Inference Consumer Audit — #947 cutover

Date: 2026-07-22 Question: Do all Signet inference consumers still work through the pi-ai + ACPX backends, and is low-latency inference (aggregate recall in particular) preserved?

Consumer model

Every daemon LLM call flows through one path:

consumer → getInferenceProvider(workload) | router.execute(...)
         → InferenceRouter → createRoutingProvider (inference-provider-factory.ts)
         → createPiModelProvider (pi-ai) | createAcpxProvider

There is no remaining direct use of the deleted per-provider factories. Two consumption styles, both satisfied by createPiModelProvider:

Style	Used by	pi-provider method
Streaming (`router.execute` requires `streamWithUsage`)	aggregate recall, hook recall injection, any routed call	`streamWithUsage` ✅
One-shot (`provider.generate` / `generateWithUsage`)	reranker-llm, structural workers, dreaming-worker (legacy)	`generate` + `generateWithUsage` ✅

router.execute hard-rejects providers without streamWithUsage (inference-router.ts:746). Because pi-provider implements it, every routed consumer — including aggregate recall — continues to work.

Aggregate recall (the flagged risk)

aggregate-recall.ts uses deps.router.execute(request, prompt, opts) for both query planning and synthesis. Traced end to end:

aggregateRecall → planQueries / synthesize → router.execute → createProvider → createPiModelProvider → streamSimple (pi-ai) → SSE over globalThis.fetch.

Latency measured (synthesis-sized call, ~700-token prompt, LM Studio qwen3.5-0.8b): 198–512ms warm. Well within aggregate recall’s 30s synthesis / 20s planning timeouts, and indistinguishable from direct HTTP — pi-ai is a thin layer over fetch, and the router’s ReadableStream pump adds negligible overhead.

The acpxHooks: "disabled" opt the router receives is honored: the pi-ai path never spawns a subprocess, so aggregate recall never pays harness-spawn cost.

Provider-name coupling

Searched for any consumer that pattern-matches provider.name (e.g. startsWith("anthropic:"), .split(":")[0]). None found outside the rate-limiter’s shouldRateLimit (which keys off the executor prefix and is unaffected). So the provider naming change is safe.

Legacy dreaming-worker

pipeline/dreaming-worker.ts still calls getSynthesisProvider(). This is the legacy daemon dreaming worker (the “broken third path” per #913) — enabled: false by default, gated behind !mutationsFrozen, already inert. It is slated for deletion in #913’s cutover, not a regression introduced here.

Verification

inference-router.test.ts: 5/5 pass (legacy credential resolution through pi-ai, including SSE).
pi-provider.live.test.ts: 5/5 live checks (generate, usage, stream, abort, timeout).
acpx-harness.live.test.ts: claude/codex reach ACP auth gate (chain works); opencode returns PONG end-to-end.
Two adversarial autoreview passes: no findings.

Conclusion

The cutover preserves every consumer contract. Aggregate recall works through pi-ai with no latency regression. The only behavioral changes are intentional (OpenRouter reasoning abstraction, keyless-local placeholder auth, /models availability probe) and are documented in the PR.