fix(discord): continuously recover missed gateway messages

[maxibotstef]

Summary

Hermes can remain nominally connected to Discord while silently missing a gateway event. The existing recovery scan runs only at startup/ready time, so a missed human message can remain unanswered until the next restart.

This change:

• runs the existing durable Discord missed-message scan periodically, with one adapter-owned loop and serialized scans;
• reuses the existing authorization, ignored-channel, mention, bot-loop, deduplication, response-ledger, and dispatch paths;
• atomically claims recovered message IDs so a live event racing recovery cannot dispatch twice;
• admits a human reply to another bot only in configured free-response scope while retaining the direct-mention and bot-loop gates;
• adds bounded configuration, lifecycle, race, durable-completion, and incident-shaped regression coverage.

Incident reproduced

An untagged human ? appeared in a configured free-response Discord channel after gateway startup. The WebSocket remained connected, but the live event was absent from gateway dispatch logs and startup-only recovery did not revisit it. This is distinct from a reconnect-only outage.

Implementation

• discord.missed_message_backfill.interval_seconds defaults to 60.
• 0 keeps the ready-time scan while disabling periodic polling.
• Invalid, negative, and non-finite intervals fail closed with a warning; intentional 0 is quiet.
• Repeated ready events coalesce onto one periodic task.
• The periodic owner awaits the existing one-shot scan task, preventing overlap.
• disconnect() cancels the periodic owner before the active scan.
• Per-pass limit and max_dispatches bounds remain in force.
• Recovered admission uses claim=True; failure/cancellation paths retain the existing claim release behavior.

Verification

Exact reviewed head: 6b3111e8f14da7f9ddf1cca09d091f4032a9c7ca

133 passed, 2 existing AsyncMock warnings in 11.19s
python3 -m py_compile: passed
git diff --check: passed
changed paths: exactly 5

Focused suites:

python -m pytest -q \
  tests/gateway/test_discord_free_response.py \
  tests/gateway/test_discord_missed_message_backfill.py \
  tests/gateway/test_discord_liveness.py \
  tests/gateway/test_voice_command.py

Independent Opus review: GO_WITH_CAVEATS, no blockers. The review specifically verified the recovered/live-event race, consecutive-pass durable completion, other-bot narrowing, interval validation, serialized task ownership, and shutdown order.

Related open PRs / overlap disclosure

• fix(discord): claim dedup ID in missed-message backfill dispatch #75067 independently fixes the recovered-message dedup claim subset. This PR includes that same bug-class correction because atomic claim ownership is required for periodic connected-state recovery; fix(discord): claim dedup ID in missed-message backfill dispatch #75067 has narrower and deeper claim-specific tests.
• fix(gateway): verify macOS restarts and recover Discord gaps #72881 combines launchd readiness work with outage recovery and same-content suppression. This PR does not add content-fingerprint suppression and targets connected-state missed events with the existing durable ledger.
• refactor(discord): extract recovery backfill + reaction hooks into DiscordRecoveryBackfillMixin (adapter.py god-file slice R2) #79651 extracts the recovery cluster from the Discord adapter. This PR intentionally targets current main; maintainers may need to sequence the behavioral fix before that refactor.

Caveats / activation checks

• The environment-variable fallback is used only when the whole missed_message_backfill config mapping is absent; config-file precedence is unchanged.
• The configured interval should be operationally sane. The production canary uses 15 seconds and watches REST/rate-limit logs and event-loop latency.
• Unit tests are L2 evidence only. Live recovery, response delivery, and duplicate suppression are verified separately after supervised activation.

Scope

Changed paths only:

1. hermes_cli/config_defaults.py
2. plugins/platforms/discord/adapter.py
3. tests/gateway/test_discord_free_response.py
4. tests/gateway/test_discord_missed_message_backfill.py
5. tests/gateway/test_voice_command.py

No scheduler, watchdog service, database, auth weakening, liveness-threshold change, broad adapter refactor, or unrelated delivery change.

[Enough1122]

AI code review — automated review for reference, author can ignore or act on any point.

fix(discord): continuously recover missed gateway messages

1. Admission behavior change: in free-response channels, a human reply that mentions another bot (but not Hermes) is now admitted (other_bots_mentioned and not raw_self_mention and (author_is_bot or not is_free_response or message.type != reply)). The intent is clear and tested, but this widens the bots response surface — Hermes will now join conversations directed at other bots in free-response channels. Consider an additional guard (e.g. only when the replied-to message is Hermes own, or when the reply reads like a question).
2. Default interval_seconds: 60 with a 6-hour window and a 100-message per-pass limit means a full window walk every minute — repeatedly paging through Discord history could hit API rate limits on busy channels. Consider a higher default, or backing off when a pass found nothing new.
3. _dispatch_recovered_message(..., claim=True) claims before dispatch — if _handle_message subsequently fails, is the claim rolled back? If not, a failed recovered dispatch permanently loses the message (it would be skipped by both the live path and later recovery passes). Verify the claim is success-only.
4. The periodic loop coalesces by design (_ensure_missed_message_backfill_task returns the in-flight task without awaiting), so an in-flight scan skips that tick; effective cadence is "at most once per interval" — matches the docstring, just flagging as intended.