v0.2.1

• Peer ranges of the harness packages now accept 0.1.7-* prereleases. The 0.2.0 range >=0.1.6-alpha.2 matched no 0.1.7 prerelease under semver's prerelease rule, so installing on DSH 0.1.7-alpha.2 reported unmet peers.

v0.2.0

• Reasoning levels per model from OpenRouter's reasoning.supported_efforts, including Off for optional reasoning and the model's default level.
• Image input for models whose OpenRouter entry lists image, within the harness image budget and offload protocol.
• reasoning_details replay across tool calls, including signed and encrypted entries.
• Prompt-cache breakpoints for anthropic/* models; x-session-id session affinity for all models.
• Failure classification with Retry-After, upstream provider detail, stream idle timeout and truncated-stream detection.
• models glob filter, toolsOnly, extraBody, retryPolicy, and route id / display name for running several instances.
• Models sorted by name (which groups them by vendor), with the price per 1M tokens in the name, such as ($0.15/$0.60), because pickers show names but not descriptions (priceInName); descriptions carry context size and price.
• Actual spend tracking: OpenRouter's usage.cost per response is appended to a JSONL ledger, attributed to the session and to every session it was delegated from. The /cost command reports session, subagent and per-model totals, and the Web UI shows the session total under the composer.
• Type-checked JSDoc, generated declarations, unit tests and a real LlmRuntime composition test.