Add claude-fable-5 to litellm proxy and bump Claude Code to 2.1.207

.github/actions/tmr-setup/action.yml

@@ -23,7 +23,7 @@ runs:
 
     - name: Install Claude Code CLI
       shell: bash
-      run: npm install -g @anthropic-ai/claude-code@2.1.160
+      run: npm install -g @anthropic-ai/claude-code@2.1.207
 
     - name: Configure git identity
       shell: bash

.github/workflows/release-tests.yml

@@ -66,7 +66,7 @@ jobs:
         run: uv sync --all-packages
 
       - name: Install Claude Code CLI
-        run: npm install -g @anthropic-ai/claude-code@2.1.160
+        run: npm install -g @anthropic-ai/claude-code@2.1.207
 
       - name: Run Docker release tests
         run: PYTEST_MAX_DURATION_SECONDS=900 uv run pytest -n 0 --timeout 600 --no-cov --cov-fail-under=0 -v --tb=short --ignore apps/minds -m '(docker or docker_sdk) and release'
@@ -117,7 +117,7 @@ jobs:
         run: brew install tmux unison
 
       - name: Install Claude Code CLI
-        run: npm install -g @anthropic-ai/claude-code@2.1.160
+        run: npm install -g @anthropic-ai/claude-code@2.1.207
 
       - name: Configure Docker Hub mirror
         if: runner.os == 'Linux'

apps/modal_litellm/README.md

@@ -91,6 +91,10 @@ even on litellm versions whose bundled price map predates a model. The model
 list lives in `apps/modal_litellm/app.py` (`LITELLM_CONFIG`) and is mirrored in
 `litellm_proxy/config.yaml`; `config_drift_test.py` fails if the two diverge.
 
+Fable ($10 / $50 per 1M input / output):
+
+- `claude-fable-5`
+
 Opus (current price tier, $5 / $25 per 1M input / output):
 
 - `claude-opus-4-8` (latest Opus)

apps/modal_litellm/app.py

@@ -68,6 +68,12 @@
 # predates a model (e.g. claude-opus-4-8 only landed in litellm's price map
 # in the 1.88.0 pre-release line). MUST stay in sync with
 # litellm_proxy/config.yaml -- config_drift_test.py enforces this.
+_FABLE_PRICING = {
+    "input_cost_per_token": 0.00001,
+    "output_cost_per_token": 0.00005,
+    "cache_creation_input_token_cost": 0.0000125,
+    "cache_read_input_token_cost": 0.000001,
+}
 _OPUS_PRICING = {
     "input_cost_per_token": 0.000005,
     "output_cost_per_token": 0.000025,
@@ -108,6 +114,8 @@ def _model_entry(model_name: str, pricing: dict[str, float]) -> dict[str, object
 
 LITELLM_CONFIG = {
     "model_list": [
+        # Fable line.
+        _model_entry("claude-fable-5", _FABLE_PRICING),
         # Current Opus line.
         _model_entry("claude-opus-4-8", _OPUS_PRICING),
         _model_entry("claude-opus-4-7", _OPUS_PRICING),

apps/modal_litellm/changelog/mngr-fable.md

@@ -0,0 +1 @@
+Add `claude-fable-5` to the LiteLLM proxy model list with inline pricing ($10 / $50 per 1M input / output tokens, cache write 1.25e-5, cache read 1e-6, mirrored from litellm's price map), so workspaces using imbue_cloud virtual keys can run Claude Fable 5 with correct cost tracking.

dev/changelog/mngr-fable.md

@@ -0,0 +1,3 @@
+- Bump the pinned Claude Code CLI version from 2.1.160 to 2.1.207 in CI workflows (`release-tests.yml`, `tmr-setup` action) and the minds e2e snapshot script, matching the new workspace pin that supports Claude Fable 5.
+
+- Add `claude-fable-5` with inline pricing ($10 / $50 per 1M input / output tokens, cache write 1.25e-5, cache read 1e-6) to the repo-root local-dev LiteLLM proxy config (`litellm_proxy/config.yaml`), kept in sync with `apps/modal_litellm/app.py` by a drift test.

libs/mngr/changelog/mngr-fable.md

@@ -0,0 +1 @@
+Bump the release-sandbox Dockerfile's pinned Claude Code version from 2.1.160 to 2.1.207, which supports the Claude Fable 5 model. Must land together with the matching `[agent_types.claude].version` bump in default-workspace-template.

libs/mngr/imbue/mngr/resources/Dockerfile

@@ -69,7 +69,7 @@ ENV PATH="/root/.local/bin:$PATH"
 # .mngr/settings.toml pin (`[agent_types.claude].version`). The release-test
 # `test_claude_code_version_matches_default_workspace_template_pin` enforces the
 # sync; bump both together when rolling a new claude release.
-ARG CLAUDE_CODE_VERSION="2.1.160"
+ARG CLAUDE_CODE_VERSION="2.1.207"
 RUN curl -fsSL https://claude.ai/install.sh > /tmp/install_claude.sh && ( if [ -n "$CLAUDE_CODE_VERSION" ]; then cat /tmp/install_claude.sh | bash -s "$CLAUDE_CODE_VERSION"; else cat /tmp/install_claude.sh | bash; fi && test -x /root/.local/bin/claude ) || ( cat /tmp/install_claude.sh && exit 1 )
 ENV CLAUDE_CODE_VERSION=${CLAUDE_CODE_VERSION}
 

libs/mngr_claude/changelog/mngr-fable.md

@@ -0,0 +1 @@
+Add `auto_disable_questions` to `ClaudeAgentConfig`: when True, appends `--disallowed-tools AskUserQuestion` to the agent invocation so unattended agents can't block on questions (recent Claude Code releases no longer treat AskUserQuestion as a permission prompt, so `auto_allow_permissions` alone no longer covers it). Appended as a separate flag; Claude accumulates repeated `--disallowed-tools` flags, so a list already supplied via `cli_args` is preserved.

libs/mngr_claude/imbue/mngr_claude/plugin.py

@@ -311,6 +311,11 @@ class ClaudeAgentConfig(AgentTypeConfig):
         description="When True, adds a PermissionRequest hook that auto-allows all permission dialogs. "
         "This means Claude Code will never pause for permission approval.",
     )
+    auto_disable_questions: bool = Field(
+        default=False,
+        description="When True, adds `--disallowed-tools AskUserQuestion` to the agent invocation to "
+        "prevent it from ever asking questions (which can cause the agent to get blocked)",
+    )
     settings_overrides: Annotated[dict[str, Any], SettingsPatchField()] = Field(
         default_factory=dict,
         description="A patch merged onto your home Claude settings at provisioning. A bare key assigns "
@@ -2360,6 +2365,9 @@ def assemble_command(
         # (Claude is last-wins) -- the accepted, documented limitation of that mode.
         cli_args = self.agent_config.cli_args
         all_extra_args = cli_args + quote_agent_args(agent_args)
+        # Claude appends & unions repeated --disallowed-tools flags.
+        if self.agent_config.auto_disable_questions:
+            all_extra_args = all_extra_args + ("--disallowed-tools", "AskUserQuestion")
         args_str = " ".join(all_extra_args) if all_extra_args else ""
 
         # Read the latest session ID from the tracking file written by the SessionStart hook.

libs/mngr_claude/imbue/mngr_claude/plugin_test.py

@@ -493,6 +493,42 @@ def test_claude_agent_assemble_command_with_cli_args_and_agent_args(
     assert parsed.create_args == ["--verbose", "--model", "opus"]
 
 
+def test_claude_agent_assemble_command_auto_disable_questions_appends_flag(
+    local_provider: LocalProviderInstance, tmp_path: Path, temp_mngr_ctx: MngrContext
+) -> None:
+    """auto_disable_questions=True should append --disallowed-tools AskUserQuestion after all other args."""
+    agent, host = make_claude_agent(
+        local_provider,
+        tmp_path,
+        temp_mngr_ctx,
+        agent_config=ClaudeAgentConfig(
+            cli_args=("--disallowed-tools", "TodoWrite"),
+            auto_disable_questions=True,
+            check_installation=False,
+        ),
+    )
+
+    command = agent.assemble_command(host=host, agent_args=(), command_override=None)
+
+    parsed = _ParsedAssembleCommand(str(command))
+    # Appended as a second --disallowed-tools flag (claude appends & unions
+    # repeated flags), so the cli_args-supplied list is preserved.
+    expected = ["--disallowed-tools", "TodoWrite", "--disallowed-tools", "AskUserQuestion"]
+    assert parsed.resume_args == expected
+    assert parsed.create_args == expected
+
+
+def test_claude_agent_assemble_command_auto_disable_questions_off_by_default(
+    local_provider: LocalProviderInstance, tmp_path: Path, temp_mngr_ctx: MngrContext
+) -> None:
+    """Default config should not add any --disallowed-tools flag."""
+    agent, host = make_claude_agent(local_provider, tmp_path, temp_mngr_ctx)
+
+    command = agent.assemble_command(host=host, agent_args=(), command_override=None)
+
+    assert "--disallowed-tools" not in shlex.split(str(command))
+
+
 def test_claude_agent_assemble_command_passes_user_settings_through(
     local_provider: LocalProviderInstance, tmp_path: Path, temp_mngr_ctx: MngrContext
 ) -> None:

libs/mngr_robinhood/changelog/mngr-fable.md

@@ -0,0 +1 @@
+Add a Claude Fable family entry ($10 / $50 per 1M input / output tokens, cache write $12.50, cache read $1.00) to the Agent SDK per-model price table, so `ResultMessage.total_cost_usd` is populated for `claude-fable-5` sessions instead of being left unset.

libs/mngr_robinhood/imbue/mngr_robinhood/_agent_sdk/pricing.py

@@ -52,6 +52,12 @@ class ModelPricing(FrozenModel):
     cache_write_usd_per_million=Decimal("6.25"),
     cache_read_usd_per_million=Decimal("0.50"),
 )
+_FABLE_PRICING: Final[ModelPricing] = ModelPricing(
+    input_usd_per_million=Decimal("10.00"),
+    output_usd_per_million=Decimal("50.00"),
+    cache_write_usd_per_million=Decimal("12.50"),
+    cache_read_usd_per_million=Decimal("1.00"),
+)
 
 # Model-id substrings mapped to their pricing, checked in order. claude model ids look like
 # ``claude-haiku-4-5-20251001`` / ``claude-sonnet-4-6`` / ``claude-opus-4-8``, so a family
@@ -60,6 +66,7 @@ class ModelPricing(FrozenModel):
     ("haiku", _HAIKU_PRICING),
     ("sonnet", _SONNET_PRICING),
     ("opus", _OPUS_PRICING),
+    ("fable", _FABLE_PRICING),
 )
 
 

libs/mngr_robinhood/imbue/mngr_robinhood/_agent_sdk/pricing_test.py

@@ -7,6 +7,7 @@ def test_resolve_model_pricing_matches_family_for_dated_and_alias_ids() -> None:
     assert resolve_model_pricing("claude-haiku-4-5-20251001") is resolve_model_pricing("haiku")
     assert resolve_model_pricing("claude-sonnet-4-6") is resolve_model_pricing("sonnet")
     assert resolve_model_pricing("claude-opus-4-8") is resolve_model_pricing("opus")
+    assert resolve_model_pricing("claude-fable-5") is resolve_model_pricing("fable")
 
 
 def test_resolve_model_pricing_is_none_for_unknown_model() -> None:

libs/mngr_usage/changelog/mngr-fable.md

@@ -0,0 +1 @@
+Add `anthropic/claude-fable-5` to the usage pricing table ($10 / $50 per 1M input / output tokens), so token-only usage sources report cost for Claude Fable 5 instead of showing it as unpriced.

libs/mngr_usage/imbue/mngr_usage/pricing.py

@@ -48,6 +48,12 @@ class PerTokenPrices(FrozenModel):
 # Anthropic per-token pricing, mirrored verbatim from apps/modal_litellm/app.py
 # (which itself mirrors litellm's map). Grouped by tier so the "same price"
 # relationship across model ids stays explicit, exactly as modal_litellm does.
+_FABLE_PRICES: Final[PerTokenPrices] = PerTokenPrices(
+    input_cost_per_token=0.00001,
+    output_cost_per_token=0.00005,
+    cache_creation_input_token_cost=0.0000125,
+    cache_read_input_token_cost=0.000001,
+)
 _OPUS_PRICES: Final[PerTokenPrices] = PerTokenPrices(
     input_cost_per_token=0.000005,
     output_cost_per_token=0.000025,
@@ -122,6 +128,7 @@ class PerTokenPrices(FrozenModel):
 # apps/modal_litellm (drift test); OpenAI stays in sync with litellm directly
 # (litellm_pricing_test).
 MODEL_PRICING: Final[dict[str, PerTokenPrices]] = {
+    "anthropic/claude-fable-5": _FABLE_PRICES,
     "anthropic/claude-opus-4-8": _OPUS_PRICES,
     "anthropic/claude-opus-4-7": _OPUS_PRICES,
     "anthropic/claude-opus-4-6": _OPUS_PRICES,

litellm_proxy/config.yaml

@@ -5,6 +5,16 @@
 # model_prices_and_context_window map) so cost tracking is correct even on
 # litellm versions whose bundled price map predates a model.
 model_list:
+  # Fable line.
+  - model_name: claude-fable-5
+    litellm_params:
+      model: anthropic/claude-fable-5
+      api_key: os.environ/ANTHROPIC_API_KEY
+      input_cost_per_token: 0.00001
+      output_cost_per_token: 0.00005
+      cache_creation_input_token_cost: 0.0000125
+      cache_read_input_token_cost: 0.000001
+
   # Current Opus line.
   - model_name: claude-opus-4-8
     litellm_params:

scripts/snapshot_minds_e2e_state.py

@@ -96,7 +96,7 @@
 # Keep these in sync with apps/minds/.nvmrc and apps/minds/package.json engines.
 _NODE_VERSION: Final[str] = "24.15.0"
 _PNPM_VERSION: Final[str] = "10.33.4"
-_CLAUDE_CODE_VERSION: Final[str] = "2.1.160"
+_CLAUDE_CODE_VERSION: Final[str] = "2.1.207"
 
 # In-sandbox entrypoint that invokes the shared e2e workspace runner the
 # pytest test also uses, but without the test's mngr-destroy cleanup. The