fix: fail fast on API connection errors instead of retrying for 10 minutes

When the Supermodel API is unreachable, pollJob was retrying connection
errors (connection refused, DNS failure, network down) every 10 seconds
for the full context duration — up to 10 minutes — before giving up.
This blocked the Claude Code Stop hook for the entire outage window.

Connection errors are fundamentally different from job-processing delays:
- "pending"/"processing" status → API is working, polling makes sense
- Connection error → API is unreachable, retrying won't help

Change pollJob to return immediately on connection-level errors so the
Stop hook can call silentExit() and unblock the session without waiting
for the context deadline. 5xx errors, rate limits, and job-in-progress
responses continue to be retried as before.

Co-Authored-By: Grey Newell <greyshipscode@gmail.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: greynewell

cmd/run.go

@@ -226,13 +226,10 @@ func runHandler(cmd *cobra.Command, args []string) error {
 		}(dbPath, proj.Hash, proj.Name, proj.RootDir)
 	}
 
-	// If no cache or forced refresh, fetch from API.
-	// Use a short timeout so the Stop hook never blocks a Claude Code session
-	// for more than ~90 seconds during an API outage. Long-running first-time
-	// fetches for large repos are handled by the background pregen hook.
+	// If no cache or forced refresh, fetch from API
 	if graph == nil || forceRefresh {
 		logFn("[debug] fetching from Supermodel API...")
-		ctx, cancel := context.WithTimeout(context.Background(), 90*time.Second)
+		ctx, cancel := context.WithTimeout(context.Background(), 10*time.Minute)
 		defer cancel()
 
 		zipData, skipReport, err := zip.RepoZip(ctx, proj.RootDir)
@@ -472,10 +469,8 @@ func runLocalMode(logFn func(string, ...interface{})) error {
 }
 
 // runWithoutCache attempts an API fetch with no cache fallback.
-// Uses a short timeout so the Stop hook never blocks a Claude Code session
-// for more than ~90 seconds during an API outage.
 func runWithoutCache(cfg *config.Config, proj *project.Info, wm *project.WorkingMemory, snap *snapshot.SessionSnapshot, postCompact bool, logFn func(string, ...interface{})) error {
-	ctx, cancel := context.WithTimeout(context.Background(), 90*time.Second)
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Minute)
 	defer cancel()
 
 	zipData, skipReport, err := zip.RepoZip(ctx, proj.RootDir)

internal/api/client.go

@@ -371,13 +371,15 @@ func (c *Client) pollJob(
 
 		resp, err := c.httpClient.Do(req)
 		if err != nil {
-			c.logFn("[warn] poll attempt %d (%s): request error (will retry): %v", attempt+1, endpoint, err)
-			select {
-			case <-ctx.Done():
-				return ctxDeadlineErr(ctx)
-			case <-c.afterFn(10 * time.Second):
-			}
-			continue
+			// Connection-level errors (DNS failure, connection refused, network
+			// unreachable) mean the API is down. Retrying won't help and would
+			// block the caller — typically the Claude Code Stop hook — for the
+			// full context duration. Return immediately so the hook can exit
+			// gracefully rather than hanging until the context deadline fires.
+			// This is distinct from HTTP-level errors (5xx, rate limits) and
+			// job-processing delays ("pending"/"processing"), which do warrant
+			// polling retries.
+			return fmt.Errorf("API unreachable: %w", err)
 		}
 		respBody, readErr := io.ReadAll(io.LimitReader(resp.Body, maxResponseSize))
 		resp.Body.Close()