[smicbee]

What does this PR do?

Prevents accepted /queue turns from being silently dropped when the gateway reaches its interrupt recursion cap.

The failure occurs after the current FIFO head has been drained and _promote_queued_event() has already staged the next item in the adapter's single pending slot. The recursion-cap path then requeues the current event by merging it into that same slot, overwriting the staged successor.

Example state transition before this fix:

initial:             slot=Q1, overflow=[Q2, Q3]
after drain/promote: pending=Q1, slot=Q2, overflow=[Q3]
recursion-cap requeue (old): slot=Q1, overflow=[Q3]  # Q2 lost

This PR restores the unprocessed current event at the FIFO head while first moving any staged successor back to the front of overflow:

recursion-cap requeue (new): slot=Q1, overflow=[Q2, Q3]

Each accepted queue item therefore remains a separate full turn in original arrival order.

Related Issue / PR overlap

No dedicated issue found.

The recursion-cap behavior also appears inside broader platform-specific PRs #72674 (Email boundaries) and #76818 (Slack recovery). This PR intentionally isolates the generic message-loss fix and regression test on current main, without taking either PR's unrelated Email or Slack changes.

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 🔒 Security fix
• 📝 Documentation update
• ✅ Tests (adding or improving test coverage)
• ♻️ Refactor (no behavior change)
• 🎯 New skill (bundled or hub)

Changes Made

• gateway/run.py
  • Add _requeue_event_at_front() to restore the drained event without overwriting an already promoted successor.
  • Use the helper at the interrupt recursion cap instead of merge_pending_message_event().
• tests/gateway/test_queue_consumption.py
  • Add a Q1 → Q2 → Q3 regression that reproduces drain, promotion, recursion-cap requeue, and complete FIFO consumption.

How to Test

scripts/run_tests.sh \
  tests/gateway/test_queue_consumption.py \
  tests/gateway/test_queue_command.py \
  tests/gateway/test_pending_drain_no_recursion.py \
  tests/gateway/test_pending_drain_race.py \
  tests/gateway/test_active_session_text_merge.py \
  tests/gateway/test_steer_fifo_overwrite.py \
  tests/gateway/test_busy_session_ack.py \
  tests/gateway/test_discord_pending_text_batch_shutdown.py \
  -q

python -m ruff check gateway/run.py tests/gateway/test_queue_consumption.py
git diff --check origin/main...HEAD

Local result on Linux:

• 35 passed, 0 failed across the 8 affected Gateway files
• Ruff: passed
• git diff --check: passed

The full repository suite was started successfully but not completed locally because it currently contains more than 20,000 tests; the focused Gateway regression set above was run through the canonical wrapper.

Checklist

Code

• I've read the Contributing Guide
• My commit message follows Conventional Commits
• I searched existing issues and PRs and documented the overlapping broader PRs above
• This PR contains only the generic queue-loss fix and its regression test
• Full repository test suite completed locally
• Added behavioral regression coverage
• Tested on Linux

Documentation & Housekeeping

• Documentation update: N/A; no user-facing behavior or configuration is added
• cli-config.yaml.example: N/A; no config keys changed
• CONTRIBUTING.md / AGENTS.md: N/A; existing FIFO architecture is preserved
• Cross-platform impact considered; change uses platform-independent Python data structures
• Tool descriptions/schemas: N/A

[Adolanium]

I checked this on current main and the bug is still there. This PR is the right fix, but it no longer applies because gateway/run.py was split into siblings. What is needed to land it:

1. Move the cap site. The recursion cap is now in gateway/run_turn.py (around line 3814). It still calls merge_pending_message_event(adapter._pending_messages, session_key, pending_event), which replaces the staged next item. Swap that call for _requeue_event_at_front(...).
2. Move the helper. _promote_queued_event, _enqueue_fifo and _queue_depth now live in gateway/run_busy.py, so _requeue_event_at_front belongs next to them. The overflow list is self._session_state(session_key).conversation.queued_events (see _overflow_queue in the same file), which matches what the helper already uses.
3. Drop the queue_message branch right after the cap call. No adapter defines queue_message any more, so that branch is dead.

The Q1, Q2, Q3 FIFO test is a good invariant test and can move over as is, next to the other tests in tests/gateway/test_queue_consumption.py.

After those changes it is a small, focused diff. I have not opened a competing PR, so this one can carry the fix.

[driptothebone]

I reproduced the recursion-cap overwrite on current main: putting the unrun head back into the adapter slot replaces the next FIFO item that was already staged. I have a private port of your fix to the split gateway modules, with real runner/adapter tests that follow ten distinct messages through the depth cap and check separate text/photo/document payloads. Would you welcome a port and those tests on this PR? I would build on your work rather than open a competing PR.

[smicbee]

@Adolanium thanks, that was exactly the right pointer. I ported the fix to the split modules and force-pushed the branch (rebased on current main, single commit bea3d26):

• _requeue_event_at_front() now lives in gateway/run_busy.py next to _promote_queued_event / _enqueue_fifo, using the queued_events overflow list.
• The cap site in gateway/run_turn.py calls it instead of merge_pending_message_event(); the dead queue_message branch and the now-unused import are removed.
• The Q1→Q2→Q3 FIFO test is carried over unchanged in tests/gateway/test_queue_consumption.py.

Local run of the 8 affected gateway test files: 37 passed, 0 failed. Diff is 3 files, +64/-4.

@driptothebone thanks for reproducing this and for the offer. I've just finished the port myself, but your real runner/adapter tests (ten messages through the depth cap, separate text/photo/document payloads) would be a valuable addition. Feel free to open a PR against my branch smicbee:fix/queue-recursion-fifo-loss or paste them here, and I'll fold them in.

[smicbee]

@Adolanium thanks again for the pointer. The port to the split modules is in (bea3d26), and @driptothebone's real runner/adapter tests are now merged into this branch as well (head 6147b53): ten messages through the depth cap, plus separate TEXT/PHOTO/DOCUMENT payloads with and without an overflow tail. With the fix, 22 tests pass in the two affected files. With gateway/ reverted to the pre-fix state, exactly the 7 new tests fail. This should be ready for another look.

[kvnloo]

FIFO/requeue review: the updated split-module port preserves event identity instead of feeding the recursion-cap item back through merge semantics. _requeue_event_at_front() first moves the already-promoted successor back to overflow, then restores the unprocessed head in the adapter slot, which gives Q1,Q2,Q3 back exactly. The real-runner tests also exercise ten turns across depth resets and TEXT/PHOTO/DOCUMENT payloads, so this is no longer only a helper-level invariant. I don't see a blocker in the current port.