes self.config.extra), and uses _scoped_gate_env (the same helper the gate accessors are built on) instead of raw os.getenv for each of the four DISCORD_ALLOW_MENTION_* flags, falling back to extra["allow_mentions"] when unset. Existing bare _build_allowed_mentions() calls (tests, and any future caller with no adapter instance) are unaffected — extra defaults to None, preserving the exact prior env-only behavior.

Tests

Extended tests/plugins/platforms/test_discord_gate_isolation.py (the established test file for issue #72348's bug class) with 16 new tests mirroring its existing conventions (two-adapter isolation, snapshot-vs-extra precedence, env-does-not-leak-into-snapshotted-adapter, and _apply_yaml_config scoped-vs-single-profile write behavior).

• tests/plugins/platforms/test_discord_gate_isolation.py — 37 passed
• tests/gateway/test_discord_*.py, tests/tools/test_discord_*.py, tests/hermes_cli/test_discord_*.py, tests/plugins/test_discord_*.py (full Discord suite) — 346 passed, 3 pre-existing failures in test_discord_send.py confirmed independent of this change (pass 11/11 when run standalone, both with and without this fix — a test-order-pollution artifact of running the whole suite together, not a regression)
• Mutation-verify: reverting the fix makes 12 of the 16 new tests fail as expected (the other 4 exercise the unscoped/single-profile path, which was never buggy and correctly stays green)
• Needed a temporary uv pip install "discord.py[voice]==2.7.1" per the standing optional-dependency gotcha, to get real dynamic verification instead of importorskip skips on the AllowedMentions-dependent tests

Competitor check

Searched DISCORD_AUTO_THREAD, DISCORD_REACTIONS, DISCORD_ALLOW_MENTION, DISCORD_REPLY_TO_MODE, "discord multiplex allow_mentions" (all PR/issue states) — no PR targets this scoping bug. One open PR (#68473) touches the same _handle_message region (adds a logger.debug call in the neighboring require_mention drop path) — different lines, different concern, no overlap.

[Enough1122]

AI code review — automated review for reference; please use your judgment.

Another consistent application of the #72348 scope-isolation pattern to the Discord adapter: DISCORD_REACTIONS/DISCORD_AUTO_THREAD join the gated env list, _reactions_enabled/_auto_thread_enabled switch to the per-profile _gate_raw accessor (adapter.py:3369-3390), _build_allowed_mentions gains a scope-aware extra.allow_mentions path (adapter.py:525-563), and _apply_yaml_config seeds the same values into seeded_extra while gating the env-bridge writes on _skip_env_bridge (adapter.py:10511-10549) — including the reverse-direction reply_to_mode leak. Test coverage is thorough: isolation between snapshotted adapters, safe defaults, process-env non-leakage, and both bridge directions for every touched key.

Points worth noting:

• The parse semantics remain asymmetric: _reactions_enabled is deny-list based (not in {"false","0","no"} → any unknown value enables reactions), while _auto_thread_enabled is allow-list based (any unknown value disables auto-threading) (adapter.py:3375-3376, 3385-3386). This mirrors the pre-existing behavior, so it's not a regression, but since both were touched it would be a good moment to decide whether an unknown value should fail safe in both directions.
• In the scoped path, _scoped_gate_env (i.e. the profile's own .env) takes precedence over extra.allow_mentions (the profile's config.yaml) in _build_allowed_mentions (adapter.py:549-552). Both are the profile's own config, so this is a defensible precedence, but it may be worth documenting that env wins over YAML when both set a scoped profile's mention policy.

Verdict: LGTM

[teknium1]

Superseded by #108440 (on main as 545e74d), with Co-authored-by credit to you. Discord's auto_thread, reactions, allow_mentions.*, history_backfill and approval_mentions bridges now go through the shared yaml_env_setter (skipped under a secondary scope, seeded into extra) and their readers consult extra under scope — the same fix you wrote here, applied uniformly across every adapter in one commit instead of per platform. Thanks, @nftpoetrist.