diff --git a/CONTEXT.md b/CONTEXT.md
index ace927a..f2868d3 100644
--- a/CONTEXT.md
+++ b/CONTEXT.md
@@ -114,7 +114,7 @@ The Dashboard renders a two-level expandable tree:
 - **Repo row**: expand/collapse with `enter`/`space`. When expanded shows Branch section, then Stack rows and standalone PR rows. A Repo opens on its own when its CI starts running or failing (its branch Stoplight, or any PR it has loaded, goes 🟡 or 🔴) and closes again when it goes quiet, so a calm dashboard sits collapsed. This only fires when that state changes, so a Repo the user opened or closed by hand stays that way until its CI moves. Collapsed Repos do not fetch PR runs, so PR-only activity on a collapsed Repo does not open it.
 - **Branch section**: non-navigable; always rendered above PR rows when a Repo is expanded.
 - **Stack row**: navigable; expand/collapse with `enter`/`space` to show its member PR rows. Its Stoplight is the most actionable of its members', and `f` and `o` act on the first failing member so a collapsed Stack still exposes what broke.
-- **PR row**: navigable; expand/collapse with `enter`/`space` to show that PR's Workflow runs. A PR inside a Stack renders one level deeper and carries its `position/size`. A Review glyph follows the PR number when there is a Review signal.
+- **PR row**: navigable; expand/collapse with `enter`/`space` to show that PR's Workflow runs and each one's Jobs. A PR inside a Stack renders one level deeper and carries its `position/size`. A Review glyph follows the PR number when there is a Review signal.
 
 Expansion and the cursor stay attached to what a row shows — the Repo by name, a Stack or PR by its number — so they survive PRs opening and closing, config reloads and active-first re-sorting.
 
@@ -130,7 +130,7 @@ Each poll cycle makes the following API calls per enabled Repo:
 - 1 × `PullRequests.List`
 - 1 × GraphQL query for Stack membership and Review status, only for an expanded Repo with 1+ open PRs. A host that rejects the Stack fields gets one Review-only retry. A failure here costs only the grouping and Review glyphs — PRs still render.
 - 1 × `ListRepositoryWorkflowRuns` per open PR (head SHA runs)
-- 1 × `ListWorkflowJobs` per branch Workflow Run
+- 1 × `ListWorkflowJobs` per Workflow Run, on the branch and on each PR, so an expanded PR row lists its Jobs the way GitHub's checks do
 
 Every GET is a conditional request: the Poller keeps one client per token for
 the life of the app, and each remembers the ETag of the last response per URL.
diff --git a/internal/github/client.go b/internal/github/client.go
index 48bdd32..5e371a4 100644
--- a/internal/github/client.go
+++ b/internal/github/client.go
@@ -194,12 +194,14 @@ func New(token string) *Client {
 //   - 1 call to list open PRs
 //
 // Expanded (q.Detail true) adds the detail the dashboard actually renders:
-//   - 1 call per branch workflow run to fetch jobs
 //   - 1 call per PR for its runs (filtered by head SHA)
-//   - 1 GraphQL call for stack membership, when 2+ PRs are open
+//   - 1 call per workflow run, on the branch and on each PR, to fetch jobs
+//   - 1 GraphQL call for stack membership and review state, when a PR is open
 //
-// A repo with 4 workflows and 11 open PRs therefore costs 2 calls collapsed
-// and 17 REST plus 1 GraphQL expanded. No ListWorkflows call in either case.
+// A repo with 4 workflows, each also running on 11 open PRs, therefore costs
+// 2 calls collapsed and 2+4+11+44 = 61 REST plus 1 GraphQL expanded. The ETag
+// cache makes every one of those that is unchanged since the last cycle free,
+// so a quiet expanded repo costs close to nothing. No ListWorkflows call.
 func (c *Client) FetchAll(ctx context.Context, q RepoQuery) (RepoData, error) {
 	var stats fetchStats
 
@@ -317,17 +319,8 @@ func (c *Client) fetchBranchRuns(ctx context.Context, q RepoQuery, stats *fetchS
 		// Jobs are only rendered under an expanded repo row, and cost one call
 		// per run, so they are fetched only when that detail is on screen.
 		if q.Detail {
-			jobs, jobsResp, err := c.gh.Actions.ListWorkflowJobs(ctx, q.Owner, q.Name, run.GetID(), &github.ListWorkflowJobsOptions{})
-			stats.observe(jobsResp)
-			if err != nil {
-				return nil, "", fmt.Errorf("listing jobs for run %d in %s/%s: %w", run.GetID(), q.Owner, q.Name, err)
-			}
-			for _, j := range jobs.Jobs {
-				wr.Jobs = append(wr.Jobs, Job{
-					Name:       j.GetName(),
-					Status:     j.GetStatus(),
-					Conclusion: j.GetConclusion(),
-				})
+			if wr.Jobs, err = c.fetchJobs(ctx, q, run.GetID(), stats); err != nil {
+				return nil, "", err
 			}
 		}
 		results = append(results, wr)
@@ -361,17 +354,37 @@ func (c *Client) fetchRunsForRef(ctx context.Context, q RepoQuery, sha string, s
 			continue
 		}
 		seen[key] = true
+		jobs, err := c.fetchJobs(ctx, q, run.GetID(), stats)
+		if err != nil {
+			return nil, err
+		}
 		results = append(results, WorkflowRun{
 			WorkflowName: workflowName(run),
 			Status:       run.GetStatus(),
 			Conclusion:   run.GetConclusion(),
 			HTMLURL:      run.GetHTMLURL(),
 			RunID:        run.GetID(),
+			Jobs:         jobs,
 		})
 	}
 	return results, nil
 }
 
+// fetchJobs returns a run's jobs: one call per run, which the ETag cache makes
+// free while the run is unchanged. Only called for expanded repos.
+func (c *Client) fetchJobs(ctx context.Context, q RepoQuery, runID int64, stats *fetchStats) ([]Job, error) {
+	jobs, resp, err := c.gh.Actions.ListWorkflowJobs(ctx, q.Owner, q.Name, runID, &github.ListWorkflowJobsOptions{})
+	stats.observe(resp)
+	if err != nil {
+		return nil, fmt.Errorf("listing jobs for run %d in %s/%s: %w", runID, q.Owner, q.Name, err)
+	}
+	out := make([]Job, 0, len(jobs.Jobs))
+	for _, j := range jobs.Jobs {
+		out = append(out, Job{Name: j.GetName(), Status: j.GetStatus(), Conclusion: j.GetConclusion()})
+	}
+	return out, nil
+}
+
 // noFailedJobsFragment is what GitHub says when rerun-failed-jobs has nothing
 // to retry individually — every job either passed or never started. Re-running
 // the whole run is the right call in that case.
diff --git a/internal/github/client_test.go b/internal/github/client_test.go
index 329e275..ff485e6 100644
--- a/internal/github/client_test.go
+++ b/internal/github/client_test.go
@@ -198,8 +198,8 @@ func TestFetchAllCollapsedCostsTwoCalls(t *testing.T) {
 	}
 }
 
-// Expanding the same repo pays for the detail it now renders: one job call per
-// deduped run, one runs call per open PR.
+// Expanding the same repo pays for the detail it now renders: one runs call per
+// open PR, and one job call per deduped run, on the branch and on every PR.
 func TestFetchAllExpandedFetchesDetail(t *testing.T) {
 	var paths []string
 	c := countingServer(t, &paths)
@@ -208,9 +208,9 @@ func TestFetchAllExpandedFetchesDetail(t *testing.T) {
 	if err != nil {
 		t.Fatalf("unexpected error: %v", err)
 	}
-	// 1 branch runs + 2 job calls + 1 PR list + 3 PR run calls.
-	if len(paths) != 7 {
-		t.Fatalf("expanded poll made %d calls (%v), want 7", len(paths), paths)
+	// 1 branch runs + 2 job calls + 1 PR list + 3 PRs × (1 runs + 2 job calls).
+	if len(paths) != 13 {
+		t.Fatalf("expanded poll made %d calls (%v), want 13", len(paths), paths)
 	}
 	jobCalls := 0
 	for _, p := range paths {
@@ -218,12 +218,19 @@ func TestFetchAllExpandedFetchesDetail(t *testing.T) {
 			jobCalls++
 		}
 	}
-	if jobCalls != 2 {
-		t.Errorf("job calls = %d, want one per branch run (2)", jobCalls)
+	if jobCalls != 8 {
+		t.Errorf("job calls = %d, want one per run: 2 branch + 3 PRs × 2 (8)", jobCalls)
 	}
 	for _, r := range data.BranchRuns {
 		if len(r.Jobs) == 0 {
 			t.Errorf("expanded run %q has no jobs", r.WorkflowName)
 		}
 	}
+	for _, pr := range data.PRRuns {
+		for _, r := range pr.Runs {
+			if len(r.Jobs) == 0 {
+				t.Errorf("PR #%d run %q has no jobs", pr.PR.Number, r.WorkflowName)
+			}
+		}
+	}
 }