feat: add Anthropic Messages to OpenAI ChatCompletions translator - #464
Open
khshanovskyi wants to merge 10 commits into
Open
khshanovskyi wants to merge 10 commits into
khshanovskyi wants to merge 10 commits into
Conversation
khshanovskyi
force-pushed
the
feat/add-openai-responses-transaltor
branch
from
July 14, 2026 09:02
f5d9eab to
d733893
Compare
khshanovskyi
changed the title
Add a FastAPI sub-app (aidial_adapter_bedrock/anthropic_translator/chat_completions/)
mounted at /to-chat-completions/anthropic, translating Anthropic Messages
requests (/v1/messages) into OpenAI Chat Completions API calls against
DIAL Core ({DIAL_URL}/openai/deployments/{deployment}/chat/completions)
and translating the streaming/non-streaming responses back into Anthropic
Messages shape. This lets an Anthropic-native client (e.g. Claude Code)
talk to a model that only speaks the OpenAI Chat Completions API.
- Support the bulk of the Anthropic Messages surface: system prompt
(string/blocks, including mid-conversation system-role turns merged
into one leading system message), user/assistant messages, tool_result/
tool_use, image/document blocks, tool_choice, reasoning-effort mapping
from output_config.effort, and cache_control propagated as DIAL Core's
custom_fields.cache_breakpoint. Unsupported features (top_k, thinking
budget, web_search/bash/text_editor/computer/code_execution tools,
mcp_servers, container) are dropped with a warning instead of being
rejected, so clients keep working. There is no count_tokens endpoint,
since Chat Completions has no token-counting API.
- Map DIAL Core and Chat Completions errors into Anthropic-shaped error
responses, for both the HTTP error body and the terminal SSE error
event.
- Promote openai from a test-only to a runtime dependency (1.107.3 ->
2.40.0) to reuse its typed Chat Completions client and SSE handling
for the outbound call to Core.
- Document the new interfaces.anthropicMessages DIAL Core config
pattern and the translator's feature-support/limitations table in
README.md.
- Add unit tests covering request/response conversion, the streaming
state machine, error mapping, and the mounted endpoints.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
khshanovskyi
force-pushed
the
feat/add-openai-responses-transaltor
branch
from
July 23, 2026 10:38
3f57ac5 to
5dcc092
Compare
Translate Anthropic's output_config.format (json_schema) to DIAL's response_format on the Chat Completions request; unsupported format types or a missing schema are dropped with a warning, consistent with the translator's other defensive-conversion behavior. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
# Conflicts: # poetry.lock
…i-responses-transaltor # Conflicts: # poetry.lock
Rolls back the capability discovery driven by the outdated
CHAT_COMPLETIONS_TRANSLATION_SPEC.md and re-implements the translator
against chatcompletion-transaltions.md, taken from the tested PoC.
DIAL Core stamps `x-dial-deployment-features` on every request it routes
here, so the `GET {DIAL_URL}/openai/models` catalog — and its per-credential
TTL/LRU cache, single-flight lock and bounded fetch — is gone, along with
the three TRANSLATOR_MODEL_CATALOG_* settings. The handler now makes exactly
one outbound call.
Only `features` is on the wire; the deployment's `defaults` and `limits` are
not, which changes three things:
- Reasoning is a single top-level `reasoning_effort` gated on
`features.reasoning_efforts`, replacing the three mutually exclusive knobs.
Intent is resolved first (`thinking.disabled` outranks an effort sent
alongside it; a budget-less `thinking` block means "think at the default"),
then degraded along none < minimal < minimum < low < medium < high < xhigh
< max, walking down before up and never crossing `none` in either
direction.
- Nothing is nested into a deployment's own `custom_fields.configuration`.
- `max_tokens` is forwarded verbatim: there is no ceiling to clamp against.
Also brings the rest of the translator in line with the PoC:
- `cache_control.ttl` becomes DIAL's absolute `expire_at`, with the longest
winning where several blocks collapse onto one message or tool.
- Cache writes are recovered from `prompt_tokens_details.cache_write_tokens`
(or `cacheWriteTokens`) and subtracted from `input_tokens` alongside cache
reads; `reasoning_tokens` surfaces as `output_tokens_details
.thinking_tokens` in both modes — the streaming path was dropping the
field `MessageDeltaUsage` declares for it.
- System-origin text merges in client order rather than grouped by kind, and
an unsupported system content block warns on the way out.
- `context_management` is warn-dropped; a stage named "thought" is
recognised as reasoning.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Review feedback: a function returning a tuple is hard to read at the call site, and the package carried comments that only restated the code. `_handle_messages` opened on `base_url, req, deployment = await prepare(...)`, which said nothing about what any of the three were. Every multi-value return in the translator is now a single one, and the handler names and types each step: - `prepare` is deleted — it only bundled `require_base_url`, `parse_request` and `resolve_model`, which the handler now calls directly. `resolve_model` is renamed `resolve_deployment` to match what it returns. - `to_chat_completions_request` takes the `ToolNameAliases` registry as a parameter instead of returning it, so its lifetime is visible where the response path reads it. - `_convert_message` returned four values that were all independently derivable; `from_chat_completions` derives each one and `_content_blocks` returns only the blocks. - `_convert_tool_choice` is split from `_convert_parallel_tool_calls`: `disable_parallel_tool_use` is read independently of `tool_choice.type`, so the split states that rather than hiding it in a tuple. - `_system_text`/`_collect_system` return a `SystemPrompt`, and `apply_stop_sequences` a `StopMatch`. - `_tool_result_output` is split into `_tool_result_message` and `_tool_result_images`. - `caller_credential` is inlined into `core_headers`, its only caller since the model catalog was removed. `emulated_stop_sequences` replaces the inline conditional in the handler and sits beside `strips_stop_parameter`, which decides the mirror case on the request side. `_blocks` now filters non-object entries, so every content-array walk is defensive rather than just the tool-result one. Comments are cut by roughly 40%: anything narrating the next statement is gone, while the decisions and incidents the code cannot show — the Vertex `thinking_level` rejection, `strict=False` against `strict=True`, the `SdkTool | StaticTool` annotation, the bool-is-an-int budget trap — stay. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
# Conflicts: # .ort.yml # README.md # aidial_adapter_bedrock/app.py # poetry.lock # pyproject.toml
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Applicable issues
Description of changes
Add a FastAPI sub-app (aidial_adapter_bedrock/anthropic_translator/chat_completions/)
mounted at /to-chat-completions/anthropic, translating Anthropic Messages
requests (/v1/messages) into OpenAI Chat Completions API calls against
DIAL Core ({DIAL_URL}/openai/deployments/{deployment}/chat/completions)
and translating the streaming/non-streaming responses back into Anthropic
Messages shape. This lets an Anthropic-native client (e.g. Claude Code)
talk to a model that only speaks the OpenAI Chat Completions API.
Support the bulk of the Anthropic Messages surface: system prompt
(string/blocks, including mid-conversation system-role turns merged
into one leading system message), user/assistant messages, tool_result/
tool_use, image/document blocks, tool_choice, reasoning-effort mapping
from output_config.effort, and cache_control propagated as DIAL Core's
custom_fields.cache_breakpoint. Unsupported features (top_k, thinking
budget, web_search/bash/text_editor/computer/code_execution tools,
mcp_servers, container) are dropped with a warning instead of being
rejected, so clients keep working. There is no count_tokens endpoint,
since Chat Completions has no token-counting API.
Map DIAL Core and Chat Completions errors into Anthropic-shaped error
responses, for both the HTTP error body and the terminal SSE error
event.
Promote openai from a test-only to a runtime dependency (1.107.3 ->
2.40.0) to reuse its typed Chat Completions client and SSE handling
for the outbound call to Core.
Document the new interfaces.anthropicMessages DIAL Core config
pattern and the translator's feature-support/limitations table in
README.md.
Add unit tests covering request/response conversion, the streaming
state machine, error mapping, and the mounted endpoints.
Checklist
By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.