[roboco-app]

Objective

Every recorded learning writes its notifications twice: once as DB rows via _persist_notifications, and once as in-memory LearningNotification objects appended by _enqueue_in_memory (roboco/services/learning.py:283-304), one row per notified agent. The in-memory copy has no production reader: get_pending_notifications and acknowledge_notification (learning.py:449-502) only filter and flip a flag, and no route, gateway verb, or service calls either (the notifications API acks through NotificationDeliveryService instead, see roboco/api/routes/notifications.py:107). Acknowledged entries are therefore never removed, and the queue grows monotonically in the orchestrator process until restart, one entry per notified agent per learning.

Note this task's original headline (CELL-scoped learnings notify the entire org because of a TODO in _fetch_notify_agents) is already fixed on master: commit 04eb61d (PR #914, 2026-08-12) filters CELL scope to the author's team. The remaining, still-real defects are the unbounded queue and the stale test that pretends that filter does not exist: test_cell_scope_runs_without_role_filter only passes because both of its fake agents are hardcoded Team.BACKEND. Fix: prune or bound the in-memory queue, or drop it in favor of the DB-backed notification path, and rewrite the stale cell-scope test.

Notes

• Evidence: roboco/services/learning.py:283-304 appends one in-memory row per notified agent on every record_learning; learning.py:449-502 never removes acknowledged entries and a repo-wide grep shows zero callers outside the module; roboco/api/routes/notifications.py:107 acks via NotificationDeliveryService, not this queue. CELL fanout was narrowed to the author's team by 04eb61d (PR [9efb36e4] CELL-scoped learning notifications bypass team filter — notify all agents #914, 2026-08-12), so the org-wide claim in the original finding is stale. tests/integration/test_learning_service.py:184-205 carries the pre-filter name, docstring, and a same-team-only assertion.
• Delegation hint: originated as a backend item - delegate into the backend cell.

Success Criteria

• The in-memory _notification_queue no longer grows without bound: acknowledged entries are pruned on acknowledge, or the queue is bounded, or it is removed in favor of the DB-backed notification path it duplicates (roboco/services/learning.py)
• tests/integration/test_learning_service.py test_cell_scope_runs_without_role_filter is rewritten to assert the CELL team filter honestly: same-team agents are notified and cross-cell agents are excluded, under a name and docstring matching the post-[9efb36e4] CELL-scoped learning notifications bypass team filter — notify all agents #914 behavior
• TEAM (same role) and ORG (all non-human agents) fanout behavior is unchanged, and make quality is green

[github-actions]

Thanks for opening your first pull request on RoboCo!

Quick checklist before review (most of these are enforced by CI, but worth a glance):

• make quality — ruff format check, ruff check, mypy, pytest (≥80% coverage), and the rest of the gate
• Panel changes pass pnpm lint and pnpm exec tsc --noEmit (run from panel/)
• No # noqa / # type: ignore shortcuts; pre-existing violations in touched files are fixed
• Added an entry under ## [Unreleased] in CHANGELOG.md
• Signed the CLA (the bot will prompt you on this PR)
• Signed your commits — master requires verified signatures (SSH signing setup)
• Updated any affected docs under docs/

See CONTRIBUTING.md for the full workflow and the Code of Conduct for the community standards we follow.

Welcome aboard — a maintainer will review shortly.

[roboco-app]

In-path PR-review gate — PASSED ✅

Reviewed by pr_reviewer (RoboCo PR reviewer). Posted by the project bot account; the gate verdict is authoritative in RoboCo.

Checked each criterion against the code. Criterion 1: learning.py:344 makes _notification_queue a deque with maxlen=_NOTIFICATION_QUEUE_MAXLEN (256); acknowledge_notification (learning.py:321-326) prunes all acknowledged entries and keeps the maxlen; the Protocol type in _learning_propagation_types.py is updated to match. Unit tests cover eviction at the cap, prune-on-ack and pending filtering. Criterion 2: test_cell_scope_runs_without_role_filter is renamed to test_cell_scope_notifies_author_team_excludes_cross_cell, with a docstring that describes the post-#914 behaviour. It adds a Team.FRONTEND agent and asserts on persisted KNOWLEDGE_SHARE rows that the same-team peer is a recipient and the cross-cell agent is not. Criterion 3: _fetch_notify_agents and the TEAM/ORG branches are untouched; make quality is left to the CI guard. The docs file is in the established docs/backend/services location. Nits, not blocking: the doc cites PR #1119 (the cell PR) rather than #1406, and a harmless comment on the ORG truncation test was removed.

Only the CEO merges this PR into master.