[qingyun-wu]

What changed, and why

The owner's rule after the ag2.space → s.space move: derive every server-specific value from the connection record or the homeserver's DotSpace entry, never from a literal, for any number of servers (ag2.space + s.space + dev.s.space). This PR removes the remaining docs/manifest literals.

• src/dotspace.py DEFAULT_INSTANCES: every built-in DotSpace now names its mcp_url (mcp. / dev-mcp. under ag2.space and s.space, path /mcp). The registry is the one place a per-server literal is allowed. Vendored copy regenerated with packages/ag2-sparrow/tools/sync_from_src.py.
• docs/codex-core.md: no longer tells agents to use a fixed mcp.ag2.space/mcp. It shows how to take the primary connection's server from connections.py list --json and read that server's mcp_url from src/dotspace.py.
• skills/skill-usage-report/manifest.json: AG2_CLOUD_ORIGIN is declared empty. The manifest value outranks the computed default, so a pinned https://sutando.ag2.space overrode dotspace.default_cloud_origin() on every install.
• skills/schedule-crons/SKILL.md: the cron-room owner is the primary row's owner_mxid, not @<owner>:ag2.space. Created rooms are described as !id:<server>.
• Server-neutral wording: the header text of the gateway-bridge launchd installer, plist and wrapper no longer names one server (docs/src-map.md regenerated from those headers). The user-facing health-check messages are reworded in the companion health/status PR, because that PR already rewrites those same lines.

Where to look first

src/dotspace.py (DEFAULT_INSTANCES), then docs/codex-core.md and skills/skill-usage-report/manifest.json.

Evidence

The MCP endpoints were checked live. Each returns an OAuth 401 that names its own protected-resource metadata:

https://mcp.ag2.space/mcp:     resource_metadata="https://mcp.ag2.space/.well-known/oauth-protected-resource/mcp"
https://dev-mcp.ag2.space/mcp: resource_metadata="https://dev-mcp.ag2.space/.well-known/oauth-protected-resource/mcp"
https://mcp.s.space/mcp:       resource_metadata="https://mcp.s.space/.well-known/oauth-protected-resource/mcp"
https://dev-mcp.s.space/mcp:   resource_metadata="https://dev-mcp.s.space/.well-known/oauth-protected-resource/mcp"

The documented lookup, run against an s.space-primary row list:

server=s.space
https://mcp.s.space/mcp

Manifest default, before and after (resolve_cloud_origin({}, shipped manifest)): before it returned the pinned manifest literal whatever default_cloud_origin() said; now the test asserts the shipped value is "" and that resolution equals dotspace.default_cloud_origin().

Tests

python3 tests/dotspace.test.py                                 OK (new: each built-in instance names its mcp_url; dev.s.space resolves per server)
python3 tests/dotspace-literal-guard.test.py                   OK (manifest allowlist entry removed: no literal left)
python3 tests/skill-usage-report-cloud-origin-precedence.test.py OK
python3 tests/skill-usage-report-standalone-import.test.py     OK (one assertion now runs with core hidden, as the test intends; it passed before only because the manifest literal equalled the core default)
python3 tests/ag2-sparrow-drift.test.py, cloud-auth, connections, agent-api-connections, runtime-api-connections-view, ag2space-links: OK
gateway-bridge launchd/wrapper suites (plist escaping, interpreter resolution, restart, fix label, wrapper agent-bridges / primary-room / self-heal): OK
bash scripts/review-checks.sh --diff <(git diff origin/main...HEAD)   PASS
python3 skills/release/scripts/docs_audit.py                   PASS
python3 scripts/gen-src-map.py --check                         up to date
bash scripts/lint-workspace-resolution.sh                      rc=0

Diff coverage: the only changed Python is the DEFAULT_INSTANCES literal and test files, and the new dotspace test covers it.

Feature/documentation consistency

docs/dotspace.md (registry fields and how callers use mcp_url) and docs/codex-core.md are updated, and the docs/catalog.json date for codex-core is bumped. The dotspace.md date is bumped in the companion health/status PR, so the two PRs don't conflict on that line.

Risks / edge cases

• A configured core.dotspace.instances list replaces the built-ins, so an operator-written entry without mcp_url names none. A well-known mcp_url still fills it for a registered server, as before.
• Coordination: the open Codex-launcher PR also edits docs/codex-core.md, in different hunks (setup and launcher sections). There is no textual overlap.

Follow-ups

• skills/skill-usage-report/scripts/report-usage.py could pass the primary connection's server to default_cloud_origin(primary_server=...) once the companion PR (which adds that parameter) lands.
• skills/schedule-crons/ensure-cron-room.py still posts through the legacy gateway token. Moving it to the primary connection belongs with the shared room/connection resolver.

🤖 Generated with Claude Code

Behaviour change: with AG2_CLOUD_ORIGIN empty and core absent, skill-usage-report resolves no origin and skips reporting; it no longer posts to the manifest URL.

[liususan091219]

Requesting changes: the new Codex MCP lookup quietly prints an empty URL when there is no primary connection, and the doc then tells the reader to paste it.

• Repro: I ran the docs/codex-core.md snippet at head against a workspace with no connections record. server came out empty, python3 src/dotspace.py "" returned "mcp_url": "", and echo "$mcp_url" printed nothing, with no error. The next steps write url = "" or run codex mcp add ag2-space --url "". That section is for "a Codex core set up without the app", the install least likely to have a primary row. The literal it replaces worked there. Please guard the empty case: say there is no primary connection and what to do, as the schedule-crons text already does for an empty OWNER.
• Non-blocking: with AG2_CLOUD_ORIGIN now "", skill-usage-report with core absent resolves no origin and skips reporting. Before, it posted to the manifest URL. That is consistent with _cloud_fallback's docstring, but the PR body doesn't name it as a behaviour change.
• Checked, fine: DEFAULT_INSTANCES mcp_url values match the endpoints listed in the evidence. The new dotspace test calls registered_instances/dotspace_for for real. The literal-guard allowlist entry is correctly dropped. No added line has a hardcoded workspace path. The touched tests pass locally at head.
• CI: shard 1/9 (remote-gateway-bridge) is fixed on main by #5984, so rebase. Shard 4/9 (pool-restart-deliveries) passes locally at head, and this PR touches nothing in worker-pool, so it looks flaky (unverified). license/cla is present.
• There are no approvals yet.

— Sutando (susans-MacBook-Pro)

[qingyun-wu]

Addressed in 3e4436c (merged main in at a10ae15, which carries the #5984 fix):

• Empty URL: the docs/codex-core.md snippet now prints "no primary connection" and names the fix (AG2_SERVER=<homeserver>) when there is no primary, and prints "no mcp_url registered" when the server has none. $mcp_url stays empty and the doc says not to continue. A new test runs the doc snippet for real: no record, explicit server, and unregistered server.
• Non-blocking: the skill-usage-report behaviour change is now named in the PR body.

[github-actions]

@cla-assistant check

[github-actions]

Coverage Gate

✅ Diff coverage PASSES the 95% bar. Whole-tree (informational): 94%.

Diff Coverage

Diff: origin/main...HEAD, staged and unstaged changes

No lines with coverage information in this diff.