kolega-code 0.37.0__tar.gz → 0.38.0__tar.gz

{kolega_code-0.37.0 → kolega_code-0.38.0}/CHANGELOG.md

@@ -6,6 +6,48 @@ This project uses GitHub Releases for detailed generated release notes. This fil
 
 ## Unreleased
 
+## 0.38.0 - 2026-09-11
+
+### Added
+
+- **DeepSeek V4.1 Flash (released).** The finished model is served under
+  DeepSeek's canonical `deepseek-flash` id and is available as **DeepSeek V4.1
+  Flash** in the model picker, with vision and `none`/`low`/`high`/`max`
+  reasoning effort (default `high`). Verified live: 1,048,576-token context and a
+  393,216-token API output ceiling. DeepSeek's Responses API ignores the built-in
+  `web_search` tool, so this model does not advertise hosted web search.
+
+### Changed
+
+- **DeepSeek now defaults to V4.1 Flash.** `deepseek-flash` is the default model
+  (and first entry in the picker) for the DeepSeek provider, matching DeepSeek's
+  own guidance that V4.1 Flash supersedes V4 Pro on performance, cost, speed, and
+  total time. `deepseek-v4-pro` remains selectable; DeepSeek routes it to V4.1
+  Flash from September 14, 2026.
+
+### Removed
+
+- **Retired DeepSeek flash ids.** DeepSeek only serves two models now — V4.1 Flash
+  and V4 Pro — so Kolega's DeepSeek catalog lists exactly those two. The retired
+  `deepseek-v4-flash` and `deepseek-v4-flash-vision-exp` entries are gone (they
+  named nothing distinct: DeepSeek routes both to V4.1 Flash). Settings, sessions,
+  and per-role overrides that still name one of them fall back to the DeepSeek
+  default (`deepseek-flash`); naming one explicitly (`--model`,
+  `KOLEGA_CODE_MODEL`) fails with a "not available for" error.
+- **DeepSeek V4.1 Flash preview id.** The expired temporary
+  `deepseek-v4.1-flash-expires-on-0910` id is no longer selectable. Settings and
+  per-role overrides that still name it fall back to the DeepSeek default
+  (`deepseek-flash`); naming it explicitly (`--model`, `KOLEGA_CODE_MODEL`) fails
+  with a "not available for" error. Switch to `deepseek-flash`, which DeepSeek
+  serves the same model under.
+
+### Fixed
+
+- **Hosted web search no longer offered where DeepSeek stopped serving it.**
+  DeepSeek silently removed server-side `web_search` with the V4.1 release, so the
+  flash models' `/web` toggle was advertising a capability that no longer exists.
+  `deepseek-v4-pro` still executes hosted search and keeps it.
+
 ## 0.37.0 - 2026-09-09
 
 ### Added

{kolega_code-0.37.0 → kolega_code-0.38.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: kolega-code
-Version: 0.37.0
+Version: 0.38.0
 Summary: Multi-agent coding in the terminal
 Project-URL: Homepage, https://kolega-ai.github.io/kolega-code/
 Project-URL: Documentation, https://kolega-ai.github.io/kolega-code/

{kolega_code-0.37.0 → kolega_code-0.38.0}/kolega_code/__init__.py

@@ -174,4 +174,4 @@ __all__ = [
     "parse_git_status_output",
 ]
 
-__version__ = "0.37.0"
+__version__ = "0.38.0"

{kolega_code-0.37.0 → kolega_code-0.38.0}/kolega_code/agent/__init__.py

@@ -51,4 +51,4 @@ __all__ = [
     "ToolExtension",
 ]
 
-__version__ = "0.37.0"
+__version__ = "0.38.0"

{kolega_code-0.37.0 → kolega_code-0.38.0}/kolega_code/agent/compression.py

@@ -157,7 +157,7 @@ class HistoryCompressor:
     # Don't bother summarizing a trivially short prefix.
     MIN_PREFIX_TO_SUMMARIZE = 3
     # Cap the summary length. The prompt targets ~600 words, but reasoning
-    # models (deepseek-v4-flash Responses, etc.) can consume the entire budget
+    # models (DeepSeek Responses models, etc.) can consume the entire budget
     # on chain-of-thought and leave no room for text output, producing an empty
     # summary. A generous ceiling gives reasoning headroom while still staying
     # well under the model's full completion budget.

{kolega_code-0.37.0 → kolega_code-0.38.0}/kolega_code/cli/provider_registry.py

@@ -70,10 +70,8 @@ MODEL_LABELS: dict[str, str] = {
     "kimi-for-coding": "Kimi for Coding",
     "kimi-for-coding-highspeed": "Kimi for Coding (HighSpeed)",
     # DeepSeek
-    "deepseek-v4.1-flash-expires-on-0910": "DeepSeek V4.1 Flash (Preview, expires Sep 10)",
+    "deepseek-flash": "DeepSeek V4.1 Flash",
     "deepseek-v4-pro": "DeepSeek V4 Pro",
-    "deepseek-v4-flash": "DeepSeek V4 Flash",
-    "deepseek-v4-flash-vision-exp": "DeepSeek V4 Flash Vision (Exp)",
     # Z.AI (GLM Coding Plan)
     "glm-5.3": "GLM-5.3",
     "glm-5.3-flash": "GLM-5.3 Flash",
@@ -164,7 +162,7 @@ MODEL_LABELS: dict[str, str] = {
 # "available set is everything, default pick is curated" split.
 PROVIDER_DEFAULT_MODEL: dict[ModelProvider, str] = {
     ModelProvider.MOONSHOT: "kimi-k3",
-    ModelProvider.DEEPSEEK: "deepseek-v4-pro",
+    ModelProvider.DEEPSEEK: "deepseek-flash",
     ModelProvider.ZAI: "glm-5.3",
     ModelProvider.KIMI_CODING: "kimi-for-coding",
     ModelProvider.ANTHROPIC: "claude-opus-5",
@@ -189,7 +187,7 @@ PROVIDER_DEFAULT_MODEL: dict[ModelProvider, str] = {
 UI_DEFAULT_PROVIDER = ModelProvider.MOONSHOT.value
 UI_DEFAULT_MODEL = "kimi-k3"
 MOONSHOT_K26_MODEL = "kimi-k2.6"
-DEEPSEEK_DEFAULT_MODEL = "deepseek-v4-pro"
+DEEPSEEK_DEFAULT_MODEL = "deepseek-flash"
 
 
 @dataclass(frozen=True)

{kolega_code-0.37.0 → kolega_code-0.38.0}/kolega_code/llm/providers/deepseek_responses.py

@@ -1,6 +1,6 @@
 """DeepSeek provider that speaks the **Responses API** (``/responses``).
 
-The whole DeepSeek catalog (``deepseek-v4-pro`` and ``deepseek-v4-flash``) speaks
+The whole DeepSeek catalog (``deepseek-flash`` and ``deepseek-v4-pro``) speaks
 the Responses API, so the ``deepseek`` provider routes here wholesale from
 ``LLMClient._provider_class`` (no model-name branching). DeepSeek models hosted on
 other providers (Fireworks/OpenRouter/Ollama-Cloud) still use Chat Completions via
@@ -49,7 +49,7 @@ from .models import GenerationParams
 from .responses_common import ResponsesProviderBase
 
 DEFAULT_BASE_URL = "https://api.deepseek.com"
-DEFAULT_MODEL = "deepseek-v4-flash"
+DEFAULT_MODEL = "deepseek-flash"
 
 
 class DeepSeekResponsesProvider(ResponsesProviderBase):

{kolega_code-0.37.0 → kolega_code-0.38.0}/kolega_code/llm/specs/accessors.py

@@ -203,9 +203,7 @@ DEEPSEEK_WIRE_OUTPUT_CAP = 64000
 
 # Routes that use the catalog output budget without the legacy Pro/chat clamp.
 _UNCLAMPED_DEEPSEEK_ROUTES = {
-    ("deepseek", "deepseek-v4-flash"),
-    ("deepseek", "deepseek-v4-flash-vision-exp"),
-    ("deepseek", "deepseek-v4.1-flash-expires-on-0910"),
+    ("deepseek", "deepseek-flash"),
 }
 
 

kolega_code-0.38.0/kolega_code/llm/specs/catalog/deepseek.py

@@ -0,0 +1,99 @@
+from kolega_code.llm.specs.types import ThinkingEffortSpec
+
+# max_completion_tokens is the MEASURED per-response output ceiling, not
+# DeepSeek's published 384000. Probed 2026-08-03 by
+# tests/agent/llm/test_deepseek_output_cap_live.py:
+#   - deepseek-v4-pro (chat), uncapped: the server cuts the stream at exactly
+#     65536 output tokens, reported honestly as finish_reason="length".
+#   - deepseek-v4-flash (Responses), uncapped: no hard cut observed — the model
+#     SELF-CENSORS ~7.4k tokens in ("response would exceed the maximum output
+#     length allowed for this chat interface") with a clean status="completed".
+#     With an explicit max_output_tokens=64000 it streams the full 64000 and
+#     reports the cap hit honestly (status="incomplete"/max_output_tokens).
+#   - Over-ceiling caps (384000) are accepted silently on both paths.
+# flash is 384000 (corrected from 65536, which was pro's measured chat ceiling
+# assumed to apply here): those probes only measured visible output, while a
+# reasoning-heavy call was later measured running to 112990 tokens in a single
+# Responses call.
+# pro never sends this value raw: the Responses path clamps to
+# DEEPSEEK_WIRE_OUTPUT_CAP=64000 (specs/accessors.py) so the client cap fires
+# before the server ceiling and truncation is reported honestly. flash passes
+# through.
+#
+# INSERTION ORDER IS LOAD-BEARING: the first entry is what the Settings model
+# picker shows first and what a provider switch falls back to
+# (ui_model_options(provider)[0]); it must stay in sync with
+# PROVIDER_DEFAULT_MODEL[DEEPSEEK] in cli/provider_registry.py.
+DEEPSEEK_SPECS = {
+    # Released DeepSeek Flash (V4.1), served under the canonical `deepseek-flash`, and
+    # the provider default. Verified live on /responses 2026-09-10: the retired preview
+    # ID (`deepseek-v4.1-flash-expires-on-0910`) as well as the older `deepseek-v4-flash`
+    # and `deepseek-v4-flash-vision-exp` aliases all now answer as
+    # `model="deepseek-flash"`, and `/models` lists exactly `deepseek-flash` and
+    # `deepseek-v4-pro`; a versioned `deepseek-v4.1-flash` is rejected ("The supported
+    # API model names are ..."). Vision (image input), function calls, and
+    # none/low/high/max reasoning effort all work on this route.
+    # NO hosted web search: DeepSeek removed it silently with the V4.1 release. Their
+    # Responses guide said "web_search / web_search_2025_08_26: Supported, executed on
+    # the server side" and listed {"type": "web_search"} under tool_choice on
+    # 2026-09-09; by 2026-09-10 the same rows read "web_search / file_search /
+    # code_interpreter / computer_use / mcp / other built-in tools: Ignored", with
+    # web_search dropped from tool_choice and its streaming events deleted. Nothing on
+    # /updates, /news, the docs home, or the FAQ mentions the removal. Live probes on
+    # 2026-09-10 returned zero web_search_call items while the model answered that it
+    # has no such tool (at effort=high it emitted bogus tool-call XML as text instead).
+    # Requests with the tool declared are accepted and silently do nothing, so Kolega
+    # must not advertise the /web toggle here. Replayed web_search_call items from
+    # older requests are still restored into context, per the same guide.
+    # Context measured at 1048576 tokens (oversized request -> "This model's maximum
+    # context length is 1048576 tokens"); keep the Flash-family 1000000 budget so the
+    # window_minus_output input budget retains margin under the hard ceiling.
+    # max_output_tokens above 393216 is rejected ("the valid range of max_tokens is
+    # [1, 393216]"), so the conservative 384000 output budget stands.
+    # Reasoning mechanics (inherited from the retired V4 Flash entries this id replaces,
+    # and still the shape this backend serves): effort must ride the Responses
+    # `reasoning` block, not the Chat-Completions `reasoning_effort` field the shared
+    # builder ignores, and this mode excludes flash from the flat reasoning_content
+    # replay path — reasoning comes back as Responses reasoning ITEMS instead. The
+    # endpoint returns no reasoning.encrypted_content (the include param is silently
+    # ignored) but exposes raw chain-of-thought as plain reasoning_text content, which
+    # the stream wrapper retains and resends next turn (mirroring Codex) so the context
+    # gauge counts it like any other content. The backend ALSO re-attaches prior CoT
+    # server-side keyed on its own function_call call_ids, billing it as input and
+    # deduping explicit copies (verified live 2026-08-04 on V4 Flash); reasoning-bearing
+    # replays with a foreign call_id 400 with "reasoning_text must be passed back",
+    # while minimal single-round histories with foreign ids are tolerated.
+    ("deepseek", "deepseek-flash"): {
+        "context_length": 1000000,
+        "max_completion_tokens": 384000,
+        "input_budget": "window_minus_output",
+        "default_temperature": 1.0,
+        "supports_vision": True,
+        "supports_hosted_web_search": False,
+        "preferred_edit_protocol": "claude_code",
+        "thinking_effort": ThinkingEffortSpec(
+            options=("none", "low", "high", "max"),
+            default="high",
+            mode="openai_responses_reasoning",
+        ),
+    },
+    ("deepseek", "deepseek-v4-pro"): {
+        "context_length": 1000000,
+        "max_completion_tokens": 65536,
+        "input_budget": "window_minus_output",
+        "default_temperature": 1.0,
+        "supports_vision": False,
+        # Hosted search still works on the V4 Pro model — re-verified live 2026-09-10
+        # (probe A body -> ws_calls=1, real answer), and it is the last first-party
+        # route that executes it. NOTE: from 12:00 Beijing Time 2026-09-14 DeepSeek
+        # routes deepseek-v4-pro requests to V4.1 Flash, whose architecture has no
+        # server-side web search; re-check this flag and the /web toggle after that.
+        "supports_hosted_web_search": True,
+        "preferred_edit_protocol": "claude_code",
+        "thinking_effort": ThinkingEffortSpec(
+            options=("none", "low", "high", "max"),
+            default="high",
+            mode="openai_responses_reasoning",
+        ),
+    },
+}

{kolega_code-0.37.0 → kolega_code-0.38.0}/pyproject.toml

@@ -8,7 +8,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "kolega-code"
-version = "0.37.0"
+version = "0.38.0"
 description = "Multi-agent coding in the terminal"
 readme = "README.md"
 license = "BUSL-1.1"

kolega_code-0.37.0/kolega_code/llm/specs/catalog/deepseek.py

@@ -1,103 +0,0 @@
-from kolega_code.llm.specs.types import ThinkingEffortSpec
-
-# max_completion_tokens is the MEASURED per-response output ceiling, not
-# DeepSeek's published 384000. Probed 2026-08-03 by
-# tests/agent/llm/test_deepseek_output_cap_live.py:
-#   - deepseek-v4-pro (chat), uncapped: the server cuts the stream at exactly
-#     65536 output tokens, reported honestly as finish_reason="length".
-#   - deepseek-v4-flash (Responses), uncapped: no hard cut observed — the model
-#     SELF-CENSORS ~7.4k tokens in ("response would exceed the maximum output
-#     length allowed for this chat interface") with a clean status="completed".
-#     With an explicit max_output_tokens=64000 it streams the full 64000 and
-#     reports the cap hit honestly (status="incomplete"/max_output_tokens).
-#   - Over-ceiling caps (384000) are accepted silently on both paths.
-# flash is 384000 (corrected from 65536, which was pro's measured chat ceiling
-# assumed to apply here): those probes only measured visible output, while a
-# reasoning-heavy call was later measured running to 112990 tokens in a single
-# Responses call.
-# pro never sends this value raw: the Responses path clamps to
-# DEEPSEEK_WIRE_OUTPUT_CAP=64000 (specs/accessors.py) so the client cap fires
-# before the server ceiling and truncation is reported honestly. flash passes
-# through.
-DEEPSEEK_SPECS = {
-    ("deepseek", "deepseek-v4-pro"): {
-        "context_length": 1000000,
-        "max_completion_tokens": 65536,
-        "input_budget": "window_minus_output",
-        "default_temperature": 1.0,
-        "supports_vision": False,
-        "supports_hosted_web_search": True,
-        "preferred_edit_protocol": "claude_code",
-        "thinking_effort": ThinkingEffortSpec(
-            options=("none", "low", "high", "max"),
-            default="high",
-            mode="openai_responses_reasoning",
-        ),
-    },
-    # deepseek-v4-flash speaks the Responses API (see DeepSeekResponsesProvider), so its
-    # reasoning effort must be emitted as the Responses `reasoning` block rather than the
-    # Chat-Completions `reasoning_effort` field the shared Responses request builder ignores.
-    # This mode also correctly excludes flash from the flat reasoning_content replay path:
-    # flash's reasoning rides in Responses reasoning ITEMS instead. The endpoint returns no
-    # reasoning.encrypted_content (the include param is silently ignored) but exposes the
-    # raw chain-of-thought as plain reasoning_text content, which the stream wrapper
-    # retains and resends next turn (mirroring Codex, the client this endpoint was adapted
-    # for) — so the context gauge counts it from the history like any other content. The
-    # backend ALSO restores prior CoT server-side keyed on its own function_call call_ids,
-    # billing it as input and deduping explicit copies (verified live 2026-08-04); deep
-    # reasoning-bearing replays with a foreign call_id 400 with "reasoning_text must be
-    # passed back", minimal single-round histories with foreign ids are tolerated.
-    ("deepseek", "deepseek-v4-flash"): {
-        "context_length": 1000000,
-        "max_completion_tokens": 384000,
-        "input_budget": "window_minus_output",
-        "default_temperature": 1.0,
-        "supports_vision": False,
-        # /responses executes {"type": "web_search"} server-side (search +
-        # open_page) and restores the searched content on replay of the
-        # web_search_call items, billed as input (verified live 2026-08-04,
-        # findings/probes/hosted_web_search_probe.py).
-        "supports_hosted_web_search": True,
-        "preferred_edit_protocol": "claude_code",
-        "thinking_effort": ThinkingEffortSpec(
-            options=("none", "low", "high", "max"),
-            default="high",
-            mode="openai_responses_reasoning",
-        ),
-    },
-    ("deepseek", "deepseek-v4-flash-vision-exp"): {
-        "context_length": 1000000,
-        "max_completion_tokens": 384000,
-        "input_budget": "window_minus_output",
-        "default_temperature": 1.0,
-        "supports_vision": True,
-        "supports_hosted_web_search": True,
-        "preferred_edit_protocol": "claude_code",
-        "thinking_effort": ThinkingEffortSpec(
-            options=("none", "low", "high", "max"),
-            default="high",
-            mode="openai_responses_reasoning",
-        ),
-    },
-    # Append previews: settings uses the first catalog model as provider fallback.
-    # Unlisted preview, verified on /responses on 2026-09-09. The ID advertises
-    # expiry on September 10; do not alias it to a stable model or make it default.
-    # Vision, hosted search, tools, and all four exposed efforts work live.
-    # Retain the Flash-family 1M context budget provisionally (not independently
-    # measured for this preview). The API accepts 384000 output tokens and reports
-    # a maximum of 393216; keep the existing conservative Flash output budget.
-    ("deepseek", "deepseek-v4.1-flash-expires-on-0910"): {
-        "context_length": 1000000,
-        "max_completion_tokens": 384000,
-        "input_budget": "window_minus_output",
-        "default_temperature": 1.0,
-        "supports_vision": True,
-        "supports_hosted_web_search": True,
-        "preferred_edit_protocol": "claude_code",
-        "thinking_effort": ThinkingEffortSpec(
-            options=("none", "low", "high", "max"),
-            default="high",
-            mode="openai_responses_reasoning",
-        ),
-    },
-}