Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- Return error when marker targets OpenAI provider but no connector available
  (instead of silently falling through to default)
- Propagate stream errors as SSE error events to client
  (instead of silently dropping them)
- Add tests for error event serialization and model override ordering

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- OpenAI connector doesn't emit NormalizedStreamEvent::Start, but
  Anthropic clients expect a message_start event. Prepend a synthetic
  Start event to the stream.
- Respect sse_keepalive_enabled flag in cross-dialect streaming path
  (matching same-dialect behavior).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Claude Code sends thinking, redacted_thinking, and other block types that
our typed AnthropicContentBlock enum didn't handle. Added custom Deserialize
with Unknown catch-all variant for both AnthropicContentBlock and
AnthropicSystemBlock. Also handles tool_result content as either string or
array of blocks (both are valid in the Anthropic API).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Config values like api_key: "${CLOUDFLARE_API_KEY}" were passed as literal
strings instead of being resolved to their environment variable values.
Added shellexpand::full() to expand both ${VAR} and ~ in the raw config
before parsing. Missing env vars now produce a clear startup error.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Anthropic requires tool_use.id to match ^[a-zA-Z0-9_-]+$. When cross-dialect
routing converts OpenAI provider responses back to Anthropic format, tool call
IDs are now sanitized and prefixed with toolu_ to ensure they pass Anthropic's
validation when Claude Code includes them in subsequent conversation history.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

When a model responds with tool_use, Claude Code sends the tool results
as a new user message without the LUNAROUTE marker. The marker extraction
now skips tool-result-only user messages and looks back to the message
that initiated the tool call chain, ensuring multi-step tool calling
stays on the same provider.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

1. Config env expansion: use non-failing mode (env_with_context_no_errors)
   so ${VAR} in comments or unresolved vars don't cause config load failures.

2. Content block deserializer: validate required fields (text, id, name,
   tool_use_id) and return proper deserialization errors instead of silently
   defaulting to empty strings for known block types.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…ifecycle

- strip_marker now skips tool-result-only messages like extract_marker does,
  preventing marker text from leaking in conversation history
- Streaming conversion tracks active content block index (Option<u32>) instead
  of a boolean, ensuring content_block_stop is emitted for both text and tool
  blocks at the correct index
- Added regression tests for both fixes

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…valid tool_result content

- Stream state now uses ActiveBlock enum (Text/Tool with index) instead of
  Option<u32>, ensuring type/index transitions always emit proper stop/start
  events even for same-index different-type scenarios
- tool_result.content with unsupported shapes (e.g. object, number) now
  returns a deserialization error instead of silently defaulting to empty
- Added debug logging when unknown content blocks are skipped during
  normalization for easier diagnosis

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- Only emit InputJsonDelta when the matching tool block is active,
  preventing invalid SSE ordering if arguments arrive before block start
- to_anthropic_tool_id now sanitizes first then checks for toolu_ prefix,
  avoiding double-prefix (toolu_toolu_*) on upstream IDs that already
  have the prefix but contain invalid characters

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…ination

- to_anthropic_tool_id now replaces invalid chars with _ instead of
  dropping them, preserving position info to avoid collisions between
  distinct upstream IDs (e.g. call.a:1 vs calla1)
- Stream error handlers now close any active content block and send
  message_stop before the error event, ensuring valid SSE sequencing
  for strict Anthropic clients (fixed in both passthrough and
  cross-dialect streaming paths)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- to_anthropic_tool_id appends a 4-char hash suffix when sanitization
  alters the original ID, ensuring distinct upstream IDs produce distinct
  Anthropic IDs even if they sanitize to the same base form
- ToolCallDelta handler now checks if the target tool block is already
  active before opening, preventing spurious stop/start events when a
  provider repeats id/name across chunks

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- Increased hash suffix from 16-bit to 32-bit for sanitized tool IDs,
  reducing collision probability from ~1/65K to ~1/4B per ID pair
- Added TODO documenting that cross-dialect path skips response recording
  (request recording works, response/tool-call recording needs design)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…version

Replace simple ActiveBlock tracking with StreamBlockState that maintains a
monotonically increasing index allocator. OpenAI tool_call_index values
(which start at 0) can collide with text delta indices; the allocator
ensures each Anthropic content_block gets a unique index regardless of
the upstream indexing scheme. Includes HashMap mapping OpenAI
tool_call_index to allocated Anthropic indices.

Added regression test for text-then-tool at same OpenAI index 0.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>