feat: add FOREMAN_KEEP_ALIVE config for worker model residency

Allow configuring how long the worker model stays resident on the Ollama
target after a request via FOREMAN_KEEP_ALIVE env var. Accepts Ollama
duration strings ("-1" forever, "0" unload, "15m", "1h", etc). Defaults
to "-1" (pin forever). The embedder warm-up is unaffected and always
uses keep_alive=-1.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

internal/worker/worker.go

@@ -16,6 +16,7 @@ import (
 	"errors"
 	"fmt"
 	"log/slog"
+	"strconv"
 	"sync"
 	"time"
 
@@ -104,6 +105,11 @@ type Worker struct {
 	dispatcher *webhook.Dispatcher
 	logger     *slog.Logger
 
+	// keepAlive is the JSON-encoded keep_alive value sent in outbound chat requests
+	// to control how long the worker model stays resident on the target. Derived from
+	// FOREMAN_KEEP_ALIVE config; does not affect the embedder.
+	keepAlive json.RawMessage
+
 	// wake is signaled when a new job is enqueued.
 	wake chan struct{}
 }
@@ -120,6 +126,7 @@ func New(
 	notifier *Notifier,
 	dispatcher *webhook.Dispatcher,
 	logger *slog.Logger,
+	keepAlive string,
 ) *Worker {
 	return &Worker{
 		store:      st,
@@ -128,10 +135,35 @@ func New(
 		notifier:   notifier,
 		dispatcher: dispatcher,
 		logger:     logger,
+		keepAlive:  encodeKeepAlive(keepAlive),
 		wake:       make(chan struct{}, 1),
 	}
 }
 
+// encodeKeepAlive converts a FOREMAN_KEEP_ALIVE config string to a json.RawMessage
+// suitable for the Ollama ChatRequest KeepAlive field.
+//
+// Why: Ollama's keep_alive field accepts either a JSON number (seconds, or -1 for
+// forever) or a JSON string duration ("15m", "1h"). Pure-numeric values and "-1"
+// are encoded as JSON numbers; everything else is encoded as a JSON string.
+// What: returns a json.RawMessage containing the appropriate JSON representation.
+// Test: assert "-1" -> `-1`, "0" -> `0`, "15m" -> `"15m"`, "3600" -> `3600`.
+func encodeKeepAlive(val string) json.RawMessage {
+	if val == "" {
+		val = "-1"
+	}
+
+	// If the value parses as an integer, emit it as a JSON number.
+	// This covers "-1", "0", "3600", etc.
+	if _, err := strconv.Atoi(val); err == nil {
+		return json.RawMessage(val)
+	}
+
+	// Otherwise, emit it as a JSON string (e.g. "15m", "1h").
+	b, _ := json.Marshal(val)
+	return json.RawMessage(b)
+}
+
 // Wake signals the worker that a new job may be available. Non-blocking.
 //
 // Why: the HTTP handlers signal the worker to check for new work immediately
@@ -241,6 +273,10 @@ func (w *Worker) executeJob(ctx context.Context, job store.Job) {
 	streamFalse := false
 	chatReq.Stream = &streamFalse
 
+	// Override keep_alive with the configured value so the worker model stays
+	// resident for the desired duration. The embedder is pinned separately.
+	chatReq.KeepAlive = w.keepAlive
+
 	// Execute the chat request.
 	resp, _, err := w.client.Chat(ctx, chatReq, false)
 	if err != nil {