Changelog¶

This page records the major version changes and important feature updates of the Argo Proxy project.

v3.0.4 (2026-05-23)¶

Fix: startup user validation model selection.

Fixed¶

• Robust model selection for startup username validation: The _fetch_first_model helper blindly picked the first model from /v1/models, which was gpt4olatest — an internal ID that ARGO's /v1/chat/completions endpoint rejects. This caused the startup username validation to always fail with a misleading "network may be unavailable" warning. Replaced with _fetch_validation_models that filters out embedding models, sorts candidates by cost (nano → mini → others), and returns a full list. validate_user_async now iterates through candidates, skipping any that return 400/model_not_found, instead of retrying the same broken model

Full Changelog: https://github.com/Oaklight/argo-proxy/compare/v3.0.3...v3.0.4
PyPI: https://pypi.org/project/argo-proxy/3.0.4/

v3.0.3 (2026-05-23)¶

Hotfix: thinking normalization in force-conversion paths.

Fixed¶

• Normalize thinking.type in all conversion paths (#123, #124): _normalize_thinking_for_upstream was only called in the passthrough path. The three force-conversion paths (_convert_non_streaming, _convert_buffered_streaming, _convert_streaming) sent thinking.type=enabled to Argo as-is, causing 502 errors on non-streaming requests and silent SSE errors on streaming requests for models that require thinking.type=adaptive (e.g. Claude Opus 4.7)

Full Changelog: https://github.com/Oaklight/argo-proxy/compare/v3.0.2...v3.0.3
PyPI: https://pypi.org/project/argo-proxy/3.0.3/

v3.0.2 (2026-05-20)¶

Hotfix: Claude Opus 4.7 support.

Fixed¶

• Strip temperature for reasoning models (#120): Claude Opus 4.7 is a reasoning model that rejects the temperature parameter with 400 invalid_request_error. argo-proxy now automatically strips temperature before forwarding requests to Opus 4.7, applied early in the dispatch pipeline so it covers both passthrough and cross-format conversion paths

Added¶

• claudeopus47 model alias: Added model registry alias mapping to argo:claude-opus-4.7 / argo:claude-4.7-opus

Full Changelog: https://github.com/Oaklight/argo-proxy/compare/v3.0.1...v3.0.2
PyPI: https://pypi.org/project/argo-proxy/3.0.2/

v3.0.1 (2026-05-06)¶

New¶

• --dump-requests CLI option: Debug flag that dumps full request and response payloads for all requests, not just errors
• llm-rosetta update detection: Added update checks for the llm-rosetta dependency to all version-check touchpoints

Fixed¶

• thinking.type normalization for upstream compatibility (#120): Fixed reverse normalization of thinking.type that broke upstream compatibility with the ARGO gateway

Changed¶

• Upgraded llm-rosetta dependency to >=0.6.0,<0.7.0
• Updated vendored zerodep modules (semver 0.4.1, yaml 0.3.1)
• Replaced PyYAML and packaging with vendored zerodep modules

CI & Docs¶

• Switched to pre-commit for lint checks
• Added lint/type-check CI workflow with pinned dev tooling versions
• Rewrote README for v3.x universal gateway architecture

Full Changelog: https://github.com/Oaklight/argo-proxy/compare/v3.0.0...v3.0.1
PyPI: https://pypi.org/project/argo-proxy/3.0.1/

v3.0.0 (2026-04-19)¶

First stable release of the v3 architecture.

After 16 beta iterations (b4–b16), argo-proxy v3 is now considered production-ready. This release includes all beta changes plus two final fixes.

Fixed¶

• max_tokens → max_completion_tokens in passthrough mode (#112): OpenAI deprecated max_tokens for newer models (GPT-4o, o1, etc.). In same-format passthrough (openai_chat → openai_chat), the deprecated parameter was forwarded as-is, causing 400 errors. Now automatically converted to max_completion_tokens
• thinking.type: "adaptive" normalization (#112): Anthropic's adaptive thinking type is not supported by all upstream gateways. In passthrough mode, now automatically converted to "enabled" with budget_tokens derived from max_tokens
• content: null normalization (#113): Assistant messages with tool_calls often have content: null (valid per OpenAI API spec). Some upstream gateways crash when iterating over null content. Now normalized to content: "" in all dispatch paths

Changed¶

• Requires llm-rosetta >= 0.5.1

Full Changelog: https://github.com/Oaklight/argo-proxy/compare/v3.0.0b16...v3.0.0
PyPI: https://pypi.org/project/argo-proxy/3.0.0/

Post v3.0.0 Updates (2026-04-19)¶

Dependency Reduction

• Replaced PyYAML and packaging with vendored zero-dependency modules (zerodep/yaml v0.3.1, zerodep/semver v0.4.0), reducing external runtime dependencies from 8 to 6

Dependency Update Detection

• Added llm-rosetta update detection to all version-check touchpoints:
  • argo-proxy update check now shows a table with both argo-proxy and llm-rosetta versions (columns: Package, Installed, Stable, Pre, Status)
  • argo-proxy --version appends dependency status when updates are available
  • Startup banner shows a warning when llm-rosetta has an update
  • /version endpoint includes a dependencies field in the JSON response

CI/CD

• Added CI workflow (.github/workflows/ci.yml) with ruff lint, format check, and ty type check
• Pinned dev tooling: ruff >= 0.15.0, ty >= 0.0.31, complexipy >= 5.2.0

v3.0.0b16 (2026-04-16)¶

Fixed¶

• Gemini parallel tool call rejection via ARGO gateway (#109): The ANL ARGO gateway rejects requests containing parallel tool calls (multiple tool_calls in a single assistant message) for Gemini models. argo-proxy now detects Gemini-family models and automatically reorders parallel tool calls into sequential assistant+tool message pairs before forwarding upstream. The reordering logic has been extracted from the legacy pipeline into utils/tool_calls.py for long-term maintainability

v3.0.0b15 (2026-04-14)¶

Fixed¶

• developer role rejected by Argo upstream (#107): The ANL Argo gateway uses an older OpenAI SDK that does not recognize the developer role (introduced for reasoning models o1/o3/o4). The gateway mishandles the unrecognized role, injecting a spurious tool_calls field and causing a misleading Unknown parameter: 'messages[0].tool_calls' error. argo-proxy now downgrades developer → system in all outbound paths (passthrough, cross-format, buffered-streaming). The mapping is safe — all OpenAI models accept system

New¶

• Metadata preservation via ConversionContext (#105): All cross-format conversion paths now thread a shared ConversionContext(options={"metadata_mode": "preserve"}) through the full request→response lifecycle. This enables lossless round-trip conversion of provider-specific metadata fields (e.g. refusal, annotations, citations, extended usage details) that were previously dropped

Changed¶

• Bumped llm-rosetta to v0.5.0: Picks up developer role mapping fix in Responses converter, tool call ID prefix mapping for cross-format conversion, Google modality token handling, and gateway admin features (Oaklight/llm-rosetta v0.4.2–v0.5.0)

v3.0.0b14 (2026-04-11)¶

New¶

• --force-conversion mode: New CLI flag and config option (force_conversion: true) that forces all requests through the full llm-rosetta conversion pipeline (source → IR → target), even when source and target providers match. This bypasses same-format passthrough and enables parameter normalization (e.g. max_tokens → max_completion_tokens for OpenAI Chat), full IR validation on every request, and more detailed runtime error/warning logs from the conversion pipeline. Disabled by default to preserve existing passthrough behavior. Configurable via --force-conversion CLI flag, FORCE_CONVERSION environment variable, or force_conversion: true in config.yaml

Changed¶

• Bumped llm-rosetta to v0.4.1: Picks up Open Responses spec support, metadata preservation for lossless A→IR→A round-trip across all 4 providers, force_conversion parameter for full pipeline execution on same-provider requests, multimodal tool result conversion (#99), and vendored validate.py with type introspection caching (Oaklight/llm-rosetta v0.4.0–v0.4.1)

v3.0.0b13 (2026-04-08)¶

New¶

• Upstream error request/response dumping: All dispatch paths (passthrough, buffered, streaming, cross-format conversion) now dump request and response payloads on any upstream 4xx/5xx error to <config_dir>/error_dumps/ for diagnostics. Dumps include source/target provider context for easier debugging of cross-provider conversion issues. Registered as error-dump in argo-proxy logs collect --type (#100, #101)

Improved¶

• config list discovers all config file variants: Previously only checked 3 hardcoded paths. Now uses glob patterns to find all config files (e.g. config.ozan.yaml, config-public.yaml) across standard search directories

Fixed¶

• input_image type rejected by Anthropic in cross-provider conversion (#99): Multimodal tool results (e.g. matplotlib plots from MCP servers) containing OpenAI's input_image content type were passed through to Anthropic without conversion, causing 400 errors. Fixed upstream in llm-rosetta via multimodal tool result conversion across all 4 providers (Oaklight/llm-rosetta#92)
• StreamContext import path: Updated import to match llm-rosetta v0.3.0 module structure

Changed¶

• Bumped llm-rosetta to v0.3.1: Picks up multimodal tool result conversion (#99), StreamContext import path fix, and streaming reliability improvements (Oaklight/llm-rosetta v0.3.0–v0.3.1)

v3.0.0b12 (2026-04-03)¶

New¶

• Configurable Anthropic non-streaming stream mode (--anthropic-stream-mode): Controls how non-streaming requests to Anthropic upstream are handled. Three modes:
  • force (default): Always force streaming upstream and aggregate SSE events back into a non-streaming response. This is the same behavior as v3.0.0b9+ and avoids Anthropic's "Streaming is required for operations that may take longer than 10 minutes" bounce-back error
  • retry: Try the request as non-streaming first. If Anthropic returns the "streaming required" bounce-back (HTTP 500), automatically retry with forced streaming. Dumps the original request to <config_dir>/stream_retry_dumps/ for diagnostics
  • passthrough: Never force streaming, pass through as-is. Useful for debugging or when requests are known to be short
• Generalized logs collect CLI: argo-proxy logs collect now supports --type {leaked-tool,stream-retry,all} to collect different categories of diagnostic logs. Default is all. New log categories can be added by registering in _DIAGNOSTIC_LOG_TYPES
• config list subcommand: argo-proxy config list lists all config files found in standard search paths, marking the active one
• Unified /version endpoint: The HTTP /version endpoint now returns both stable and pre-release version info (matching CLI update check behavior), with latest_stable, latest_pre, update_commands (CLI + pip), and changelog URL

Improved¶

• CLI output cleanup: All CLI subcommand output (config, update, logs) now uses clean print() instead of log_info() with timestamps, producing cleaner terminal output
• Dual update commands: Version check surfaces (--version, startup banner, update check, /version endpoint) now show both argo-proxy update install (primary) and pip install --upgrade (alternative) as upgrade paths

Changed¶

• ANTHROPIC_STREAM_MODE environment variable: New env var to set the Anthropic stream mode without CLI flags. Accepts force, retry, or passthrough
• anthropic_stream_mode config field: Can also be set in config.yaml (persisted only when non-default). Placed in the "Upstream" config section
• Legacy code consolidation: All legacy-only modules (types/, tool_calls/, legacy endpoints, legacy utils) moved under _legacy/ top-level package. Embeddings passthrough extracted to endpoints/passthrough.py for universal mode. utils/models.py cleaned up — dead code removed, legacy-only functions moved to _legacy/utils/models.py

v3.0.0b11 (2026-03-31)¶

Improved¶

• CLI typo detection: Misspelled subcommands now show a helpful suggestion instead of silently falling through to the serve handler with the typo treated as a config file path (e.g. argo-proxy server → unknown command 'server'. Did you mean 'serve'?). Uses difflib.get_close_matches with cutoff 0.9; file-path-like arguments (containing ., /, \, or matching an existing file) are excluded from fuzzy matching
• Reduced default log history truncation: Verbose request logs now keep the last 3 conversation history items (down from 5) to reduce noise from large message entries

New¶

• max_log_history config option: Controls how many conversation history items are kept in verbose request logs. Default: 3. Set higher if you need more context in logs for debugging long conversations

v3.0.0b10 (2026-03-29)¶

Features¶

• File logging with gzip rotation: New log_to_file config option enables structured file logging to ~/.local/share/argoproxy/logs/ with automatic daily rotation and gzip compression of old files. Console output remains unchanged
• Diagnostic logging for streaming chunks: Verbose mode now logs the first upstream chunk and stream completion stats (chunk count, duration) for easier debugging

Improved¶

• Request logging — history truncation and Responses API support (@caidao22): Conversation history in logs is now truncated to the last 5 items (configurable via max_history_items) with a [... N earlier items omitted ...] placeholder. Added sanitization for Responses API input items (input_text/output_text truncation). Request summary now includes input=N for Responses API requests. log_converted_request gated behind verbose check to skip unnecessary deep copies
• Per-request user tagging in logs: When USERNAME_PASSTHROUGH=true, all dispatch log entries now include user=<identity> so operators can trace requests back to individual users for debugging. Uses contextvars to tag all log calls within a request's lifetime — covers [ORIGINAL], [CONVERTED], [UPSTREAM ERROR], [ModelRegistry], and all dispatch-level messages

Fixed¶

• Auth credential fallback in dispatch upstream headers: When clients (e.g. Codex CLI) send no auth headers, upstream requests were missing credentials entirely, causing 401 errors. Added _extract_client_credential() with provider-aware header priority (Anthropic prefers x-api-key, OpenAI prefers Authorization: Bearer) and _build_upstream_headers() with fallback_user parameter — without --username-passthrough, always uses config.user; with it, extracts from client request first, falls back to config.user
• Graceful client disconnect handling: Client disconnects during streaming no longer produce noisy tracebacks — ConnectionResetError is caught and logged cleanly
• Catch aiohttp ClientConnectionResetError in dispatch: aiohttp's ClientConnectionResetError (raised when writing to a closing transport) is not a subclass of stdlib ConnectionResetError, so it was falling through to the generic exception handler and logging as ERROR. Now caught explicitly alongside stdlib connection errors and logged as a warning

Changed¶

• Bumped llm-rosetta to v0.2.6: Picks up orphaned tool_call IR-level fix, parallel tool_call_index propagation, Responses streaming item_id tracking, non-function tool name preservation, orphaned tool_choice stripping, and stream event ordering fixes (Oaklight/llm-rosetta v0.2.6)

v3.0.0b9 (2026-03-29)¶

Features¶

• config init subcommand: Interactive config creation wizard, accessible via argo-proxy config init [path] [--force]. Supports custom file paths and overwrite confirmation for existing configs. Reuses the same interactive flow as first-time setup
• config env subcommand: Show or switch upstream ARGO API environment via argo-proxy config env [prod|dev|test]. Three environments available: production (apps.inside.anl.gov), development (apps-dev.inside.anl.gov, default), and test (apps-test.inside.anl.gov)

Changed¶

• ReadTheDocs links in CLI help: Each subcommand's --help now includes a Docs: URL pointing to the corresponding ReadTheDocs page
• Updated config.sample.yaml and configuration docs: Reorganized sample config with logical section grouping matching _format_config_yaml() output. Added upstream environment table (prod/dev/test URLs), conditional fields documentation, and removed legacy endpoint examples

Fixed¶

• Force streaming for Anthropic non-streaming requests: Anthropic API requires streaming for operations that may exceed 10 minutes, causing 500 errors ("Streaming is required for operations that may take longer than 10 minutes") for non-streaming requests from clients like Claude Code. Solved transparently at the proxy layer: all non-streaming requests targeting Anthropic are now sent with stream: true upstream, SSE events are aggregated into a complete Messages API response, and standard JSON is returned to the client. Covers both same-format passthrough (Anthropic → Anthropic) and cross-format conversion (e.g. OpenAI → Anthropic) paths
• config migrate produces canonical v3 format: Rewrote CLI config migrate to use the standard load_config → ArgoConfig.from_dict → to_persistent_dict → _format_config_yaml pipeline instead of raw dict manipulation. Migrated output now matches the format produced by a fresh create_config() init flow. Infers argo_base_url from legacy individual URL fields (argo_url, argo_stream_url, argo_embedding_url), drops stale/unknown fields (e.g. mode, num_workers), removes deprecated keys, and cleans up v3 configs with dirty derived fields. to_persistent_dict() now conditionally persists non-default optional fields (use_legacy_argo, skip_url_validation, native_openai_base_url, native_anthropic_base_url)
• Banner version check for pre-release versions: When the installed version is a pre-release (e.g. 3.0.0b7), the startup banner now compares against the latest pre-release on PyPI instead of only the latest stable version. Previously it would incorrectly show "(Latest)" even when a newer pre-release was available. Extracted shared get_pypi_versions() async function and deduplicated version-fetching logic between the banner display and update check handler

v3.0.0b8 (2026-03-27)¶

Features¶

• ARGO authentication warning detection: Detects the upstream AUTHENTICATION NOTICE FROM ARGO warning injected into responses when a username is not registered, and returns a 403 Forbidden error with code argo_auth_warning instead of forwarding the polluted content to the client. Covers all endpoint types (OpenAI, Anthropic, legacy ARGO) in both streaming and non-streaming modes
• Startup username validation: During initialization, argo-proxy now validates the configured username against the upstream ARGO API by making a test request. If the username is not registered, it rejects the input and re-prompts — invalid usernames can no longer slip through to runtime
• Interactive init flow improvements: First-time setup now prompts for the upstream base URL before username/port, validates upstream connectivity (GET /v1/models) immediately, and loops until a reachable URL is given. The verbose mode prompt is moved to the end, after all validations pass
• Config YAML formatting: Saved config files now use logical section grouping (Core settings, Upstream, Network & validation, Image processing) with comments and blank lines for readability

Changed¶

• Streaming passthrough buffering: Passthrough streaming handlers (native_openai, native_anthropic, dispatch) now buffer initial chunks (up to 2 KB or the first complete SSE event) before committing to the streaming response, enabling early detection of upstream auth warnings
• Added argo_auth_warning error code: New error code added to ResponseError literal type for programmatic detection of ARGO authentication issues
• Extended validate_user_async() transport utility: New async function in utils/transports.py that validates a username by sending a lightweight chat request and inspecting the response for authentication warnings
• Config persistence cleanup: save_config() now uses to_persistent_dict() which excludes runtime-derived fields (mode, native_openai_base_url, native_anthropic_base_url). New configs are created with config_version: "3". Config is saved to the user-specified path instead of always defaulting to ~/.config/argoproxy/config.yaml
• Removed tqdm progress bar from startup: URL validation no longer shows a progress bar — replaced with a simple log message. The redundant mode display block at the end of validate_config() is also removed (already shown in the banner)
• Refactored config.py into config/ subpackage: Split the ~1000-line monolithic config module into model.py (ArgoConfig dataclass), io.py (load/save/migrate), validation.py (user/port/URL checks), and interactive.py (user prompts and create_config)
• Refactored cli.py into cli/ subpackage: Split the ~1050-line monolithic CLI module into parser.py (argparse construction), display.py (banner and version check), and handlers.py (subcommand handlers)

v3.0.0b7 (2026-03-23)¶

Fixed¶

• Orphaned tool_calls cause 400 on strict upstreams: When a tool call is interrupted (e.g. user cancels mid-execution in Claude Code), the tool_calls entry remains in conversation history without a matching role: "tool" response. OpenAI and Anthropic APIs strictly reject this. Now fixed via fix_orphaned_tool_calls() from llm-rosetta — applied in both passthrough and cross-format paths for all three strict-pairing providers (OpenAI Chat, OpenAI Responses, Anthropic) (#82)
• Missing user field in cross-format requests: Cross-format IR round-trips produce a fresh request body that drops the user field, causing upstream auth/tracking to lose the user identity. Now injected automatically in both streaming and non-streaming conversion paths
• Google GenAI SDK auth via x-goog-api-key header: The Google GenAI SDK (used by Gemini CLI) sends API keys via the x-goog-api-key header, which was not extracted by _build_upstream_headers(). Now supported alongside Authorization: Bearer, x-api-key, and ?key= query parameter
• aiohttp 3.12 startup crash: TCPConnector.__init__() no longer accepts tcp_nodelay (removed upstream in aiohttp 3.12). Removed the parameter from OptimizedHTTPSession

Changed¶

• Bumped llm-rosetta to v0.2.5: Picks up full-stack Google GenAI camelCase support, cross-format image passthrough fixes, tool_call_id reconciliation, input_schema type default for parameterless tools, and built-in tool handling (Oaklight/llm-rosetta v0.2.4–v0.2.5)
• Refactored dispatch.py: Merged 4 identical SSE formatters into 2, extracted _write_sse_chunks() / _ensure_user_field() helpers, defined _STREAMING_HEADERS constant, added Callable type hint to _SSE_FORMATTERS
• Modernized typing imports: Replaced typing.List, Dict, Tuple, Set with Python 3.10+ built-in generics via ruff UP006 (525 auto-fixes across 31 files); removed typing_extensions.Literal in favor of typing.Literal
• Added ruff and ty configuration: pyproject.toml now includes [tool.ruff] (target-version py310, select E/F/UP, ignore UP007/E501) and [tool.ty] (python-version 3.10, unresolved-import ignore) sections
• Fixed 35 ty type-check diagnostics: Corrected AbstractResolver.resolve override signature, narrowed Optional types with assert guards and cast(), added missing Union members to return type annotations, fixed dict[str, str] → dict[str, Any] for log entries
• Cleaned up performance.py: Removed optimize_event_loop() (no-op defaults), fake tqdm progress bar, inspect.signature feature check, CPU-based connection scaling; simplified to fixed defaults with env var overrides (-80 lines)

v3.0.0b6 (2026-03-22)¶

Changed¶

• Bumped llm-rosetta to v0.2.3: Picks up tool schema sanitization across all converters (#80), $ref/$defs resolution (#80), streaming tool call argument accumulation fixes (#81), and sanitize_schema extraction to shared base module (#66)
• Updated sanitize_schema import path: dispatch.py now imports sanitize_schema from llm_rosetta.converters.base.tools instead of llm_rosetta.converters.openai_chat.tool_ops (follows upstream refactoring in Oaklight/llm-rosetta#66)

v3.0.0b5 (2026-03-22)¶

Fixed¶

• Same-format passthrough tool schema sanitization: Added _sanitize_tool_schemas() to sanitize tool parameter schemas in the passthrough path (when source_provider == target_provider). Upstreams like Vertex AI reject unsupported JSON Schema keywords even in passthrough mode (#76)
• Stream termination fallback: Added fallback in _convert_streaming() — when the upstream stream ends without sending a StreamEndEvent trigger (empty-choices chunk), synthesize termination events (message_delta, message_stop) so the client receives a complete Anthropic SSE event sequence (#79)
• Bumped llm-rosetta to v0.2.2: Picks up Anthropic content_block_stop fix and upstream preflight chunk handling (Oaklight/llm-rosetta#77)

v3.0.0b4 (2026-03-21)¶

Initial v3 beta release — complete architecture rewrite.

Major Features¶

• Universal API Gateway (llm-rosetta): Transformed argo-proxy from a single-format ARGO gateway proxy into a universal API translator. All 4 client API formats are now supported — users can talk to any upstream model using any client format:

  Client Format	Endpoint
  OpenAI Chat Completions	POST /v1/chat/completions
  OpenAI Responses API	POST /v1/responses
  Anthropic Messages	POST /v1/messages
  Google GenAI	POST /v1beta/models/{model}:generateContent

• Smart Upstream Routing: Claude models are automatically routed to the native Anthropic endpoint (avoiding tool call leakage on OpenAI-compat), while all other models (GPT, Gemini) go to the native OpenAI Chat endpoint.

• Same-Format Passthrough: When the client format matches the upstream format, requests and responses are piped through without conversion for maximum performance and fidelity.

• Cross-Format Conversion: When formats differ (e.g., Anthropic client + GPT model, or OpenAI client + Claude model), llm-rosetta handles bidirectional conversion via its Intermediate Representation (IR), for both streaming and non-streaming requests.

New Features¶

• Universal Dispatch Module (endpoints/dispatch.py): Central entry point for all API requests with model resolution, format detection, image preprocessing, auth header translation, and error handling.

• Cross-Format Auth Header Translation: Automatic conversion between x-api-key (Anthropic) and Authorization: Bearer (OpenAI) headers when routing across formats.

• Flexible Model Name Resolution: Enhanced _model_lookup_candidates() to accept all common model name formats:

  • Anthropic dash format: claude-sonnet-4-6
  • Argo dot format: argo:claude-sonnet-4.6
  • Internal IDs: claudesonnet46
  • Anthropic model IDs with dates: claude-sonnet-4-6-20250514

• Native Upstream Model List: Model registry now fetches from the native OpenAI /v1/models endpoint instead of the legacy ARGO models API.

• CLI Subcommands: Restructured CLI with hierarchical subcommands:

  • argo-proxy serve — start the server (default, backward compatible)
  • argo-proxy config {edit,validate,show,migrate} — manage configuration
  • argo-proxy logs collect — collect diagnostic logs
  • argo-proxy update {check,install} — check for and install updates from PyPI (supports --pre for pre-release)
  • argo-proxy models — list all available upstream models with their aliases, grouped by type (Chat/Embedding) and provider, with --json output support
  • --no-banner flag to suppress ASCII banner

• Simplified Base URL Configuration: Only argo_base_url is needed — native_openai_base_url and native_anthropic_base_url are automatically derived. Explicit overrides are still supported and trailing slashes are normalized.

• Config Migration (argo-proxy config migrate): Automatic migration of v1/v2 config files to v3 format with .bak backup.

• Mode-Aware Config Display: argo-proxy config validate and config show now display native endpoint URLs and mode: universal in v3 mode, or legacy ARGO gateway URLs in legacy mode.

• Native Endpoint Connectivity Check: URL validation at startup tests the native OpenAI /v1/models endpoint (GET) in universal mode, instead of the legacy ARGO chat/embedding POST endpoints.

Breaking Changes¶

• Native endpoints are now default: use_native_openai and use_native_anthropic config keys are deprecated and ignored. Native upstream routing is always active.
• Legacy mode is opt-in: Use --legacy-argo flag or use_legacy_argo: true in config to enable the old ARGO gateway pipeline.
• Removed --native-openai / --native-anthropic CLI flags: No longer needed since native mode is the default.
• Legacy endpoints moved to _legacy/: chat.py, completions.py, embed.py, responses.py, native_anthropic.py relocated to endpoints/_legacy/.

Deprecations¶

• tool_calls/handler.py, tool_calls/output_handle.py — replaced by llm-rosetta converters
• types/chat_completion.py, types/completions.py, types/embedding.py, types/responses.py — replaced by llm-rosetta IR types
• tool_calls/converters.py — removed entirely (unused)

Dependencies¶

• Added llm-rosetta as a core dependency for cross-format conversion

Documentation¶

• Comprehensive v3 documentation rewrite covering all pages
• New CLI Tools Guide with configuration instructions for Claude Code, Codex CLI, Aider, Gemini CLI, OpenCode, and generic SDK usage
• Updated endpoint docs with universal routing logic and all 4 API formats
• Rewrote CLI reference for subcommand-based interface
• Rewrote configuration guide for v3 format (argo_base_url, modes, migration)
• Simplified native passthrough pages with v3 notes pointing to new docs

v2.8.9 (2026-03-21)¶

Bug Fixes¶

• Leaked Tool Parser JSON Literals: Fixed ast.literal_eval failures when Claude emits JSON-style false/true/null instead of Python's False/True/None in leaked tool calls. Added word-boundary-aware regex repair strategy. (#70, contributed by @mvictoras)

v2.8.8.post1 (2026-03-07)¶

Bug Fixes¶

• Native Endpoint Model Passthrough: Fixed model name resolution on native endpoints (/v1/messages and native OpenAI passthrough) incorrectly falling back to the default model (gpt-4o) when encountering unrecognized model names. Added as_is parameter to resolve_model_name() so that native passthrough endpoints forward the caller's original model name to the upstream API unchanged. (#68, fixes #67)

v2.8.8 (2026-03-05)¶

Bug Fixes¶

• Native Anthropic Model Resolution: Fixed model name resolution on the native Anthropic endpoint (/v1/messages) to resolve bare model names (e.g., claude-4.6-sonnet) the same way as all other endpoints. Previously, only argo:-prefixed names were resolved, causing bare names to be forwarded as-is to the upstream API, resulting in 400 errors.

Features¶

• Configurable Connection Test Timeout: Added connection_test_timeout configuration option (default: 5 seconds) to control the per-URL timeout during startup connectivity validation. Previously hardcoded at 2 seconds, which was insufficient for high-latency connections such as layered SSH tunnels. Configurable via config.yaml.

v2.8.7.post1 (2026-02-22)¶

Bug Fixes¶

• Unclosed aiohttp Connector: Fixed connector_owner logic in get_upstream_model_list_async() and validate_api_async() that caused "Unclosed connector" warnings during startup. When no custom DNS resolver was provided, connector_owner=False was incorrectly set, preventing the auto-created TCPConnector from being closed with the session.

v2.8.7 (2026-02-22)¶

Features¶

• Flexible Model Name Resolution: Improve model name resolution flexibility - support slash separator, case-insensitive matching, and automatic argo: prefix handling (#65, contributed by @keceli)
• DNS Resolution Overrides (@michel2323): Custom DNS resolver (StaticOverrideResolver) implementing aiohttp.abc.AbstractResolver, similar to curl --resolve. Maps specific host:port combinations to IP addresses, enabling access to Argo API through SSH tunnels while preserving TLS certificate validation. Configured via resolve_overrides in config.yaml.
• Skip URL Validation (@michel2323): New skip_url_validation config option (also via SKIP_URL_VALIDATION env var) to skip startup connectivity checks — useful for non-interactive environments (CI/CD, containers) where network may be unavailable.

Refactoring¶

• urllib → aiohttp Migration: Migrated validate_api_async and get_upstream_model_list from synchronous urllib.request to native async aiohttp.ClientSession, with support for resolve_overrides parameter for custom DNS resolution.

Documentation¶

• Model Naming Guide: Rewrite models page as naming scheme guide
• CLI Docs Link: Add model naming docs link to CLI startup prompt
• DNS Resolution Overrides: Added DNS Resolution Overrides documentation page

Bug Fixes¶

• CLI Verbose Flag Override: CLI --verbose/--quiet flags now properly override the verbose setting in config.yaml (#64)
• Leaked tool parsing: Replaced quote-aware brace counting with a more robust candidate-end-position + repair strategy approach for parsing leaked Claude tool calls. This handles edge cases with nested quotes, escaped characters, and malformed output more reliably. (Contributed by n-getty)

v2.8.6 (2026-02-18)¶

Features¶

• Model List Refresh Endpoint: Added POST /refresh endpoint to reload the model list from upstream without restarting the instance
  • Returns before/after model statistics for easy comparison
  • Useful for picking up newly added upstream models without downtime

v2.8.5 (2026-02-15)¶

Bug Fixes¶

• Native Anthropic Tool Validation (@caidao22): Fixed incorrect input format for tool validation on native Anthropic endpoint (#62)
  • Added input_format parameter to handle_tools/handle_tools_native to correctly parse tools arriving via the native Anthropic endpoint
  • Allowed ToolChoice._handle_anthropic to accept string shorthand values (auto, any, none)
  • Short-circuit tool conversion when input format already matches target model type

v2.8.4 (2026-02-14)¶

Refactor¶

• Base URL Consolidation: Refactored URL configuration to use a single argo_base_url as the root for all endpoint URLs
  • Introduced argo_base_url configuration field — all endpoint URLs (chat, stream, embed, models, native OpenAI) are now derived from this single base URL
  • Simplified _argo_dev_base and _argo_prod_base to root URLs (e.g., https://apps-dev.inside.anl.gov/argoapi) instead of full API paths
  • Individual endpoint URLs (argo_url, argo_stream_url, etc.) can still override the derived values for backward compatibility
  • Added config_version field for tracking configuration format; legacy configs without this field will log a migration hint
  • Updated config.sample.yaml with the new recommended configuration format

Features¶

• Native Anthropic Passthrough Mode: Added --native-anthropic CLI flag and use_native_anthropic configuration option for Anthropic-compatible API passthrough, similar to the existing native OpenAI passthrough
  • Exposes /v1/messages endpoint for direct Anthropic API compatibility
  • Supports both streaming and non-streaming message requests
  • Added native_anthropic_base_url configuration for customizing the upstream Anthropic endpoint
  • URL is derived from argo_base_url by default for consistency with other endpoints

Improvements¶

• Examples Directory Reorganization: Restructured examples into anthropic/, openai/, and argo/ subdirectories
  • Each provider directory contains sdk_based/ and rest_based/ subdirectories
  • Added Anthropic function calling examples (SDK and REST)
  • Added Anthropic image comprehension examples with base64 encoding
  • Unified API key handling across all examples using API_KEY environment variable

v2.8.3 (2026-02-08)¶

Features¶

• Improved Claude Leaked Tool Call Parsing: Added new robust parser for extracting leaked tool calls from Claude model responses

  • New leaked_tool_parser module with quote-aware brace counting for accurate dict boundary detection
  • Handle Anthropic content array format (mixed text+tool_use blocks)
  • Extract ALL leaked tool calls, not just the first one
  • Proper handling of braces inside strings (code snippets)
  • Support for escaped quotes in tool arguments
  • Added comprehensive unit tests (24 tests)

• Attack Logging Module: Added dedicated security logging for malicious HTTP requests

  • Create attack_logger.py module for security logging
  • Log concise warning messages with attack type and source IP
  • Save detailed attack logs to {config_dir}/attack_logs/ directory
  • Classify attacks: struts2_ognl, directory_traversal, ssti_probe, etc.
  • Store logs as compressed JSONL files organized by date

Bug Fixes¶

• Streaming UTF-8 Handling: Fixed incomplete UTF-8 sequences in streaming responses
  • Add StreamDecoder utility class to safely decode UTF-8 byte streams where multi-byte characters may be split across network packets
  • Update chat, completions, and responses endpoints to use StreamDecoder
  • Fixes intermittent ValueError: 'utf-8' codec can't decode byte errors

v2.8.2 (2026-01-31)¶

Refactor¶

• Logging System Overhaul: Replaced loguru with standard library logging for better compatibility; centralized request logging utilities with improved formatting and log levels

Bug Fixes¶

• Large Image Payload Support: Increased aiohttp client_max_size to 100MB to handle large image uploads

• Token Counting Accuracy (@Neil Getty): Fixed missing token counts for tool-related content

  • Added token counting for tool_calls and tools definitions
  • Improved recursive text extraction for nested content
  • Fixed count_tokens to handle None values gracefully

• Claude Max Tokens Fix: Fixed 500 upstream errors caused by excessive max_tokens in non-streaming tool call requests

  • Enforced max_tokens limit for Claude models to prevent issues with clients (like OpenCode) requesting 32k+ tokens with large tool definitions

Improvements¶

• Centralized pseudo_stream handling and improved logging consistency across endpoints
• Improved upstream error handling across all endpoints with better error messages and recovery

v2.8.1 (2026-01-24)¶

Features¶

• Leaked Tool Call Detection and Collection: Added comprehensive leaked tool call debugging features
• Added --enable-leaked-tool-fix CLI flag to enable AST-based leaked tool call detection and automatic fixing
• Added --collect-leaked-logs CLI flag to collect all leaked tool call logs into timestamped tar.gz archive
• Implemented automatic log directory management with compression when exceeding 50MB threshold
• Enhanced ToolInterceptor to support request data logging and conditional leaked tool fixing

• Added comprehensive logging and user guidance for sharing collected logs with maintainers

• Streaming Completions Support: Added full streaming support for legacy completions endpoint

• Implemented pseudo-streaming and real streaming handlers for completions endpoint
• Added usage chunk generation for completions API type
• Integrated token counting for accurate usage reporting in streaming mode

• Refactored transform functions to support both sync and async OpenAI compatibility

• Streaming Usage Stats (@Neil Getty): Implemented streaming usage statistics for pseudo stream responses in chat completions endpoint

• Accumulate total response content during pseudo chunk generation
• Calculate completion tokens using count_tokens_async function
• Send usage statistics chunk at end of stream with prompt/completion/total tokens
• Maintain compatibility with existing streaming infrastructure using SSE

Bug Fixes¶

• Consistent Chunk IDs: Ensured consistent chunk IDs across streaming responses
• Added chunk_id parameter to transform_chat_completions_streaming_async function
• Generated shared ID for all chunks in a stream to maintain consistency
• Passed shared chunk_id to all streaming chunks including tool calls and usage

• Ensured only first chunk gets role: assistant by tracking is_first_chunk state

• Role Field Handling: Fixed None value handling for role field in chat completion types

• Made role field optional to prevent validation errors when None is provided

• Maintained default value of "assistant" for backward compatibility

• First Chunk Handling: Improved first chunk handling in streaming responses

• Added is_first_chunk parameter to include assistant role in initial delta
• Fixed tool calls handling in streaming transformation

• Collected total response content for accurate token counting

• Claude Tool Calls (@Neil Getty): Fixed tool calls handling in pseudo stream responses

• Set finish_reason to "tool_calls" when tool calls are present
• Added robust parsing for leaked tool calls in Claude text content using ast.literal_eval
• Implemented balanced dictionary detection to extract tool calls from text
• Clean up text content by removing leaked tool call strings

Refactor¶

• Usage Counting: Unified usage calculation logic across all endpoints
• Extracted usage calculation logic into shared calculate_completion_tokens_async function
• Created unified create_usage factory function for different API types
• Implemented generate_usage_chunk for streaming responses
• Removed duplicate token counting code from chat, completions, embed, and responses endpoints

• Standardized usage object creation across all API formats (chat_completion, completion, response, embedding)

• Development Scripts: Migrated OpenAI example scripts to use environment variables

• Updated scripts to load environment variables from .env files
• Improved configuration management and flexibility

Enhancements¶

• Tool Call Examples: Enhanced OpenAI chat completions example with tool call support
• Added TOOL_CALL environment variable to switch between regular chat and tool call modes
• Implemented tool call payload with get_stock_price function example

• Maintained backward compatibility with existing message-based chat functionality

• Test Results: Added OpenAI function call streaming test results for comparison

• Created new test result files for OpenAI function call chunk streaming
• Enabled comparison between OpenAI and local streaming implementations

Maintenance¶

• Documentation Cleanup: Removed legacy Sphinx and ReadTheDocs documentation system
• Completely removed the legacy documentation system built with Sphinx and configured for ReadTheDocs
• Deleted entire docs/ directory including Sphinx configuration, source files, and build scripts

• Simplified build process and overall maintenance

• Project Documentation: Updated project name and badge formatting in README

v2.8.0 (2026-01-02)¶

Major Features¶

• Gemini Native Function Calling: Added full native function calling support for Google Gemini models
• Model Statistics Enhancement: Enhanced model registry with detailed family breakdown and alias counting
• Comprehensive Test Suite: Added complete test coverage for all API endpoints

New Features¶

• Gemini Function Calling:

  • Native tool calling support for Google/Gemini models in chat completions
  • Google-specific tool call format conversion and processing
  • Parallel to sequential tool call conversion for Gemini API compatibility
  • Tool call ID generation and argument serialization for Google format
  • Enhanced tool choice handling with proper string/dict conversion

• Model Registry Enhancements:

  • Added get_model_stats() method for detailed model statistics
  • Model family classification (OpenAI, Anthropic, Google, unknown)
  • Comprehensive model registry display with family breakdown
  • Chat and embedding model statistics with alias counts
  • Tree-structured logging for better readability

• Testing Infrastructure:

  • Complete test suite for chat completions, embeddings, and function calling
  • Tests for streaming, temperature control, conversation context
  • Function calling tests for single and multiple tool scenarios
  • Legacy completions endpoint testing with various parameters

Bug Fixes¶

• Embedding Index: Fixed embedding index assignment to use sequential numbering instead of hardcoded 0

Improvements¶

• Tool Call Processing: Enhanced tool call conversion between OpenAI and Google formats
• Development Scripts: Added comprehensive Gemini function calling examples and test scripts
• Model Display: Improved startup model registry display with family breakdown

v2.7.10 (2026-01-02)¶

Major Features¶

• Gemini Native Function Calling: Added full native function calling support for Google Gemini models
• Model Statistics Enhancement: Enhanced model registry with detailed family breakdown and alias counting
• Comprehensive Test Suite: Added complete test coverage for all API endpoints

New Features¶

• Gemini Function Calling:

  • Native tool calling support for Google/Gemini models in chat completions
  • Google-specific tool call format conversion and processing
  • Parallel to sequential tool call conversion for Gemini API compatibility
  • Tool call ID generation and argument serialization for Google format
  • Enhanced tool choice handling with proper string/dict conversion

• Model Registry Enhancements:

  • Added get_model_stats() method for detailed model statistics
  • Model family classification (OpenAI, Anthropic, Google, unknown)
  • Comprehensive model registry display with family breakdown
  • Chat and embedding model statistics with alias counts
  • Tree-structured logging for better readability

• Testing Infrastructure:

  • Complete test suite for chat completions, embeddings, and function calling
  • Tests for streaming, temperature control, conversation context
  • Function calling tests for single and multiple tool scenarios
  • Legacy completions endpoint testing with various parameters

Bug Fixes¶

• Model Fetching: Enhanced error handling in model fetching with detailed logging for HTTP, URL, JSON, and unknown errors
• API Compatibility: Added support for both old and new API formats in Model class with backward compatibility
• Network Reliability: Added request timeouts and improved error recovery with fallback to built-in model list
• Embedding Index: Fixed embedding index assignment to use sequential numbering instead of hardcoded 0

Improvements¶

• Error Handling: Improved error logging with detailed HTTP, URL, and JSON error messages
• API Format Detection: Added automatic detection and logging of API format versions
• Tool Call Processing: Enhanced tool call conversion between OpenAI and Google formats
• Development Scripts: Added comprehensive Gemini function calling examples and test scripts

v2.7.9 (2025-12-14)¶

Major Features¶

• Native OpenAI Passthrough Mode: Added complete native OpenAI endpoint passthrough functionality
• Dynamic Version Configuration: Updated build system to use dynamic version configuration from argoproxy.__version__
• Enhanced Startup Experience: Added ASCII banner with version information and styled model registry display

New Features¶

• Native OpenAI Mode:

  • Added use_native_openai configuration flag and --native-openai CLI option
  • Implemented pure passthrough for chat/completions, completions, and embeddings endpoints
  • Added support for model name mapping even in native mode (argo:gpt-4o aliases)
  • Integrated tool call processing for native OpenAI mode with model family detection
  • Added image processing support with automatic URL download and base64 conversion

• Startup Enhancements:

  • Added ASCII banner for Argo Proxy with version information
  • Implemented styled model registry display with model count
  • Added startup banner with update availability check
  • Enhanced configuration mode logging with visual indicators

• Tool Call Improvements:

  • Added streaming response handling with data prefix parsing in function calling examples
  • Implemented [DONE] message detection for stream termination
  • Added environment variable support for stream control in examples

Improvements¶

• Configuration Management:

  • Added native OpenAI mode warnings and status indicators
  • Improved mode logging logic to properly handle native vs standard mode
  • Enhanced logging consistency across modules with standardized formatting

• Documentation:

  • Added comprehensive native OpenAI passthrough documentation
  • Updated feature comparison tables and endpoint availability information
  • Added tool call processing documentation with model compatibility matrix
  • Included troubleshooting section with common error messages and solutions

• Build System:

  • Migrated from static to dynamic version configuration
  • Updated pyproject.toml to read version from argoproxy.__version__ attribute

Bug Fixes¶

• Fixed streaming response handling in function calling chat examples
• Improved error handling for non-JSON data in streams
• Enhanced model name resolution formatting for better readability

v2.7.8 (2025-11-07)¶

New Features¶

• Image Processing Enhancement: Added support for JPEG (JPG), PNG, WebP, and GIF image formats
• Image URL Conversion: Implemented automatic conversion from image URLs to base64
• Image Payload Control: Added enable_payload_control configuration option for image payload size and concurrency control
• Image Format Validation: Added image format and magic bytes validation functionality

Improvements¶

• Improved image processing payload size handling logic
• Optimized image content type preservation in payload
• Enhanced base64 truncation logging functionality
• Added Pillow dependency for image processing support

Documentation Updates¶

• Added vision and image input usage examples
• Added comprehensive base64 usage documentation and examples
• Updated version requirement documentation

v2.7.8a0 (2025-11-05)¶

Pre-release Version¶

• Pre-release testing version for image processing functionality

v2.7.7 (2025-07-23)¶

Major Features¶

• Tool Call System Refactor: Complete rewrite of tool call processing system with multi-provider compatibility
• Native Tool Support: Enabled native tool processing for OpenAI and Anthropic models
• Cross-Provider Conversion: Implemented tool call format conversion between OpenAI and Anthropic
• Model-Specific Prompts: Implemented specific prompt skeletons for different model families

New Features¶

• Added pseudo_stream_override flag for streaming response control
• Implemented username passthrough functionality
• Added message content normalization logic
• Added user message checks for Gemini and Claude models

Improvements¶

• Enhanced tool call serialization handling
• Improved model determination logic
• Optimized input validation and error handling
• Updated streaming mode descriptions and tool call documentation

v2.7.6 (2025-07-17)¶

New Features¶

• Real Streaming: Added real_stream functionality for testing
• Configuration Enhancement: Added configuration diff display functionality
• Base URL Configuration: Support for base URL configuration
• Network Troubleshooting: Added network troubleshooting guidance

Improvements¶

• Renamed fake_stream to pseudo_stream for better clarity
• Enhanced server shutdown process
• Improved error handling and streaming logic
• Optimized configuration loading and URL construction

Documentation Updates¶

• Updated tool call documentation
• Improved toolregistry documentation with examples and integration instructions
• Added detailed documentation links

v2.7.5 (2025-07-14)¶

Major Features¶

• Function Calling Support: Added function calling functionality to chat completion endpoint
• Tool Call Framework: Implemented complete tool call processing framework
• Multi-Model Support: Extended support for more chat and embedding models

Technical Improvements¶

• Added Pydantic models for response structuring
• Implemented OpenAI compatibility enhancements including usage tracking
• Refactored type system with new completion and usage types

v2.7.5.post1 (2025-07-14)¶

Patch Version¶

• Fixed minor issues related to function calling
• Improved dependency management

v2.7.5.alpha1 (2025-06-30)¶

Pre-release Version¶

• Alpha testing version for function calling functionality

v2.7.4.post2 (2025-06-30)¶

Patch Version¶

• Tool Call Feature Development: Extensive tool call related feature development and refactoring
• Added function calling examples and tool choice types
• Implemented tool call templates and conversion functionality
• Enhanced streaming tool call support
• Improved tool call processing logic

v2.7.4.post1 (2025-06-28)¶

Patch Version¶

• Added model availability timeout input functionality
• Improved URL validation efficiency
• Simplified model availability handling
• Enhanced configuration loading verbosity control
• Temporarily disabled streamchat endpoint

v2.7.4 (2025-06-27)¶

Improvements¶

• Optimized streaming processing
• Enhanced error handling mechanisms
• Improved configuration management

v2.7.3 (2025-06-26)¶

New Features¶

• Added multi-entry processing support
• Improved message separation logic
• Added message deduplication and merging functionality

v2.7.2.post1 (2025-06-26)¶

Patch Version¶

• Removed model-specific transformation functionality
• Simplified chat processing logic

v2.7.2 (2025-06-25)¶

Improvements¶

• Updated Claude model configuration
• Optimized model pattern matching
• Improved version handling and help formatting

v2.7.1.post1 (2025-06-20)¶

Patch Version¶

• Fixed embedding endpoint input processing issues
• Improved request data preparation logic

v2.7.0.post2 (2025-06-17)¶

Patch Version¶

• Fixed case-sensitive header comparison issue in streaming responses
• Updated logger configuration

v2.7.0.post1 (2025-06-16)¶

Patch Version¶

• Added version check functionality
• Enhanced version endpoint response
• Simplified install command message

v2.7.0 (2025-06-15)¶

Architecture Refactor¶

• Response Endpoint: Added brand new /v1/responses endpoint
• SSE Support: Implemented Server-Sent Events (SSE) streaming
• Event-Driven: Introduced event-driven response workflow
• Type System: Complete refactor of type system using Pydantic models

New Features¶

• Added streaming and non-streaming response conversion functions
• Implemented response text completion event handling
• Added multiple OpenAI client example scripts

Improvements¶

• Optimized import statement organization
• Enhanced endpoint compatibility
• Improved error handling and logging

v2.6.1 (2025-06-11)¶

Improvements¶

• Updated configuration file search order and usage details
• Improved CLI usage and option formatting
• Removed num-worker parameter

v2.6.0 (2025-06-11)¶

Server Migration¶

• Sanic to aiohttp Migration: Complete rewrite of server implementation
• Logging System: Replaced Sanic logger with loguru
• Async Processing: Improved async JSON parsing and request handling

Configuration Improvements¶

• Removed timeout logic, simplified configuration
• Updated configuration validation and path handling
• Improved API validation workflow

Documentation Updates¶

• Updated installation instructions
• Added version badges and installation guide
• Improved timeout documentation and examples

v2.5.1 (2025-05-30)¶

Improvements¶

• Updated documentation and configuration file search order
• Improved CLI usage instructions
• Optimized editor selection logic

v2.5.1-alpha (2025-05-30)¶

Pre-release Version¶

• Alpha testing version for v2.5.1

v2.5.0 (2025-05-29)¶

CLI Refactor¶

• New CLI Interface: Replaced Sanic server with CLI interface
• Configuration Management: Enhanced configuration handling and validation
• Editor Integration: Added --edit option to open configuration in editor

Package Management¶

• Added Makefile for package management
• Updated pyproject.toml configuration

Feature Enhancements¶

• Added version command
• Improved type hints and error handling
• Enhanced log level configuration

v2.4 (2025-04-20)¶

Model Support¶

• Updated model aliases and Argo tags
• Simplified model availability import and usage
• Unified model mapping architecture

Improvements¶

• Enhanced model parsing logic
• Improved model handling compatibility
• Added unified prompt token calculation

v2.3 (2025-04-06)¶

Embedding Feature Enhancement¶

• Improved OpenAI response structure
• Enhanced prompt data processing
• Added OpenAI compatible response conversion
• Added token counting functionality
• Added tiktoken encoding and tools

v2.2 (2025-03-31)¶

New Features¶

• Added scripts for syncing GitHub and GitLab repositories
• Centralized API validation logic
• Added API validation functions

Improvements¶

• Prevented invalid username selection
• Improved setup instructions and configuration tables
• Enhanced documentation structure and content

v2.1 (2025-03-12)¶

Improvements¶

• Replaced problematic prompts with new complete stream implementation
• Disabled system message to user message conversion
• Added Argo stream URL to configuration
• Updated models and simplified chat examples

v2.0 (2025-03-11)¶

Major Refactor¶

• Project Structure: Reorganized entire project directory structure
• Modularization: Split functionality into independent endpoint modules
• Type Safety: Added comprehensive type hints

New Features¶

• Added embedding endpoint support
• Implemented model availability checking
• Added status and extra endpoints

Compatibility¶

• Improved OpenAI API compatibility
• Added model remapping functionality
• Enhanced error handling

v1.1 (2025-01-06)¶

Improvements¶

• Added model-specific system message conversion functionality
• Improved o1-preview model handling

v1.0 (2024-12-29)¶

Initial Stable Version¶

• Basic Proxy Functionality: Implemented basic ARGO API proxy
• OpenAI Compatibility: Provided OpenAI API format compatibility
• Chat Completion: Supported chat completion endpoint
• Streaming: Basic streaming response support

Core Features¶

• Configuration file support
• Basic logging
• Example scripts
• Model remapping and default settings

v0.3.0 (2024-12-12)¶

Modularization Improvements¶

• Completion Module: Added completion module and enhanced chat routing
• Compatibility: Ensured prompts are in list format in proxy_request
• Refactor: Separated chat and embedding modules, updated configuration

v0.2.0 (2024-12-02)¶

Infrastructure Improvements¶

• API Testing: Tested API URLs and increased Gunicorn timeout
• Bug Fixes: Fixed cases where prompts are lists
• Docker Optimization: Optimized Dockerfile, enhanced configuration
• Configuration Files: Updated Dockerfile, added config.yaml, modified app.py

v0.1.0 (2024-12-01)¶

Initial Release¶

• Response Handling Refactor: Refactored response handling and added prompt token counting
• OpenAI Compatibility: Enhanced proxy OpenAI compatibility and error handling
• Basic Endpoints: Added new endpoints
• Project Initialization: Initialized gitignore, local build support
• Core Functionality: Argo proxy ready, requires VPN connection

Version Notes¶

• Major Version: Indicates major architectural changes or incompatible updates
• Minor Version: Indicates new feature additions or important improvements
• Patch Version: Indicates bug fixes and minor improvements
• Suffix Versions: Such as .post1, .alpha1 etc. indicate patches or pre-release versions

Contributing¶

If you find any issues or have suggestions for improvements, please submit an issue or submit a pull request.