plat 008

← Pulse platform issue index

PLAT-008 — phase costs omit input and can use the wrong rate card

Coordination	Value
Assigned agent	Unassigned
Ticket state	runtime_verified
Last synchronized	2026-08-05

Claim this ticket in this file before implementation. During active work, update this fragment rather than the shared index; synchronize the index once at handoff, review, or completion.

• Priority: P1
• Owner: cost observer/phase ledger pricing
• Source finding: HARNESS-PHASE-COST-PRICING-2026-08-03
• Source database: Workflow/build-in-public/db/db.sqlite
• Problem: phase rows omitted input cost and one Opus row used Sonnet output and cached-input rates.
• Impact: historical and daily spend is materially understated and changes without a workload change.
• Implementation (2026-08-03): Claude transcript usage now reports total prompt input (fresh + cache create + cache read), while retaining the raw cache components. Phase persistence carries cache read/write separately, recalculates every component from the effective model, and stamps pricing_model_id plus pricing_version. Claude Opus 5 and Sonnet 5 golden cases prove distinct input/output/cache-read/cache-write rates and totals.
• Acceptance: one immutable model identity selects one versioned rate card; input, output, reasoning, cache-read, and cache-write components reconcile to the total. Golden tests cover both Opus and Sonnet on adjacent dates.

Tectonicus follow-up — 2026-08-05

Tectonicus Pulse found a separate reconciliation symptom after the rate-card repair: costs/phase/daily and costs/execution disagree by 3.1× under a single opaque workflow-builder bucket. The workflow has no producer, plan step, or configuration surface that controls that grouping, so it is platform-owned.

This does not refute the verified rate-card repair. It shows that correct per-call pricing alone is insufficient when summaries cannot reconcile the same execution across phase/daily and execution views. Coordinate with PLAT-009 and PLAT-031: add a real per-execution reconciliation test that proves each view's scope, includes every charge once, and labels intentionally cumulative views as non-comparable rather than presenting them as a run total.

Scope-presentation repair — 2026-08-05

Implemented; needs a fresh workflow/Pulse run for runtime re-verification. get_cost_summary(run_folder=...) now states that its execution totals cover only that run and renders Builder/Pulse usage as a cumulative reference which must not be added to it. The dashboard likewise presents Selected run cost separately from Builder/Pulse spend across recorded days, rather than silently forming a misleading mixed-scope “total cost.” This prevents the observed 3.1× comparison from being presented as a run-cost mismatch while PLAT-009/031 continue to improve ledger reconciliation itself.

Focused verification: the workflow-review sequence tests and the frontend production build pass.