Make reasoning policy provider-neutral and client-aware (#1148)

Preserve mid-conversation system messages through provider conversion (#1125)

Fix Auto Mode classifier responses when stream is omitted (#1098)

Canonicalize provider failure and retry ownership (#1046)

Introduce a typed application boundary for provider execution (#1045)

Establish single-owner runtime with stream-safe provider hot swaps (#1036)

Restore protocol-correct provider errors without client retry loops (#1033)

Move runtime packages under src namespace (#1029)

Preserve messaging transcript on stop

Fix pre-start stream failures returning HTTP 200 (#1026)

Refactor provider runtime ownership (#925)

Refactor messaging workflow architecture (#852)

fix: accept system role messages

Added claude-code native model picker

Validate configured models at startup

Major refactor: API, providers, messaging, and Anthropic protocol

Add per-model thinking toggles

feat: Anthropic web server tools, provider metadata, messaging hardening

Architecture refactor: core anthropic, runtime, smoke tiers, remove providers.common

Fixes for issue 113 and 116

Per claude model mapping (#66)

Improve deterministic error surfacing across stream and API

Improved optimizations to decrease llm calls further and increase throughput

new linter rules and fixes

Phase 7: Directory restructuring (messaging/ and tests/)