fix(agents): stop dropping an agent's reply when another agent posts first (#757)

[lilyshen0722]

Closes #757.

The bug

Two BYO-wrapper agents mentioned in one message: whichever finishes second has its entire reply silently discarded, and the event is still acked so nothing surfaces the loss. Observed live in pod 6a65c153a0087c7e86f426ee — Sam has been hand-sequencing "@critic-codex — you only", "@strategist-fable — you only" all morning to work around it.

The wrapper's self-post detection asked "did any bot post during my turn?" as a proxy for "did I post?":

if (!preSpawnIds.has(String(m._id || m.id)) && m.isBot) { … }

Equivalent in a single-agent pod. Wrong in the multi-agent rooms that are the core use case.

Why the fix is server-side

The obvious fix — have the wrapper derive its own bot username and match on it — is unsafe. The server applies resolveAgentType's LEGACY_AGENT_MAP (server-only) before buildAgentUsername, and instanceId is owner-scoped and assigned at install, not the CLI --name. A client that guesses wrong never matches → guaranteed double-post, the exact failure the mechanism exists to prevent.

So the server answers it. getRecentMessages(podId, limit, selfUserId) stamps an explicit self boolean per message, wired through both agent-facing message routes. Omitting the argument emits no self key at all — that absence is how the wrapper detects an older server and degrades to the legacy any-bot test rather than breaking.

Three siblings found while verifying

Each real, each fixed, each with a test:

Mongo fallback had isBot missing from the projection	.populate('userId', 'username profilePicture') → isBot always false → detection dead outright, guaranteed double-posting whenever PG is unavailable. Untested because every backend test mocks getRecentMessages.
PG path fabricated isBot from a username substring	'-bot' / endsWith 'bot' / 'openclaw-' → a human named talbot or abbot reads as a bot and suppresses an agent's reply. The mirrored is_bot column is authoritative.
hasRecentDuplicateUrls had no author filter	A human posting a link suppressed the agent's next heartbeat post.

Observability

Suppression now names the message id, its author, and whether the decision came from self or the legacy fallback. #757 went unnoticed for as long as it did because that log line said only agent posted via tool this turn.

Test

cli 17/17 suites, 186 tests. New backend suite 8/8. Both proven load-bearing by reverting the fix:

• #757: ANOTHER agent posting during the spawn must NOT suppress this reply — fails on pre-fix code, passes after
• the 3 backend behaviours each have a test that fails when that behaviour is reverted
• the fix(cli): consistent wrapper-agent behavior — mount skill by default + kill double-post via self-post detection #624 double-post guard and the fix(agents): agents must post as themselves — never via an operator profile #702 human-author carve-out stay green in both states

Backend lint warnings are at baseline; the +6 errors are the import/no-unresolved + import/extensions pair on 3 imports, which every backend test file in the repo carries.

🤖 Generated with Claude Code