feat: Ollama live drift leg via CI-provisioned daemon

.github/workflows/fix-drift.yml

@@ -62,6 +62,27 @@ jobs:
           git config user.email "drift-bot@copilotkit.ai"
           git checkout -B "fix/drift-$(date +%Y-%m-%d)-${RUN_ID}"
 
+      # Provision a local Ollama daemon so the OLLAMA_HOST-gated live drift leg
+      # (src/__tests__/drift/ollama.drift.ts) is detected here too. Ollama needs
+      # NO API key — the "credential" is just a running daemon. Pull the smallest
+      # chat+generate-capable model (qwen2:0.5b, ~350MB); OLLAMA_MODEL points the
+      # leg at it. COST: install + model pull adds ~2-4 min to the fix run.
+      - name: Provision Ollama daemon (live drift leg)
+        run: |
+          set -euo pipefail
+          # Download the installer to disk first, then execute it — avoids piping
+          # a remote, mutable script straight into a shell (no `curl | sh`).
+          curl -fsSL https://ollama.com/install.sh -o "${RUNNER_TEMP}/ollama-install.sh"
+          sh "${RUNNER_TEMP}/ollama-install.sh"
+          ollama serve > /tmp/ollama-serve.log 2>&1 &
+          for _ in $(seq 1 30); do
+            if curl -sf http://127.0.0.1:11434/api/version >/dev/null 2>&1; then
+              echo "ollama daemon ready"; break
+            fi
+            sleep 1
+          done
+          ollama pull qwen2:0.5b
+
       # Step 1: Detect drift and produce report.
       #
       # FIX #F3 (round-4) — write the report OUTSIDE the repo checkout, to
@@ -78,6 +99,9 @@ jobs:
           ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
           GOOGLE_API_KEY: ${{ secrets.GOOGLE_API_KEY }}
           OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }}
+          # Un-gates the Ollama live drift leg against the daemon provisioned above.
+          OLLAMA_HOST: 127.0.0.1:11434
+          OLLAMA_MODEL: qwen2:0.5b
           PRE_FIX_REPORT: ${{ runner.temp }}/drift-report.json
         run: |
           set +e
@@ -210,6 +234,10 @@ jobs:
           ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
           GOOGLE_API_KEY: ${{ secrets.GOOGLE_API_KEY }}
           OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }}
+          # Keep the Ollama live leg un-gated so post-fix verification exercises it
+          # (the daemon provisioned earlier in this job is still running).
+          OLLAMA_HOST: 127.0.0.1:11434
+          OLLAMA_MODEL: qwen2:0.5b
         run: pnpm test:drift
 
       # Step 4c: Re-collect drift AUTHORITATIVELY against the LIVE SDK, writing
@@ -234,6 +262,10 @@ jobs:
           ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
           GOOGLE_API_KEY: ${{ secrets.GOOGLE_API_KEY }}
           OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }}
+          # Keep the Ollama live leg un-gated so the authoritative re-collect
+          # exercises it (the daemon provisioned earlier in this job is running).
+          OLLAMA_HOST: 127.0.0.1:11434
+          OLLAMA_MODEL: qwen2:0.5b
           POST_FIX_REPORT: ${{ runner.temp }}/drift-report.post-fix.json
         run: |
           set +e

.github/workflows/test-drift.yml

@@ -148,6 +148,28 @@ jobs:
           PREFLIGHT
           npx tsx "$PREFLIGHT_FILE"
 
+      # Provision a local Ollama daemon so the OLLAMA_HOST-gated live drift leg
+      # (src/__tests__/drift/ollama.drift.ts) actually runs. Ollama needs NO API
+      # key — the "credential" is just a running daemon. We pull the smallest
+      # chat+generate-capable model (qwen2:0.5b, ~350MB) and point the leg at it
+      # via OLLAMA_MODEL to keep the added cost minimal. COST: the install + model
+      # pull adds ~2-4 min to every drift run.
+      - name: Provision Ollama daemon (live drift leg)
+        run: |
+          set -euo pipefail
+          # Download the installer to disk first, then execute it — avoids piping
+          # a remote, mutable script straight into a shell (no `curl | sh`).
+          curl -fsSL https://ollama.com/install.sh -o "${RUNNER_TEMP}/ollama-install.sh"
+          sh "${RUNNER_TEMP}/ollama-install.sh"
+          ollama serve > /tmp/ollama-serve.log 2>&1 &
+          for _ in $(seq 1 30); do
+            if curl -sf http://127.0.0.1:11434/api/version >/dev/null 2>&1; then
+              echo "ollama daemon ready"; break
+            fi
+            sleep 1
+          done
+          ollama pull qwen2:0.5b
+
       # Run the collector behind a "retry before alert" wrapper. A single
       # critical run can be a transient real-API hiccup (a streaming call
       # failing mid-flight), NOT a format change — so the wrapper re-runs the
@@ -163,6 +185,9 @@ jobs:
           ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
           GOOGLE_API_KEY: ${{ secrets.GOOGLE_API_KEY }}
           OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }}
+          # Un-gates the Ollama live drift leg against the daemon provisioned above.
+          OLLAMA_HOST: 127.0.0.1:11434
+          OLLAMA_MODEL: qwen2:0.5b
         run: |
           set +e
           npx tsx scripts/drift-retry.ts

src/__tests__/drift/ollama.drift.ts

@@ -16,7 +16,20 @@ import { httpPost, startDriftServer, stopDriftServer } from "./helpers.js";
 // Environment-based opt-in (consistent with other drift files)
 // ---------------------------------------------------------------------------
 
-const OLLAMA_HOST = process.env.OLLAMA_HOST ?? "http://localhost:11434";
+// Accept both a full base URL ("http://127.0.0.1:11434") and Ollama's native
+// host:port form ("127.0.0.1:11434" — the value the daemon itself uses). The
+// test issues real HTTP requests, so a scheme is required; prefix http:// when
+// one is absent.
+const RAW_OLLAMA_HOST = process.env.OLLAMA_HOST ?? "http://localhost:11434";
+const OLLAMA_HOST = /^https?:\/\//.test(RAW_OLLAMA_HOST)
+  ? RAW_OLLAMA_HOST
+  : `http://${RAW_OLLAMA_HOST}`;
+
+// The model to exercise against the live daemon. Defaults to "llama3.2" for a
+// local developer run, but CI provisions a much smaller model (to keep the
+// daemon pull cheap) and points the leg at it via OLLAMA_MODEL. Both /api/chat
+// and /api/generate use the same model.
+const OLLAMA_MODEL = process.env.OLLAMA_MODEL ?? "llama3.2";
 
 // ---------------------------------------------------------------------------
 // Server lifecycle
@@ -112,7 +125,7 @@ describe.skipIf(!process.env.OLLAMA_HOST)("Ollama drift", () => {
     const sdkShape = ollamaChatResponseShape();
 
     const body = {
-      model: "llama3.2",
+      model: OLLAMA_MODEL,
       messages: [{ role: "user", content: "Say hello" }],
       stream: false,
     };
@@ -143,7 +156,7 @@ describe.skipIf(!process.env.OLLAMA_HOST)("Ollama drift", () => {
     const sdkChunkShape = ollamaChatStreamChunkShape();
 
     const body = {
-      model: "llama3.2",
+      model: OLLAMA_MODEL,
       messages: [{ role: "user", content: "Say hello" }],
       stream: true,
     };
@@ -181,7 +194,7 @@ describe.skipIf(!process.env.OLLAMA_HOST)("Ollama drift", () => {
     const sdkShape = ollamaGenerateResponseShape();
 
     const body = {
-      model: "llama3.2",
+      model: OLLAMA_MODEL,
       prompt: "Say hello",
       stream: false,
     };