v0.9.2

Bug Fixes

• cost: A task timeout with no preserved turn is unrecorded spend, not free
  (#63,
  93c7fc0)

• cost: Book spend on the error and timeout paths, flag what is unpriced
  (#63,
  93c7fc0)

• cost: Flag every hard-killed task as a cost floor, not just the empty ones
  (#63,
  93c7fc0)

• evalboard: Honest scoped counts, and one definition of a run's scope
  (#69,
  0bdac0b)

• litellm: Gate cost_log_tags on agent capability, not route (fixes non-Claude crash)
  (#66,
  4131a2a)

• litellm: Make the orphaned-spend warning actually fire
  (#66,
  4131a2a)

• litellm: Per-attempt cost-log scoping + single run-id accessor + no-match warning
  (#66,
  4131a2a)

• litellm: Pin each open-weight model to a vetted provider set (no silent fallback)
  (#66,
  4131a2a)

• litellm: Proxy-authoritative token buckets + all-priced gate + transactional join
  (#66,
  4131a2a)

• litellm: Sanitize cost headers, reject non-finite cost, drop debug scaffolding
  (#66,
  4131a2a)

• orchestrator: Recover the in-flight turn's spend on a hard kill
  (#63,
  93c7fc0)

• pricing: Add the claude-opus-5 rate so killed turns stop booking zero
  (#63,
  93c7fc0)

• pricing: Add the five unpriced codex tiers still on OpenAI's rate card
  (#63,
  93c7fc0)

• pricing: Correct every wrong rate-card entry and close the alias gaps
  (#63,
  93c7fc0)

• pricing: Refresh the rate card and correct gemini-3-flash-preview
  (#63,
  93c7fc0)

• reports: Count errors as misses and stop losing cost on error paths
  (#63,
  93c7fc0)

• reports: Count errors as misses in one canonical pass rate
  (#63,
  93c7fc0)

Code Style

• evalboard: Drop the swatch dots and the scope caption from the header
  (#69,
  0bdac0b)

Documentation

• cost: Describe the per-turn backfill as the net it is
  (#63,
  93c7fc0)

• cost: Describe the unpriced-crash mechanism accurately and keep comments framework-general
  (#63,
  93c7fc0)

• litellm: Correct the cost contract after cutting per-message distribution
  (#66,
  4131a2a)

• litellm: Document LITELLM_COST_LOG wiring + correct the reconciliation-cost contract
  (#66,
  4131a2a)

Features

• cost: Publish one accurate total on every reporting surface
  (#63,
  93c7fc0)

• docker: Bind-mount the LiteLLM cost log so --driver docker joins actual cost
  (#66,
  4131a2a)

• evalboard: Compare every harness on the overview, and scope the whole page to one
  (#69,
  0bdac0b)

• evalboard: Compare harnesses on the overview, and identify each run
  (#69,
  0bdac0b)

• evalboard: Lift the harness scope to the page header, in vendor colors
  (#69,
  0bdac0b)

• evalboard: Make each turn's provider-call table a collapsed dropdown
  (#66,
  4131a2a)

• evalboard: Mark a partly-priced run total as a floor, not the bill
  (#63,
  93c7fc0)

• evalboard: One set of pass-rate cutoffs, and a run table that pages through all history
  (#69,
  0bdac0b)

• evalboard: Per-call cost/cache table from provider_call_costs (replaces inline)
  (#66,
  4131a2a)

• evalboard: Read the canonical pass rate and surface incomplete cost
  (#63,
  93c7fc0)

• evalboard: Say which harness, model, and framework version a run used
  (#69,
  0bdac0b)

• litellm: Actual per-call cost + cache accounting for the open-weight backend
  (#66,
  4131a2a)

Refactoring

• cost: Correct the simulator-cost bound and drop the unread variant error share
  (#63,
  93c7fc0)

• cost: Cut the commentary and drop unreachable rate-card keys
  (#63,
  93c7fc0)

• cost: Define the unpriced-row test once, and only for new runs
  (#63,
  93c7fc0)

• cost: Total_cost_usd means the whole bill everywhere
  (#63,
  93c7fc0)

• evalboard: Call the UiPath harness Delegate
  (#69,
  0bdac0b)

• litellm: Cut per-message distribution; turn-level join + per-call audit record
  (#66,
  4131a2a)

• litellm: Drop the provider field/column — unavailable on the streaming path
  (#66,
  4131a2a)

• litellm: Stream the cost log + de-duplicate the OpenRouter config comment
  (#66,
  4131a2a)

Testing

• litellm: Cover config shape, join ordering, and defensive cost branches
  (#66,
  4131a2a)