Show and tell: a runnable OpenRouter server-tool completion-boundary comparison

[canblmz1]

I maintain prefix-safe-json (PSJ). I built a small TanStack AI / OpenRouter comparison for applications that want original stream completion evidence before a server tool records a side effect.

The example runs real chat() server tools and the installed HTTP/SSE parsers with @tanstack/ai@0.69.0, @tanstack/ai-openrouter@0.22.0, @openrouter/sdk@0.13.20, Zod 4.6.5 and PSJ 0.5.4. The SQL-shaped input is only recorded in memory.

Authored fixture	Native effects	Guarded effects
Complete JSON and normal final usage trailer	1	1
Malformed JSON, native output limit and usage trailer	1	0
Complete JSON, native output limit and usage trailer	1	0
Complete JSON, native output limit without usage trailer	1	0
Complete JSON, normalized length termination	1	0

The optional guarded wrapper waits for a tee of the original HTTP evidence, binds the decision to the execution context's toolCallId, and consumes authority once before the counter changes. The test also checks the paired result sent to the next model request and the final streamed response. It covers one call per response.

To try it with Node 22+ and npm, download the source-only trial ZIP, extract it, and run:

node scripts/try-npm-demo.mjs tanstack

No API key, live model request, repository clone or PSJ build is needed. The launcher retains the independent npm project for editing. For a PSJ-free native baseline before adding the optional comparison, the standalone gist provides both paths.

The malformed-input fallback is already removed in pinned upstream source; I did not build or run all of upstream main. That correction is useful independently of PSJ. The complete-JSON/output-limit cases test an optional caller policy, not production frequency or a requirement for every application. This investigation also found two bugs in PSJ's own 0.5.3 OpenRouter adapter; 0.5.4 fixes them.

These are selected authored mock-transport checks. Production transport interception needs cancellation, timeouts and bounded buffering; application authorization, semantic validation and durable side effects remain caller-owned.

If you have a SQL or file tool where this boundary would help, I'd welcome a harmless complete/interrupted counter test in your own project. Reply with versions and the outcome or a concrete blocker; I'm happy to help connect the original evidence to your actual execution boundary.