Claude Code newer versions send a `betas` list in the request body (e.g.
"interleaved-thinking-2025-05-14"). This landed in `__pydantic_extra__`
and triggered `_openai_reject_native_only_top_level_fields`, surfacing
as "Invalid request sent to provider." for all OpenAI Chat upstreams
(kimi, nvidia_nim).
Declare `betas` on both `MessagesRequest` and `TokenCountRequest` with
`exclude=True` so it is accepted from clients but never forwarded to any
provider body.
## What
Fixes the 422 error returned by the proxy when Claude Code v2.1.128+
sends PDF context as Anthropic `document` blocks, plus two related
DeepSeek-only issues uncovered while debugging.
Closes#357.
## Why
Claude Code now attaches PDFs/files as `document` content blocks. The
`Message.content` union didn't include that variant, so every follow-up
turn after a PDF read returned `422 Unprocessable Content` from the
proxy itself (before the request ever reached DeepSeek).
## Changes
- `api/models/anthropic.py`: new `ContentBlockDocument`, added to
`Message.content` union.
- `providers/deepseek/request.py`:
- `_strip_unsupported_attachment_blocks` — silently drops
`image`/`document` blocks for DeepSeek.
- `_serialize_tool_result_content` + `_normalize_tool_result_content` —
coerces `tool_result.content` to a string per DeepSeek's API contract.
- Hooked into `build_request_body` (strip before validation, normalize
before send).
- `tests/providers/test_deepseek.py`: +4 tests, 1 legacy test updated.
## Verification
- `uv run ruff format` ✅
- `uv run ruff check` ✅
- `uv run ty check` ✅
- `uv run pytest` ✅ (1194 passed)
- Manual: `claude-deepseek` against a PDF no longer 422s.
## Scope / non-goals
- No README change (per CONTRIBUTING).
- Image/document blocks for DeepSeek are stripped, not converted to text
— Claude Code already ships the extracted text in the paired
tool_result.
- No changes to other providers.
---------
Co-authored-by: Alishahryar1 <alishahryar2@gmail.com>
## What changed
This updates the command parsing helpers to recognize leading
environment assignments only when they match shell-style variable syntax
like `FOO=bar`. The previous implementation treated any leading token
containing `=` as an env assignment, which could misclassify real
commands or file paths that happen to include `=`. Both
`_strip_env_assignments()` and `extract_command_prefix()` now share the
same validation helper so their behavior stays consistent.
## Why it matters
These helpers are used by fast-path request optimizations, so incorrect
prefix detection can bypass normal handling with the wrong result.
Restricting env stripping to valid assignment tokens fixes false
positives without changing the public API or the response format.
## How it was tested
I validated the updated logic against representative inputs such as
`FOO=bar git status`, plain commands like `git status`, and malformed or
path-like tokens containing `=` that should not be treated as env
assignments. The change is isolated to parsing logic and does not
introduce new dependencies.
---------
Co-authored-by: genoshide <genoshide@users.noreply.github.com>
## What changed
This updates `extract_filepaths_from_command()` to skip leading
environment variable assignments before detecting the actual command.
The same parsing pattern was already used in `extract_command_prefix()`,
but filepath extraction still treated the first token as the command
even when it was an env assignment.
## Why it matters
Commands like `FOO=bar cat README.md` or `DEBUG=1 grep pattern file.txt`
were incorrectly classified because `FOO=bar` was treated as the
executable name. That caused the optimization path to return an empty
`<filepaths>` block instead of the file paths the command actually
reads.
## How it was tested
Tested the updated logic against env-prefixed read commands such as
`FOO=bar cat README.md` and `DEBUG=1 grep needle app.py`. Also verified
existing behavior remains unchanged for plain commands like `cat
file.txt`, `grep pattern file.txt`, and listing commands like `ls`.
Co-authored-by: genoshide <genoshide@users.noreply.github.com>
- Point DeepSeek at api.deepseek.com/anthropic with x-api-key headers
- Native request builder, DeepSeek-specific thinking/block sanitization
- Drop deepseek from OpenAI-chat server-tool preflight; update tests and docs
- Default smoke model deepseek-v4-pro; re-export dump_raw_messages_request
Consolidates the incremental refactor work into a single change set: modular web tools (api/web_tools), native Anthropic request building and SSE block policy, OpenAI conversion and error handling, provider transports and rate limiting, messaging handler and tree queue, safe logging, smoke tests, and broad test coverage.
Expand is_title_generation_request to match sentence-case/JSON title prompts
in addition to legacy new-conversation-topic copy. Add unit test for the
current session-title system text shape.
Add proxy support for providers based on
[doc](https://www.python-httpx.org/advanced/proxies/):
- Add per-provider proxy support (HTTP and SOCKS5) for all 4 providers:
nvidia_nim, open_router, lmstudio, llamacpp
- Each provider gets its own env var (NVIDIA_NIM_PROXY,
OPENROUTER_PROXY, LMSTUDIO_PROXY, LLAMACPP_PROXY) for independent proxy
configuration
---------
Co-authored-by: Alishahryar1 <alishahryar2@gmail.com>
## Summary
* add native DeepSeek provider support via the shared OpenAI-compatible
provider base
* allow `deepseek/...` model prefixes in config validation
* add `DEEPSEEK_API_KEY` and `DEEPSEEK_BASE_URL` settings
* add DeepSeek entries to `.env.example` and `config/env.example`
* implement `DeepSeekProvider` and register it in provider dependencies
* add a DeepSeek request builder with DeepSeek-specific thinking payload
handling
* preserve Anthropic thinking blocks as `reasoning_content` for
DeepSeek-compatible continuation flows
* update `claude-pick` to discover DeepSeek models from the DeepSeek API
* document DeepSeek usage in `README.md`
* add tests for config validation, provider dependency wiring, request
building, and streaming behavior
## Motivation
DeepSeek exposes an OpenAI-compatible API and can be used directly
without routing through OpenRouter. This lets users spend their existing
DeepSeek balance through the proxy while keeping the same Claude Code
workflow and per-model provider mapping.
## Example
```dotenv
DEEPSEEK_API_KEY="sk-..."
DEEPSEEK_BASE_URL="https://api.deepseek.com"
MODEL_OPUS="deepseek/deepseek-reasoner"
MODEL_SONNET="deepseek/deepseek-chat"
MODEL_HAIKU="deepseek/deepseek-chat"
MODEL="deepseek/deepseek-chat"
---------
Co-authored-by: Alishahryar1 <alishahryar2@gmail.com>
The max_sessions cap in CLISessionManager was the only thing enforcing
a limit on concurrent CLI processes. Now that provider concurrency is
controlled at the streaming layer (PROVIDER_MAX_CONCURRENCY semaphore),
the CLI session pool cap is redundant and removed entirely.
Changes:
- cli/manager.py: remove max_sessions param, cap check, _cleanup_idle_sessions_unlocked, max_sessions from get_stats()
- config/settings.py: remove max_cli_sessions field
- api/app.py: remove max_sessions=settings.max_cli_sessions from CLISessionManager constructor
- messaging/handler.py: remove "Waiting for slot" status check; stats display no longer shows Max CLI
- .env.example: remove MAX_CLI_SESSIONS line
- tests/cli/test_cli.py: remove max_sessions args and assertion from manager tests
- tests/cli/test_cli_manager_edge_cases.py: remove two tests for cap/cleanup behavior
- tests/api/test_app_lifespan_and_errors.py: remove max_cli_sessions from all SimpleNamespace settings
- tests/config/test_config.py: remove max_cli_sessions isinstance assertion
- tests/conftest.py: remove max_sessions from mock stats
- tests/messaging/test_handler.py: merge slot/capacity tests into single new-conversation test; remove Max CLI assertion from stats test
- tests/messaging/test_handler_markdown_and_status_edges.py: remove "Waiting for slot" assertion; drop max_sessions from all stats mocks
https://claude.ai/code/session_014mrF1WMNgmNjtPBuoQHsbg
Adds max_concurrency cap to GlobalRateLimiter using asyncio.Semaphore.
A request now waits for a concurrency slot before the sliding window rate
limit check, so at most N streams are open to the provider simultaneously,
even when the rate window would allow more.
Changes:
- providers/rate_limit.py: max_concurrency param, _concurrency_sem, concurrency_slot() asynccontextmanager
- providers/openai_compat.py: pass max_concurrency to limiter; wrap execute_with_retry + stream iteration in concurrency_slot()
- providers/base.py: max_concurrency field on ProviderConfig
- config/settings.py: provider_max_concurrency setting (PROVIDER_MAX_CONCURRENCY env var, default None = unlimited)
- api/dependencies.py: pass provider_max_concurrency into all three provider ProviderConfig instantiations
- .env.example: document PROVIDER_MAX_CONCURRENCY (commented out)
- tests/providers/test_provider_rate_limit.py: 5 new tests covering concurrency limit enforcement, slot release on exception, noop when unconfigured
- tests/api/test_dependencies.py: add provider_max_concurrency=None to mock settings helper
https://claude.ai/code/session_014mrF1WMNgmNjtPBuoQHsbg
Alishahryar1