Repro

$ mm agent register planner --description "Architecture & roadmap"
Agent registered: planner
- Namespace: agent-runtime:planner
- Shared namespace: shared
$ mm agent register coder
Agent registered: coder
- Namespace: agent-runtime:coder
- Shared namespace: shared
$ mm agent list
Agents: 0

Shared: shared (0 chunk(s)) — Shared knowledge base for all agents

Both registers reported success and wrote a namespace_metadata
row, but the listing showed zero agents. The user-visible symptom
is "did the register actually take effect?"; the answer is yes,
the listing was just blind to namespaces that had no chunks yet.

The Web UI's GET /namespaces route had the same blind spot via
the same storage method — registered-but-empty namespaces never
appeared in the namespaces tab.

Root cause

list_namespace_meta in sqlite_namespace.py was

SELECT c.namespace, COUNT(*), m.description, m.color
FROM chunks c
LEFT JOIN namespace_metadata m ON c.namespace = m.namespace
GROUP BY c.namespace

Iterating from the chunks side meant a namespace_metadata row
with zero matching chunks rows produced zero output rows. Newly-
registered agents have zero chunks until someone writes into their
namespace, so mm agent register planner && mm agent list was
guaranteed to lose planner.

shared continued to show up in mm agent list only because the
CLI fetches it separately via get_namespace_meta(SHARED_NAMESPACE)
(agent_cmd.py:168), which reads namespace_metadata directly.
That asymmetry made the bug class non-obvious.

Fix

Source list_namespace_meta from the union of
namespace_metadata.namespace and chunks.namespace, then LEFT
JOIN both sides for chunk count + description/color:

SELECT
    ns.namespace,
    COALESCE(c.chunk_count, 0),
    COALESCE(m.description, ''),
    COALESCE(m.color, '')
FROM (
    SELECT namespace FROM namespace_metadata
    UNION
    SELECT DISTINCT namespace FROM chunks
) ns
LEFT JOIN (SELECT namespace, COUNT(*) AS chunk_count FROM chunks GROUP BY namespace) c
    ON c.namespace = ns.namespace
LEFT JOIN namespace_metadata m ON m.namespace = ns.namespace
ORDER BY ns.namespace

Three states surface correctly now:

No caller-side signature change. The same return shape
({namespace, chunk_count, description, color}) is preserved, so
the Web GET /namespaces route picks up the fix transparently.

Why tests didn't catch this

tests/test_agent_cmd.py stubs the storage with AsyncMock
returning canned list_namespace_meta payloads — the SQL bug
never tripped because the SQL never ran. This is the
feedback_storage_artifact_false_pass.md pattern in memory:
asserting on processing-stage spies instead of the actual
storage-layer artifact gives a false PASS.

Two regression tests added in
tests/test_server_tools_org.py::TestNamespace against the real
SQLite backend:

1. test_list_namespace_meta_includes_registered_empty_namespace
   — exact reproducer (set_namespace_meta then
   list_namespace_meta with no upsert_chunks in between).
2. test_list_namespace_meta_unions_chunks_and_metadata —
   enumerates all three union states in one fixture so a
   regression on either side fails loudly.

How it was found

External tester running the v0.1.28 multi-agent test scenarios
walkthrough (memtomem-docs/memtomem/testing/multi-agent-test-scenarios.md,
private repo PR #3) caught this at scenario 1 step 1 — exactly the
class of bug the scenarios doc was written to surface. This is a
data point for the doc's value: gate-2-precondition (external e2e
report) effectively delivered a real bug report on first contact.

Test plan

• uv run pytest packages/memtomem/tests/test_server_tools_org.py::TestNamespace -q — 21 passed (incl. 2 new)
• uv run pytest packages/memtomem/tests/test_agent_cmd.py packages/memtomem/tests/test_web_routes_extended.py -q — 72 passed
• uv run ruff check + format --check — clean
• Manual: rerun the repro after merge — mm agent register planner && mm agent list should show Agents: 1.

🤖 Generated with Claude Code