fix: resolve $0.00 session costs in tapes deck

[bdougie]

Summary

• Cost calculation fallback: Track last-seen model in sessionFromNodes loop and use it when response nodes have empty model fields. This fixes the root cause : 2042 of 2048 of my response nodes since Feb 19 had empty models, causing cost calculation to skip them entirely.
• Missing pricing: Add claude-haiku-4.6 to DefaultPricing() so sessions using it get costed correctly.
• Streaming model fallback: In enqueueStreamedResponse, fall back to the request model when the reconstructed response has an empty model, preventing future empty-model nodes.

Test plan

• New Ginkgo tests for empty-model fallback (last-seen model used, no-model-yet skipped)
• New pricing tests for claude-haiku-4.6 resolution (exact and date-suffixed)
• All 125 deck tests pass
• All 55 proxy tests pass (including Anthropic streaming model capture)
• go build ./... succeeds

---

Continue Tasks: ✅ 1 no changes — View all

[oppegard]

I'm not really familiar with much of the codebase yet, so let me know if this is a non-issue.

Codex found this issue:

Apply empty-model fallback in detail costing
buildSessionSummaryFromNodes now reuses the previous model when n.Model is empty, but buildSessionMessages still calls costForNode, which returns zero cost for empty-model nodes. In sessions with streamed responses missing Model, the summary and overview totals will now include those costs while per-message and grouped-message totals remain at zero, producing inconsistent cost reporting for the same session (summary vs detail view).

I had it write a simple failing test to show the inconsistency. I ran code coverage on the deck package as well, and it appears we don't have any tests exercising buildSessionMessages() in query.go.

It("keeps summary and message totals consistent", func() {
		pricing := DefaultPricing()
		q := &Query{pricing: pricing}

		nodes := []*ent.Node{
			{ID: "node-1", Role: "user", Model: "claude-opus-4-6-20260219"},
			{
				ID:               "node-2",
				Role:             "assistant",
				Model:            "",
				CompletionTokens: intPtr(500000),
			},
		}

		summary, _, _, err := q.buildSessionSummaryFromNodes(nodes)
		Expect(err).NotTo(HaveOccurred())

		messages, _ := q.buildSessionMessages(nodes)
		messageTotal := 0.0
		for _, msg := range messages {
			messageTotal += msg.TotalCost
		}

		Expect(summary.TotalCost).To(BeNumerically("~", messageTotal, 1e-12))
	})

[bdougie]

Good catch — this is a real inconsistency. buildSessionMessages was calling costForNode which bails on empty models, while buildSessionSummaryFromNodes had the last-seen model fallback. Summary says "$X" but per-message detail sums to "$0" for those nodes.

Fixed in the latest push:

• Renamed costForNode → costForModel (accepts a model string directly)
• Added the same lastModel tracking to buildSessionMessages
• Resolved model is now set on SessionMessage.Model too, so the detail view shows the correct model
• Added your consistency test (summary.TotalCost ≈ Σ message.TotalCost) — 126 deck tests and 55 proxy tests pass

[continue]

No docs update needed. This PR fixes an internal bug in session cost calculation where response nodes with empty model fields were being skipped. The changes are:

1. Cost calculation fallback: Uses last-seen model when response nodes have empty model fields
2. Missing pricing: Adds claude-haiku-4.6 to the internal pricing table
3. Streaming model fallback: Falls back to request model when reconstructed response has empty model

These are all internal implementation fixes that don't change any user-facing configuration, commands, or documented behavior. The cost display in tapes deck will now show accurate costs instead of $0.00 for affected sessions, but no documentation changes are required since cost calculation was already a documented/expected feature.

PR #133 was merged: fix: resolve $0.00 session costs in tapes deck
PR URL: #133
Merged by: bdougie