fix(templates): publish agent.upsert with the hasStream flag on launch

[selfcontained]

What

apps/server/src/routes/templates.ts was the only agent.upsert publish site in the repo that sent the raw AgentRecord instead of withStreamFlag(...). This threads withStreamFlag into TemplateRouteDeps and wraps the launch publish, plus a colocated test.

Why it's tech debt

There are 31 agent.upsert publish sites (server.ts, server/mcp-handlers.ts x7, server/mcp-review-handlers.ts x2, server/agent-lifecycle-runtime.ts x7, routes/release.ts, routes/agents/{terminal,crud,lifecycle}-routes.ts, routes/templates.ts). Thirty wrapped the agent; this one didn't. Verified with grep -rn --text -A3 'type: "agent.upsert"' apps/server/src/.

Is it a behavior change? Yes, but an unobservable one — and it closes a latent hole

• Web's applyAgentUpsert (apps/web/src/hooks/use-sse.ts) replaces the cached agent wholesale, preserving only submittedReviewId. So the unflagged payload dropped hasStream from the cached row.
• The only consumer is agents-view.tsx:275 — focusedAgent?.hasStream ?? false — so undefined and false were already indistinguishable.
• Could the launched agent ever already have a stream? Traced it: templateService.launchTemplate always calls agentManager.createAgent, so the record is always brand new. createAgent fires onAgentCreated early (agents/manager.ts:369, which publishes a flagged upsert with hasStream: false), then runs launchWithSetupScript, then re-fetches and returns the final record — which is what this route publishes. A stream only exists once the agent's CLI drives a CDP screencast via StreamManager.startStream, which cannot happen inside that window. So there is no user-visible symptom today; the fix makes the payload correct by construction rather than correct by luck, and stops the second publish from being able to clobber the first one's flag.
• No release-notes entry: nothing user-visible changed.

Deliberately NOT changed (near-misses)

• The HTTP response body. return { agent: result.agent } stays unflagged. crud-routes.ts:333 does the same (return reply.code(201).send({ agent }) after a flagged publish); lifecycle-routes.ts:63 is the one that flags its response. That inconsistency is pre-existing and is about the REST payload, not the SSE payload this entry is about — folding it in would widen the blast radius to the launch endpoint's public response shape for no benefit.
• server.ts:417's onAgentCreated publish. It looks like this route's publish is redundant with it, but it isn't: the listener fires with the record as of insert, before media seeding, pin/prompt assembly and process launch, and the route publishes the re-fetched final record. Removing either one loses information.
• TemplateRouteDeps.publishUiEvent: (event: unknown) => void. Eight route dep types declare it as unknown and server.ts casts event as UiEvent at 13 wiring sites; only routes/reviews.ts:18 uses the precise (event: UiEvent) => void. That looseness is exactly why an unflagged agent.upsert type-checked, but tightening it touches 9 modules and belongs in its own PR. Backlogged.

What the included set has in common

Exactly one line changed at a publish site: the payload's agent value. The other 30 sites are untouched; the deps type gains one field with the same signature already used by routes/agents/shared.ts:32, routes/release.ts:247, server/mcp-handlers.ts:101, server/mcp-review-handlers.ts:97 and server/agent-lifecycle-runtime.ts:15.

Test

New apps/server/test/template-launch-publish.test.ts — bare Fastify + registerTemplateRoutes with a stubbed templateService and a spy publishUiEvent (the quick-phrases-routes.test.ts shape; the existing template-routes.test.ts boots the full app and can't observe published events). Confirmed non-vacuous: reverting the one-line fix fails the first test and leaves the second passing.

Checks

pnpm run check clean · pnpm run test 2986 passed / 9 skipped, plus 1490 and 60 in the other two suites · pnpm run test:e2e 186 passed / 12 skipped. Server-only change, so finalize:web and web vitest do not apply.

Next run

Backlog entry for AgentRecord (server) vs Agent (web) divergence — the last big server↔web wire-type gap. This PR unblocks it: hasStream is now attached at every agent.upsert publish site, so a shared agent wire type no longer has to keep it optional on that account.

🤖 Generated with Claude Code