fix(runner): a guard-skipped step is no longer a silent log line

The strict echo-skip left only a logger.Warn, so a step the guard discarded was
indistinguishable in the trace from a picker that legitimately declined. Any
yield or actions-per-hour figure computed from model traces mixed the two.
Every path that ends a step without a model-chosen action now records its own
outcome.

Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX

internal/runner/llm_source.go

@@ -13,6 +13,7 @@ import (
 	"log/slog"
 	"regexp"
 	"strings"
+	"time"
 
 	"github.com/priyanshujain/sanderling/internal/llmclient"
 	"github.com/priyanshujain/sanderling/internal/trace"
@@ -51,6 +52,9 @@ type llmSource struct {
 	instructions string
 	logger       *slog.Logger
 	history      *actionHistory
+	// recorder persists one record per step: what was sent, what came back, and
+	// how the step ended. Nil only in unit tests that never select.
+	recorder llmCallRecorder
 
 	// lastSource/lastReasoning describe the most recent NextAction so the runner
 	// can stamp the trace. lastSource is "llm" only when the LLM (not setup)
@@ -71,10 +75,18 @@ type llmSelection struct {
 	chosenAction string
 }
 
+// llmCallRecorder persists one selection record per step. *trace.Writer
+// implements it.
+type llmCallRecorder interface {
+	WriteLLMCall(call trace.LLMCall) error
+}
+
 // NextAction returns the step's action. Setup precedence is preserved by
 // running the JS path first (the llm marker is inert there, so a null result
-// means setup yielded nothing); the LLM selection then takes over.
-func (s *llmSource) NextAction(ctx context.Context) (verifier.Action, error) {
+// means setup yielded nothing); the LLM selection then takes over. Every way a
+// step can end lands one record keyed to stepIndex, so a step the guards threw
+// away is never confused with one the picker declined.
+func (s *llmSource) NextAction(ctx context.Context, stepIndex int) (verifier.Action, error) {
 	s.lastSource = ""
 	s.lastReasoning = ""
 	s.lastChoice = 0
@@ -85,17 +97,20 @@ func (s *llmSource) NextAction(ctx context.Context) (verifier.Action, error) {
 	// setup (e.g. login) first but never the weighted picker.
 	action, err := s.verifier.SetupAction()
 	if err == nil {
+		s.record(stepIndex, trace.LLMCall{Outcome: trace.LLMOutcomeSetupAction})
 		s.history.add(describeAction(action))
 		return action, nil
 	}
 	if !errors.Is(err, verifier.ErrNoAction) {
+		s.record(stepIndex, trace.LLMCall{Outcome: trace.LLMOutcomeSetupFailed, Error: err.Error()})
 		return verifier.Action{}, err
 	}
 
-	selection, ok := s.selectViaLLM(ctx)
-	if !ok {
-		// Any failure (HTTP error, unusable output, invalid choice, echo
-		// mismatch) skips the step; the next step re-observes and tries again.
+	selection, call := s.selectViaLLM(ctx)
+	s.record(stepIndex, call)
+	if call.Outcome != trace.LLMOutcomeSelected {
+		// Every other outcome skips the step; the next step re-observes and
+		// tries again. The record says which one it was.
 		return verifier.Action{}, verifier.ErrNoAction
 	}
 	s.lastSource = "llm"
@@ -107,33 +122,57 @@ func (s *llmSource) NextAction(ctx context.Context) (verifier.Action, error) {
 }
 
 // selectViaLLM runs one multimodal call and maps the chosen number to an action.
-// It returns ok=false on any error/empty/invalid output, logging the cause; the
-// caller turns that into a skipped step.
-func (s *llmSource) selectViaLLM(ctx context.Context) (llmSelection, bool) {
+// The returned record is complete whichever way the selection ended: its
+// Outcome is trace.LLMOutcomeSelected exactly when the returned selection is
+// usable.
+func (s *llmSource) selectViaLLM(ctx context.Context) (llmSelection, trace.LLMCall) {
+	call := trace.LLMCall{Timestamp: time.Now(), Model: s.model}
 	candidates := s.verifier.Candidates()
 	if len(candidates) == 0 {
-		return llmSelection{}, false
+		call.Outcome = trace.LLMOutcomeNoCandidates
+		return llmSelection{}, call
 	}
+	call.Candidates = recordCandidates(candidates)
 
-	response, err := s.client.ChatCompletion(ctx, s.buildRequest(candidates))
+	request, screenshotReference := s.buildRequest(candidates)
+	call.Screenshot = screenshotReference
+	call.SystemPrompt, call.UserPrompt = promptTexts(request)
+
+	requestStart := time.Now()
+	response, err := s.client.ChatCompletion(ctx, request)
+	call.LatencyMillis = time.Since(requestStart).Milliseconds()
 	if err != nil {
 		s.logger.Warn("llm action selection failed", "err", err)
-		return llmSelection{}, false
+		call.Outcome = trace.LLMOutcomeRequestFailed
+		call.Error = err.Error()
+		return llmSelection{}, call
 	}
+	call.ServedModel = response.Model
+	call.PromptTokens = response.Usage.PromptTokens
+	call.CompletionTokens = response.Usage.CompletionTokens
+	call.TotalTokens = response.Usage.TotalTokens
 	if len(response.Choices) == 0 {
 		s.logger.Warn("llm returned no choices")
-		return llmSelection{}, false
+		call.Outcome = trace.LLMOutcomeNoChoices
+		return llmSelection{}, call
 	}
+	call.RawResponse = response.Choices[0].Message.Content
 
-	output, err := parseChoice(response.Choices[0].Message.Content)
+	output, err := parseChoice(call.RawResponse)
 	if err != nil {
 		s.logger.Warn("llm output unusable", "err", err)
-		return llmSelection{}, false
+		call.Outcome = trace.LLMOutcomeUnparsableResponse
+		call.Error = err.Error()
+		return llmSelection{}, call
 	}
+	call.Choice = output.Choice
+	call.EchoedAction = output.ChosenAction
+	call.Reasoning = output.Reasoning
 	// choice is 1-based into the numbered list.
 	if output.Choice < 1 || output.Choice > len(candidates) {
 		s.logger.Warn("llm choice out of range", "choice", output.Choice, "candidates", len(candidates))
-		return llmSelection{}, false
+		call.Outcome = trace.LLMOutcomeChoiceOutOfRange
+		return llmSelection{}, call
 	}
 	candidate := candidates[output.Choice-1]
 	// Strict skip: the echoed action must match the numbered entry, so a model
@@ -143,19 +182,79 @@ func (s *llmSource) selectViaLLM(ctx context.Context) (llmSelection, bool) {
 	if stripWeightSuffix(output.ChosenAction) != candidate.Description {
 		s.logger.Warn("llm chosen_action mismatch; skipping",
 			"choice", output.Choice, "echoed", output.ChosenAction, "candidate", candidate.Description)
-		return llmSelection{}, false
+		call.Outcome = trace.LLMOutcomeEchoMismatch
+		return llmSelection{}, call
 	}
 	action, err := s.actionForCandidate(candidate, output.Text)
 	if err != nil {
 		s.logger.Warn("building action from candidate failed", "choice", output.Choice, "err", err)
-		return llmSelection{}, false
+		call.Outcome = trace.LLMOutcomeActionBuildFailed
+		call.Error = err.Error()
+		return llmSelection{}, call
 	}
+	call.Outcome = trace.LLMOutcomeSelected
 	return llmSelection{
 		action:       action,
 		reasoning:    output.Reasoning,
 		choice:       output.Choice,
 		chosenAction: candidate.Description,
-	}, true
+	}, call
+}
+
+// record stamps the step this selection belongs to and appends the record. A
+// write failure must not kill the run, but it does mean the step is
+// unattributable, so it is logged.
+func (s *llmSource) record(stepIndex int, call trace.LLMCall) {
+	if s.recorder == nil {
+		return
+	}
+	call.Step = stepIndex
+	if call.Timestamp.IsZero() {
+		call.Timestamp = time.Now()
+	}
+	if err := s.recorder.WriteLLMCall(call); err != nil {
+		s.logger.Warn("llm call record failed", "step", stepIndex, "err", err)
+	}
+}
+
+// recordCandidates snapshots the numbered list as the prompt rendered it, so a
+// campaign that varies candidate labelling can recover the labels each call saw.
+func recordCandidates(candidates []verifier.ActionCandidate) []trace.LLMCandidate {
+	recorded := make([]trace.LLMCandidate, 0, len(candidates))
+	for _, candidate := range candidates {
+		entry := trace.LLMCandidate{
+			Index:       candidate.Index,
+			Kind:        string(candidate.Kind),
+			Description: candidate.Description,
+			Label:       candidate.Label,
+		}
+		if candidate.Weighted {
+			entry.Weight = candidate.Weight
+		}
+		recorded = append(recorded, entry)
+	}
+	return recorded
+}
+
+// promptTexts reads the text parts back off the built request, so the record is
+// what went on the wire rather than a second rendering of the same templates.
+func promptTexts(request llmclient.Request) (systemPrompt, userPrompt string) {
+	for _, message := range request.Messages {
+		var parts []string
+		for _, part := range message.Content {
+			if part.Type == "text" {
+				parts = append(parts, part.Text)
+			}
+		}
+		text := strings.Join(parts, "\n")
+		switch message.Role {
+		case "system":
+			systemPrompt = text
+		case "user":
+			userPrompt = text
+		}
+	}
+	return systemPrompt, userPrompt
 }
 
 // actionForCandidate turns a chosen candidate into the executable action. The
@@ -182,14 +281,23 @@ func (s *llmSource) actionForCandidate(candidate verifier.ActionCandidate, text
 // buildRequest assembles the one-shot multimodal request: a system frame, the
 // numbered candidate list plus recent-action memory, and the downscaled
 // screenshot. The strict json_schema response format pins the ranked output.
-func (s *llmSource) buildRequest(candidates []verifier.ActionCandidate) llmclient.Request {
+//
+// The second return is the run-relative path of the screenshot attached, empty
+// when none was, so a record can never claim an image the call did not carry.
+// It names the step the image was captured at, which is the last observed step
+// rather than the current one whenever an observation was skipped.
+func (s *llmSource) buildRequest(candidates []verifier.ActionCandidate) (llmclient.Request, string) {
 	userParts := []llmclient.ContentPart{llmclient.TextPart(s.userPrompt(candidates))}
+	screenshotReference := ""
 	if screenshot := s.verifier.Screenshot(); len(screenshot) > 0 {
 		if dataURL, ok := screenshotDataURL(screenshot, llmMaxImageEdge); ok {
 			userParts = append(userParts, llmclient.ImagePart(dataURL))
+			if step := s.verifier.SnapshotStep(); step > 0 {
+				screenshotReference = trace.ScreenshotReference(step)
+			}
 		}
 	}
-	return llmclient.Request{
+	request := llmclient.Request{
 		Model: s.model,
 		Messages: []llmclient.Message{
 			{Role: "system", Content: []llmclient.ContentPart{llmclient.TextPart(s.systemPrompt())}},
@@ -197,6 +305,7 @@ func (s *llmSource) buildRequest(candidates []verifier.ActionCandidate) llmclien
 		},
 		ResponseFormat: choiceResponseFormat(len(candidates)),
 	}
+	return request, screenshotReference
 }
 
 // systemPrompt is the base framing plus any spec-level instructions, appended as

internal/runner/llm_source_test.go

@@ -5,21 +5,27 @@ import (
 	"context"
 	"encoding/json"
 	"errors"
+	"fmt"
 	"image"
 	"image/png"
 	"io"
 	"log/slog"
 	"net/http"
 	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"regexp"
 	"slices"
 	"strconv"
 	"strings"
 	"testing"
+	"time"
 
 	"github.com/priyanshujain/sanderling/internal/driver"
 	mockdriver "github.com/priyanshujain/sanderling/internal/driver/mock"
 	"github.com/priyanshujain/sanderling/internal/hierarchy"
 	"github.com/priyanshujain/sanderling/internal/llmclient"
+	"github.com/priyanshujain/sanderling/internal/trace"
 	"github.com/priyanshujain/sanderling/internal/verifier"
 )
 
@@ -168,6 +174,9 @@ type fakeOpenRouter struct {
 	chosenAction string
 	text         string
 	reasoning    string
+	usage        llmclient.Usage
+	servedModel  string
+	delay        time.Duration
 	lastRequest  map[string]any
 }
 
@@ -184,8 +193,11 @@ func newFakeOpenRouter(t *testing.T) *fakeOpenRouter {
 			"text":          fake.text,
 		})
 		response, _ := json.Marshal(llmclient.Response{
+			Model:   fake.servedModel,
 			Choices: []llmclient.Choice{{Message: llmclient.ResponseMessage{Content: string(content)}}},
+			Usage:   fake.usage,
 		})
+		time.Sleep(fake.delay)
 		w.Header().Set("Content-Type", "application/json")
 		_, _ = w.Write(response)
 	}))
@@ -193,7 +205,41 @@ func newFakeOpenRouter(t *testing.T) *fakeOpenRouter {
 	return fake
 }
 
+// fakeCallRecorder keeps the per-step selection records in memory so a test can
+// assert on them without opening the run directory.
+type fakeCallRecorder struct {
+	calls []trace.LLMCall
+}
+
+func (r *fakeCallRecorder) WriteLLMCall(call trace.LLMCall) error {
+	r.calls = append(r.calls, call)
+	return nil
+}
+
+func recordedCalls(t *testing.T, source *llmSource) []trace.LLMCall {
+	t.Helper()
+	recorder, ok := source.recorder.(*fakeCallRecorder)
+	if !ok {
+		t.Fatalf("recorder = %T, want *fakeCallRecorder", source.recorder)
+	}
+	return recorder.calls
+}
+
+func lastCall(t *testing.T, source *llmSource) trace.LLMCall {
+	t.Helper()
+	calls := recordedCalls(t, source)
+	if len(calls) == 0 {
+		t.Fatal("no selection record written")
+	}
+	return calls[len(calls)-1]
+}
+
 func newLLMSource(t *testing.T, fake *fakeOpenRouter) (*llmSource, *verifier.Verifier) {
+	t.Helper()
+	return newLLMSourceWithSpec(t, fake, llmFixtureSpec)
+}
+
+func newLLMSourceWithSpec(t *testing.T, fake *fakeOpenRouter, spec string) (*llmSource, *verifier.Verifier) {
 	t.Helper()
 	t.Setenv("OPENROUTER_API_KEY", "test-key")
 	t.Setenv("OPENROUTER_BASE_URL", fake.server.URL)
@@ -206,7 +252,7 @@ func newLLMSource(t *testing.T, fake *fakeOpenRouter) (*llmSource, *verifier.Ver
 	if err != nil {
 		t.Fatal(err)
 	}
-	if err := verifierInstance.Load(bundleSpec(t, llmFixtureSpec)); err != nil {
+	if err := verifierInstance.Load(bundleSpec(t, spec)); err != nil {
 		t.Fatal(err)
 	}
 	if _, ok := verifierInstance.LLMConfig(); !ok {
@@ -219,21 +265,48 @@ func newLLMSource(t *testing.T, fake *fakeOpenRouter) (*llmSource, *verifier.Ver
 		model:    "test/model",
 		logger:   slog.New(slog.NewTextHandler(io.Discard, nil)),
 		history:  newActionHistory(llmHistorySize),
+		recorder: &fakeCallRecorder{},
 	}
 	return source, verifierInstance
 }
 
 func pushLLMSnapshot(t *testing.T, v *verifier.Verifier) {
+	t.Helper()
+	pushLLMSnapshotAtStep(t, v, 0)
+}
+
+func pushLLMSnapshotAtStep(t *testing.T, v *verifier.Verifier, stepIndex int) {
 	t.Helper()
 	tree, err := hierarchy.Parse(llmTreeJSON)
 	if err != nil {
 		t.Fatal(err)
 	}
-	if err := v.PushSnapshot(verifier.SnapshotInput{Tree: tree, ScreenshotPNG: tinyPNG(t)}); err != nil {
+	if err := v.PushSnapshot(verifier.SnapshotInput{
+		Tree:          tree,
+		ScreenshotPNG: tinyPNG(t),
+		StepIndex:     stepIndex,
+	}); err != nil {
 		t.Fatalf("PushSnapshot: %v", err)
 	}
 }
 
+func readLLMCalls(t *testing.T, directory string) []trace.LLMCall {
+	t.Helper()
+	body, err := os.ReadFile(filepath.Join(directory, trace.LLMCallFileName))
+	if err != nil {
+		t.Fatalf("read %s: %v", trace.LLMCallFileName, err)
+	}
+	var calls []trace.LLMCall
+	for _, line := range strings.Split(strings.TrimSpace(string(body)), "\n") {
+		var call trace.LLMCall
+		if err := json.Unmarshal([]byte(line), &call); err != nil {
+			t.Fatalf("decode record %q: %v", line, err)
+		}
+		calls = append(calls, call)
+	}
+	return calls
+}
+
 func candidateByKind(t *testing.T, candidates []verifier.ActionCandidate, kind verifier.ActionKind) verifier.ActionCandidate {
 	t.Helper()
 	for _, candidate := range candidates {
@@ -385,7 +458,7 @@ func TestLLMSourceDrivesExecutedActions(t *testing.T) {
 	fake.chosenAction = tap.Description
 	fake.reasoning = "tap submit"
 	fake.text = ""
-	action, err := source.NextAction(context.Background())
+	action, err := source.NextAction(context.Background(), 1)
 	if err != nil {
 		t.Fatalf("NextAction: %v", err)
 	}
@@ -411,7 +484,7 @@ func TestLLMSourceDrivesExecutedActions(t *testing.T) {
 	fake.chosenAction = typing.Description
 	fake.reasoning = "type a name"
 	fake.text = "Priya"
-	action, err = source.NextAction(context.Background())
+	action, err = source.NextAction(context.Background(), 1)
 	if err != nil {
 		t.Fatalf("NextAction: %v", err)
 	}
@@ -440,7 +513,7 @@ func TestLLMSourceSkipsOnOutOfRangeChoice(t *testing.T) {
 
 	fake.choice = 9999
 	fake.chosenAction = "whatever"
-	_, err := source.NextAction(context.Background())
+	_, err := source.NextAction(context.Background(), 1)
 	if !errors.Is(err, verifier.ErrNoAction) {
 		t.Fatalf("NextAction err = %v, want ErrNoAction for an out-of-range choice", err)
 	}
@@ -460,7 +533,7 @@ func TestLLMSourceAcceptsEchoWithWeightSuffix(t *testing.T) {
 	tap := candidateByKind(t, candidates, verifier.ActionKindTap)
 	fake.choice = tap.Index
 	fake.chosenAction = tap.Description + "  (w" + strconv.Itoa(tap.Weight) + ")"
-	action, err := source.NextAction(context.Background())
+	action, err := source.NextAction(context.Background(), 1)
 	if err != nil {
 		t.Fatalf("NextAction: %v", err)
 	}
@@ -497,7 +570,7 @@ func TestLLMSourceStrictSkipsOnEchoMismatch(t *testing.T) {
 	tap := candidateByKind(t, candidates, verifier.ActionKindTap)
 	fake.choice = tap.Index
 	fake.chosenAction = "Tap \"Something Else\""
-	_, err := source.NextAction(context.Background())
+	_, err := source.NextAction(context.Background(), 1)
 	if !errors.Is(err, verifier.ErrNoAction) {
 		t.Fatalf("NextAction err = %v, want ErrNoAction on chosen_action mismatch", err)
 	}
@@ -516,12 +589,247 @@ func TestLLMSourceSkipsOnHTTPError(t *testing.T) {
 	pushLLMSnapshot(t, verifierInstance)
 
 	fake.choice = 1
-	_, err := source.NextAction(context.Background())
+	_, err := source.NextAction(context.Background(), 1)
 	if !errors.Is(err, verifier.ErrNoAction) {
 		t.Fatalf("NextAction err = %v, want ErrNoAction on HTTP failure", err)
 	}
 }
 
+// llmSetupFixtureSpec drives the first action from setup, so the model is never
+// consulted for that step.
+const llmSetupFixtureSpec = `
+import { llm, always, actions, taps, typing, weighted, Tap } from "@sanderling/spec";
+globalThis.properties = { ok: always(() => true) };
+globalThis.setup = actions(() => [Tap({ on: "id:Submit" })]);
+globalThis.actions = weighted([1, taps], [1, typing]);
+globalThis.generator = llm({ model: "test/model" });
+`
+
+// TestLLMCallRecordSeparatesGuardSkipFromDecline pins the reason these records
+// exist. A step the echo guard threw away and a step where the picker had
+// nothing to choose both used to leave nothing behind but a log line, so any
+// defect-yield or actions-per-hour figure computed from a model run silently
+// mixed the two with each other and with steps that acted.
+func TestLLMCallRecordSeparatesGuardSkipFromDecline(t *testing.T) {
+	const stepIndex = 7
+
+	guardSkipped := func(t *testing.T) trace.LLMCall {
+		fake := newFakeOpenRouter(t)
+		source, verifierInstance := newLLMSource(t, fake)
+		pushLLMSnapshot(t, verifierInstance)
+		tap := candidateByKind(t, verifierInstance.Candidates(), verifier.ActionKindTap)
+		fake.choice = tap.Index
+		fake.chosenAction = `Tap "Something Else"`
+		if _, err := source.NextAction(context.Background(), stepIndex); !errors.Is(err, verifier.ErrNoAction) {
+			t.Fatalf("NextAction err = %v, want ErrNoAction on echo mismatch", err)
+		}
+		return lastCall(t, source)
+	}
+	declined := func(t *testing.T) trace.LLMCall {
+		fake := newFakeOpenRouter(t)
+		// No snapshot pushed, so the action tree yields nothing to pick from.
+		source, _ := newLLMSource(t, fake)
+		if _, err := source.NextAction(context.Background(), stepIndex); !errors.Is(err, verifier.ErrNoAction) {
+			t.Fatalf("NextAction err = %v, want ErrNoAction with no candidates", err)
+		}
+		return lastCall(t, source)
+	}
+	acted := func(t *testing.T) trace.LLMCall {
+		fake := newFakeOpenRouter(t)
+		source, verifierInstance := newLLMSource(t, fake)
+		pushLLMSnapshot(t, verifierInstance)
+		tap := candidateByKind(t, verifierInstance.Candidates(), verifier.ActionKindTap)
+		fake.choice = tap.Index
+		fake.chosenAction = tap.Description
+		if _, err := source.NextAction(context.Background(), stepIndex); err != nil {
+			t.Fatalf("NextAction: %v", err)
+		}
+		return lastCall(t, source)
+	}
+
+	skip, decline, pick := guardSkipped(t), declined(t), acted(t)
+	if skip.Outcome != trace.LLMOutcomeEchoMismatch {
+		t.Errorf("guard-skipped outcome = %q, want %q", skip.Outcome, trace.LLMOutcomeEchoMismatch)
+	}
+	if decline.Outcome != trace.LLMOutcomeNoCandidates {
+		t.Errorf("declined outcome = %q, want %q", decline.Outcome, trace.LLMOutcomeNoCandidates)
+	}
+	if pick.Outcome != trace.LLMOutcomeSelected {
+		t.Errorf("executed outcome = %q, want %q", pick.Outcome, trace.LLMOutcomeSelected)
+	}
+	for _, call := range []trace.LLMCall{skip, decline, pick} {
+		if call.Step != stepIndex {
+			t.Errorf("record step = %d, want %d so it joins its trace line", call.Step, stepIndex)
+		}
+		if call.Timestamp.IsZero() {
+			t.Error("record carries no timestamp")
+		}
+	}
+	// The guard skip must carry what only the dropped log line used to hold.
+	if skip.Choice == 0 || skip.EchoedAction != `Tap "Something Else"` || skip.RawResponse == "" {
+		t.Errorf("guard-skip record = %+v, want the choice, the echo, and the raw response", skip)
+	}
+	if len(skip.Candidates) == 0 {
+		t.Error("guard-skip record must keep the candidate list the mismatch is judged against")
+	}
+	// A decline never reached the provider, so it must not look like a call.
+	if decline.RawResponse != "" || decline.UserPrompt != "" || len(decline.Candidates) != 0 {
+		t.Errorf("decline record = %+v, want no prompt, candidates, or response", decline)
+	}
+}
+
+// TestLLMCallRecordsCandidateListAsShown covers the companion experiment that
+// varies how candidates are labelled: the labels are an independent variable, so
+// each call must be recoverable with the exact numbered list it saw.
+func TestLLMCallRecordsCandidateListAsShown(t *testing.T) {
+	fake := newFakeOpenRouter(t)
+	source, verifierInstance := newLLMSource(t, fake)
+	source.instructions = "hunt for double submits"
+	pushLLMSnapshot(t, verifierInstance)
+	tap := candidateByKind(t, verifierInstance.Candidates(), verifier.ActionKindTap)
+	fake.choice = tap.Index
+	fake.chosenAction = tap.Description
+	if _, err := source.NextAction(context.Background(), 1); err != nil {
+		t.Fatalf("NextAction: %v", err)
+	}
+
+	call := lastCall(t, source)
+	if len(call.Candidates) == 0 {
+		t.Fatal("no candidate list recorded")
+	}
+	for _, candidate := range call.Candidates {
+		line := fmt.Sprintf("%d. %s", candidate.Index, candidate.Description)
+		if !strings.Contains(call.UserPrompt, line) {
+			t.Errorf("recorded candidate %q is not a line of the prompt:\n%s", line, call.UserPrompt)
+		}
+		if candidate.Weight > 0 && !strings.Contains(call.UserPrompt, fmt.Sprintf("%s  (w%d)", candidate.Description, candidate.Weight)) {
+			t.Errorf("recorded weight %d for %q is not the weight the prompt showed:\n%s",
+				candidate.Weight, candidate.Description, call.UserPrompt)
+		}
+	}
+	numberedLine := regexp.MustCompile(`(?m)^\d+\. `)
+	if shown := len(numberedLine.FindAllString(call.UserPrompt, -1)); shown != len(call.Candidates) {
+		t.Errorf("prompt showed %d numbered lines but %d candidates were recorded", shown, len(call.Candidates))
+	}
+	if got := candidateLabels(call.Candidates); !slices.Contains(got, tap.Label) {
+		t.Errorf("recorded labels = %v, want the target label %q among them", got, tap.Label)
+	}
+	// The system prompt is assembled from a constant plus spec instructions, so
+	// the record has to be the assembled text, not the constant.
+	if call.SystemPrompt != source.systemPrompt() {
+		t.Errorf("recorded system prompt = %q, want the assembled prompt", call.SystemPrompt)
+	}
+	if !strings.Contains(call.SystemPrompt, "hunt for double submits") {
+		t.Error("recorded system prompt dropped the spec instructions")
+	}
+	if call.Model != "test/model" {
+		t.Errorf("recorded model = %q, want test/model", call.Model)
+	}
+}
+
+func candidateLabels(candidates []trace.LLMCandidate) []string {
+	labels := make([]string, 0, len(candidates))
+	for _, candidate := range candidates {
+		labels = append(labels, candidate.Label)
+	}
+	return labels
+}
+
+func TestLLMCallRecordsSetupDrivenStep(t *testing.T) {
+	fake := newFakeOpenRouter(t)
+	source, verifierInstance := newLLMSourceWithSpec(t, fake, llmSetupFixtureSpec)
+	pushLLMSnapshot(t, verifierInstance)
+
+	action, err := source.NextAction(context.Background(), 2)
+	if err != nil {
+		t.Fatalf("NextAction: %v", err)
+	}
+	if action.On != "id:Submit" {
+		t.Fatalf("action = %+v, want the setup tap on id:Submit", action)
+	}
+	call := lastCall(t, source)
+	if call.Outcome != trace.LLMOutcomeSetupAction {
+		t.Errorf("outcome = %q, want %q", call.Outcome, trace.LLMOutcomeSetupAction)
+	}
+	if call.UserPrompt != "" || call.RawResponse != "" || call.TotalTokens != 0 {
+		t.Errorf("setup-driven step = %+v, want no model call recorded", call)
+	}
+	if source.lastSource != "" {
+		t.Errorf("lastSource = %q, want empty: setup chose the action, not the model", source.lastSource)
+	}
+}
+
+func TestLLMCallFileRecordsUsageLatencyAndScreenshot(t *testing.T) {
+	fake := newFakeOpenRouter(t)
+	fake.usage = llmclient.Usage{PromptTokens: 1200, CompletionTokens: 34, TotalTokens: 1234}
+	fake.servedModel = "vendor/model-2026-05"
+	fake.delay = 15 * time.Millisecond
+	source, verifierInstance := newLLMSource(t, fake)
+
+	directory := t.TempDir()
+	writer, err := trace.NewWriter(directory)
+	if err != nil {
+		t.Fatal(err)
+	}
+	source.recorder = writer
+
+	pushLLMSnapshotAtStep(t, verifierInstance, 4)
+	tap := candidateByKind(t, verifierInstance.Candidates(), verifier.ActionKindTap)
+	fake.choice = tap.Index
+	fake.chosenAction = tap.Description
+	if _, err := source.NextAction(context.Background(), 4); err != nil {
+		t.Fatalf("NextAction: %v", err)
+	}
+	if err := writer.Close(); err != nil {
+		t.Fatal(err)
+	}
+
+	calls := readLLMCalls(t, directory)
+	if len(calls) != 1 {
+		t.Fatalf("recorded %d calls, want 1", len(calls))
+	}
+	call := calls[0]
+	if call.PromptTokens != 1200 || call.CompletionTokens != 34 || call.TotalTokens != 1234 {
+		t.Errorf("tokens = %d/%d/%d, want 1200/34/1234",
+			call.PromptTokens, call.CompletionTokens, call.TotalTokens)
+	}
+	if call.LatencyMillis < 15 {
+		t.Errorf("latency = %dms, want at least the server's 15ms", call.LatencyMillis)
+	}
+	if call.ServedModel != "vendor/model-2026-05" {
+		t.Errorf("served model = %q, want the id the provider reported", call.ServedModel)
+	}
+	if want := trace.ScreenshotReference(4); call.Screenshot != want {
+		t.Errorf("screenshot = %q, want %q", call.Screenshot, want)
+	}
+	if call.Reasoning == "" || call.EchoedAction != tap.Description {
+		t.Errorf("record = %+v, want the parsed reasoning and echo", call)
+	}
+}
+
+// TestLLMCallScreenshotNamesObservedStep guards the one case where the image
+// sent is not the current step's: the runner skips PushSnapshot on a
+// transitional observation, so the model still sees the last observed screen.
+func TestLLMCallScreenshotNamesObservedStep(t *testing.T) {
+	fake := newFakeOpenRouter(t)
+	source, verifierInstance := newLLMSource(t, fake)
+	pushLLMSnapshotAtStep(t, verifierInstance, 4)
+	tap := candidateByKind(t, verifierInstance.Candidates(), verifier.ActionKindTap)
+	fake.choice = tap.Index
+	fake.chosenAction = tap.Description
+	if _, err := source.NextAction(context.Background(), 6); err != nil {
+		t.Fatalf("NextAction: %v", err)
+	}
+
+	call := lastCall(t, source)
+	if call.Step != 6 {
+		t.Errorf("record step = %d, want the current step 6", call.Step)
+	}
+	if want := trace.ScreenshotReference(4); call.Screenshot != want {
+		t.Errorf("screenshot = %q, want %q: the image sent was step 4's", call.Screenshot, want)
+	}
+}
+
 func TestDownscalePNGShrinksLongEdge(t *testing.T) {
 	large := image.NewRGBA(image.Rect(0, 0, 2048, 1024))
 	var buffer bytes.Buffer

internal/runner/source.go

@@ -15,9 +15,11 @@ import (
 // ActionSource resolves the next action for a step. Both runtimes (the goja
 // picker and the web/V8 picker) implement it so the runner loop has one path
 // and no per-step driver type assertion. NextAction returns verifier.ErrNoAction
-// when the picker declined to act this tick.
+// when the picker declined to act this tick. stepIndex is the trace line the
+// decision belongs to, so a source that keeps its own records can key them to it
+// rather than counting steps a second time.
 type ActionSource interface {
-	NextAction(ctx context.Context) (verifier.Action, error)
+	NextAction(ctx context.Context, stepIndex int) (verifier.Action, error)
 }
 
 // ExtractorSource yields per-step extractor overrides the runner applies after
@@ -34,7 +36,7 @@ type gojaSource struct {
 	verifier *verifier.Verifier
 }
 
-func (s gojaSource) NextAction(context.Context) (verifier.Action, error) {
+func (s gojaSource) NextAction(context.Context, int) (verifier.Action, error) {
 	return s.verifier.NextAction()
 }
 
@@ -49,7 +51,7 @@ type webSource struct {
 	web driver.WebDriver
 }
 
-func (s webSource) NextAction(ctx context.Context) (verifier.Action, error) {
+func (s webSource) NextAction(ctx context.Context, _ int) (verifier.Action, error) {
 	raw, err := s.web.NextActionFromV8(ctx)
 	if err != nil {
 		return verifier.Action{}, fmt.Errorf("v8 next action: %w", err)
@@ -108,5 +110,10 @@ func pickSources(options Options) (ActionSource, ExtractorSource, error) {
 		logger:       logger,
 		history:      newActionHistory(llmHistorySize),
 	}
+	// Assigned only when present so the interface field stays nil rather than
+	// holding a typed nil pointer that would panic on the first record.
+	if options.TraceWriter != nil {
+		action.recorder = options.TraceWriter
+	}
 	return action, extractor, nil
 }