Wire Peregrine to cf-orch coordinator (CF_ORCH_URL) for stress test readiness #106

New issue

Closed

opened 2026-04-18 22:38:40 -07:00 by pyr0ball · 0 comments

pyr0ball commented

2026-04-18 22:38:40 -07:00

Owner

Context

The go-live stress test harness (circuitforge-infra#12) requires all three public apps to be wired to cf-orch via CF_ORCH_URL. Kiwi is already wired. Peregrine is not.

Until this is done, the stress test coordinator sidecar will show zero allocations for Peregrine endpoints, and the test results will be misleading.

What needs to happen

Identify all inference call sites in Peregrine (cover letter gen, recipe suggest, enrich, etc.)
Replace direct ollama/vllm calls with CFOrchClient.allocate() (or async variant)
Add CF_ORCH_URL to compose.cloud.yml environment block
Confirm GET /api/analytics/summary shows allocations during a test run

Acceptance Criteria

CF_ORCH_URL set in compose.cloud.yml
All LLM inference calls route through CFOrchClient
Coordinator dashboard (:7700) shows active leases from Peregrine under load
Stress test sidecar CSV shows non-zero allocation count for Peregrine endpoints

Reference

Design spec: circuitforge-plans/shared/superpowers/specs/2026-04-18-go-live-stress-test-design.md
Harness PR: Circuit-Forge/circuitforge-infra#12

## Context The go-live stress test harness (circuitforge-infra#12) requires all three public apps to be wired to cf-orch via `CF_ORCH_URL`. Kiwi is already wired. Peregrine is not. Until this is done, the stress test coordinator sidecar will show zero allocations for Peregrine endpoints, and the test results will be misleading. ## What needs to happen 1. Identify all inference call sites in Peregrine (cover letter gen, recipe suggest, enrich, etc.) 2. Replace direct ollama/vllm calls with `CFOrchClient.allocate()` (or async variant) 3. Add `CF_ORCH_URL` to `compose.cloud.yml` environment block 4. Confirm `GET /api/analytics/summary` shows allocations during a test run ## Acceptance Criteria - [ ] `CF_ORCH_URL` set in `compose.cloud.yml` - [ ] All LLM inference calls route through `CFOrchClient` - [ ] Coordinator dashboard (:7700) shows active leases from Peregrine under load - [ ] Stress test sidecar CSV shows non-zero allocation count for Peregrine endpoints ## Reference Design spec: `circuitforge-plans/shared/superpowers/specs/2026-04-18-go-live-stress-test-design.md` Harness PR: Circuit-Forge/circuitforge-infra#12