roboshed 0.1.1.dev2__tar.gz → 0.1.1.dev3__tar.gz

{roboshed-0.1.1.dev2 → roboshed-0.1.1.dev3}/CHANGELOG.md

@@ -2,6 +2,44 @@
 
 ## Unreleased
 
+## 0.1.1.dev3 - 2026-09-13
+
+- Migrate capability bindings to the central typed contexts. Existing capability
+  arguments, owner configuration, endpoint overrides, defaults, and maintenance
+  order are preserved. Each build creates fresh tool state and reports actual
+  resources without initializing model clients. The RoboSprawl recipe now uses
+  `LLMEndpointRoute` for the selectable orchestrator while its Librarian endpoint
+  remains fixed.
+
+- Breaking: email factories now use the central `EmailContext`; attachment
+  resolvers bind a `Path` directly. `get_work_with_email` keeps its arguments and
+  requires a complete `EmailService`, which now inherits `ExternalDependency`
+  and implements `check()` using its read-only probe. Email dependencies are
+  inspectable without mailbox access and can be monitored without an agent.
+  Mailbox operations, permission checks, defaults, and error messages are preserved.
+  See the [context migration guide](../../docs/shed-tool-contexts.md#email-services-and-contexts).
+
+- Breaking: bind built-in file and maintenance factories to concrete typed contexts
+  instead of `Ctx`; direct patch stages accept a `Path` or `TruncationSpec`.
+  Existing `get_*` helper arguments remain supported. Command and summary contexts
+  report their actual resources through `tool.external_dependencies()`. Compaction
+  counters belong to the context: rebinding/copying shares them; constructing a new
+  context creates fresh state. See the [tool-context migration](../../docs/shed-tool-contexts.md).
+
+- Breaking: remove callback-based `inspect_dependencies`; inspect the configured
+  `DeployableAgent.external_dependencies()` instead; it constructs unstarted
+  agents using normal capability builders, without temporary sandbox isolation.
+  The health monitor accepts the resulting resources together with standalone
+  dependencies such as selectable models, deduplicating the combined sequence. Each resource owns its
+  synchronous `check() -> bool`; remove checker registrations and replace the
+  three category-specific probe helpers with `check_dependency(resource)` when
+  a sanitized observation is needed. Timeout, concurrency, metadata filtering,
+  and cached health-record behavior remain unchanged. See the README health guide.
+
+- Breaking: replace the setter-based `RoboSprawl` class with `robosprawl(sandbox, ...)`.
+  Supply deployment choices directly and unpack the returned root/background
+  agents. See [the recipe guide](../../docs/agent-factories.md).
+
 ## 0.1.1.dev2 - 2026-09-12
 
 - Breaking: move permission, sandbox, watched-agent, and default model inputs

{roboshed-0.1.1.dev2 → roboshed-0.1.1.dev3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: roboshed
-Version: 0.1.1.dev2
+Version: 0.1.1.dev3
 Import-Name: roboshed
 Summary: roboshed: optional building blocks for Roboz
 Project-URL: Repository, https://github.com/Tachion-Oy/roboz
@@ -11,7 +11,7 @@ Classifier: Development Status :: 2 - Pre-Alpha
 Classifier: Typing :: Typed
 Requires-Python: >=3.13
 Requires-Dist: pydantic>=2.12.4
-Requires-Dist: roboz<0.2.0,>=0.1.2.dev4
+Requires-Dist: roboz<0.2.0,>=0.1.2.dev5
 Description-Content-Type: text/markdown
 
 # roboshed
@@ -65,29 +65,30 @@ role definitions.
 Each capability declares the typed attributes it reads from its owning
 `DeployableAgent`. Runtime controls remain separate. Use `set_attributes()` to
 supply standalone permission policies and sandbox inputs before building.
+Capability builders construct the central typed tool contexts; each build gets
+fresh runtime state while retaining the selected endpoint objects. Build-based
+resource inspection uses those tools without initializing model clients.
 
 See the [factory and migration guide](../../docs/agent-factories.md).
 
 ## RoboSprawl deployment recipe
 
-`roboshed.deployments.robosprawl.RoboSprawl` is the concrete lazy persistent
-orchestrator and Librarian recipe. Construct `RoboSprawl()` without inputs;
-supply the required scoped sandbox, endpoint getter, and memory endpoint with
-`set_sandbox()`, `set_endpoint_getter()`, and `set_memory_endpoint()`.
-`set_additional_capabilities()`, `set_specialists()`, `set_interaction_mode()`,
-and `set_event_sinks()` supply optional inputs (empty sequences and
-`interaction_mode=None` by default). A `None` mode inherits core's current output
-setting, falling back to CLI when none is bound. Argument-free `build()` returns
-a fresh root/background-agent tuple
-or reports missing required inputs. `robosprawl` is an alias for this class.
+The recipe still requires migration to the concrete endpoint contract in this
+branch; importing it currently fails on the removed lazy-reference API. The
+standalone `orchestrator` and `librarian` constructors and their capabilities are
+migrated. The description below records the recipe behavior to preserve.
+
+`roboshed.deployments.robosprawl.robosprawl` is the concrete lazy persistent
+orchestrator and Librarian recipe. Call it with an already-scoped sandbox,
+`endpoint_getter`, `memory_endpoint`, `additional_capabilities`, `specialists`,
+`interaction_mode`, and optional `event_sinks`. It returns a fresh root and
+background-agent tuple.
 
 The recipe loads project memory and supplies project locations through initial
 messages. Its root follows the selected model getter; the Librarian uses its
-separate memory endpoint. Configuration and building start no agents or provider
-clients and create no project directories. Use a fresh recipe per new run;
-setters snapshot the sandbox and sequence containers, while supplied endpoints
-and capability/child objects remain caller-owned. The application owns scope
-selection and runtime lifecycle. See the [migration example](../../docs/agent-factories.md#fixed-robosprawl-recipe).
+separate memory endpoint. Construction starts no agents and creates no
+directories before the build requires its persistence sinks. The application
+owns scope selection and runtime lifecycle.
 
 ## Conversation compaction
 
@@ -108,8 +109,9 @@ Include this tool in an agent's `default_tools` and pass that agent's owning
 `system_prompt` and `skill_message` override the full continuation instructions.
 The tool preserves the contiguous bootstrap prefix and folds the remaining
 history, including previous summaries, into a new continuation message. Its
-status also carries the summary for event persistence. Each constructed tool
-owns its compaction count; constructing one per agent keeps counters independent.
+status also carries the summary for event persistence. Each capability build
+creates a fresh compaction context; tools copied or rebound to that context share
+its count, while separate builds keep counters independent.
 
 Successful status reports describe the compacted history's current usage and
 headroom. Summaries are budgeted below the configured threshold and endpoint
@@ -132,12 +134,12 @@ memory pipeline; core provides the mechanisms they use.
 
 ## Context API migration
 
-Low-level tool factories now use `roboz.Ctx(**values)` directly, with service,
-endpoint, and executable objects supplied without wrappers. The specialized
-context classes have been removed. Existing tool builders retain their keyword
-arguments, defaults, permission checks, cancellation, and timeout behavior.
-See the [migration guide](https://github.com/Tachion-Oy/roboz/blob/main/docs/context-migration.md)
-for low-level context fields and state ownership.
+Low-level tool factories use concrete context classes from `roboshed.tools`.
+Their typed constructors own required fields, defaults, validation, and fresh
+state. Direct resources such as endpoints may also be factory contexts. Existing
+tool builders retain their keyword arguments, permission checks, cancellation,
+and timeout behavior. See the
+[context guide](../../docs/shed-tool-contexts.md) for the complete mapping.
 
 
 ## Deployable agent graphs
@@ -177,9 +179,10 @@ capabilities, and sinks remain caller-owned. No execution state is retained on
 the definition. The old deployment wrapper, factories, host protocols, and
 result bundles are removed.
 Use core's `roboz.llm.ModelSelector` for lazy model selection.
-`roboz.dependencies` supplies exact dependency registration and binding;
-`roboshed.dependency_health` supplies isolated `inspect_dependencies`, probes,
-and monitoring without a web framework or provider SDK.
+`roboz.dependencies` supplies the resource contract and ordered deduplication.
+`DeployableAgent.external_dependencies()` builds an unstarted graph and inspects
+its resources. `roboshed.dependency_health` monitors resource-owned checks without a
+web framework or provider SDK.
 Permission policies treat configured folder names literally. Health scheduling
 retries observation failures; timed-out workers retain their concurrency slots
 until completion.
@@ -190,3 +193,88 @@ HUD file-link/markdown contract. The external RoboSprawl application may select
 it through `Capability(auto_loaded_skills=(robosprawl,))`. Concrete paths,
 endpoints, extra capabilities, and specialist definitions remain application
 choices. The `robosprawl` recipe assembles and builds a fresh agent graph.
+
+
+## Dependency health
+
+Call `definition.external_dependencies()` on the configured `DeployableAgent`.
+It constructs fresh agents without event sinks and delegates to their existing
+tool dependency inspection, including foreground and background descendants and
+unloaded skills. It returns a deduplicated `tuple[ExternalDependency, ...]` and
+does not invoke agents or request endpoint initialization or availability checks.
+Capability builders run normally, including any construction effects they own.
+
+The former `inspect_dependencies` callback helper and its temporary sandbox are
+removed. The optional definition method requires configuration sufficient for
+normal construction; the monitor never calls it automatically. Agents do not need
+to be running. Existing runtime agents can still be inspected directly. If inputs
+are substituted for discovery, they must produce the resource declarations used
+by the actual deployment.
+
+The monitor accepts resources independently of agents. Combine agent resources
+with other resources explicitly, for example:
+
+```python
+monitor = DependencyHealthMonitor((
+    *definition.external_dependencies(),
+    *selectable_models,
+    transcription_endpoint,
+))
+```
+
+Here `selectable_models` is a sequence of concrete `LLMEndpoint` objects; they need
+not be attached to an agent or selected yet. The monitor keeps the first resource
+for each dependency ID across the combined sequence. Resources are captured when
+the monitor is constructed; create a new monitor if the resource set changes.
+
+Pass those resources directly to `DependencyHealthMonitor(resources)`. Its
+constructor creates pending records without performing checks. Explicit
+`run_once()` or scheduled observation calls each resource's synchronous
+`check() -> bool` in a worker thread. Checks own their service-specific behavior
+and any required client initialization; the monitor supplies bounded concurrency,
+timeouts, scheduling, and cached observations.
+
+```python
+import asyncio
+import sys
+
+from roboz.dependencies import ExecutableDependency
+from roboshed.dependency_health import DependencyHealthMonitor, DependencyStatus
+
+program = ExecutableDependency(sys.executable)
+monitor = DependencyHealthMonitor((program,))
+assert monitor.records()[0].status is DependencyStatus.PENDING
+asyncio.run(monitor.run_once())
+assert monitor.records()[0].status is DependencyStatus.AVAILABLE
+```
+
+`check_dependency(resource)` performs one synchronous check and returns a
+`DependencyCheckResult`: `True` means available, `False` becomes `model_unavailable`
+for models or `not_found` for other resources, and exceptions become sanitized
+reason codes. Other return values produce `protocol_error`. Provider exception
+payloads are not exposed through records. Record schemas and metadata filtering
+are unchanged. Async checker callbacks and registration records are removed;
+implement the synchronous method on the resource instead.
+
+Replace `check_executable`, `check_openai_compatible_endpoint`, and
+`check_network_service` with `check_dependency` when a sanitized health result
+is needed, or use `resource.check()` for the primitive boolean/exception contract.
+The old helper names have no compatibility aliases. Capability bindings and the
+RoboSprawl deployment recipe now use the concrete context and endpoint contracts.
+
+## Concrete tool contexts
+
+File-command, guard, editing, maintenance, and email factories now use concrete context
+classes exported from `roboshed.tools`. Existing `get_run_file_command`,
+`get_apply_patch`, and `get_compactify_messages_when_needed_tool` keyword arguments
+are retained. Direct factory users should follow the
+[context migration guide](../../docs/shed-tool-contexts.md), including the context
+ownership rules for compaction counters. Command and summary resources are
+reported through `tool.external_dependencies()` without running external work.
+
+
+Email contexts live in the same module. `get_work_with_email` keeps its existing
+arguments; direct email execution uses `EmailContext`, and attachment resolvers
+accept `Path`. `EmailService` defines every provider operation and the resource
+identity/metadata contract. Its availability check calls the existing read-only
+probe. See the [email context contract](../../docs/shed-tool-contexts.md#email-services-and-contexts).

{roboshed-0.1.1.dev2 → roboshed-0.1.1.dev3}/README.md

@@ -49,29 +49,30 @@ role definitions.
 Each capability declares the typed attributes it reads from its owning
 `DeployableAgent`. Runtime controls remain separate. Use `set_attributes()` to
 supply standalone permission policies and sandbox inputs before building.
+Capability builders construct the central typed tool contexts; each build gets
+fresh runtime state while retaining the selected endpoint objects. Build-based
+resource inspection uses those tools without initializing model clients.
 
 See the [factory and migration guide](../../docs/agent-factories.md).
 
 ## RoboSprawl deployment recipe
 
-`roboshed.deployments.robosprawl.RoboSprawl` is the concrete lazy persistent
-orchestrator and Librarian recipe. Construct `RoboSprawl()` without inputs;
-supply the required scoped sandbox, endpoint getter, and memory endpoint with
-`set_sandbox()`, `set_endpoint_getter()`, and `set_memory_endpoint()`.
-`set_additional_capabilities()`, `set_specialists()`, `set_interaction_mode()`,
-and `set_event_sinks()` supply optional inputs (empty sequences and
-`interaction_mode=None` by default). A `None` mode inherits core's current output
-setting, falling back to CLI when none is bound. Argument-free `build()` returns
-a fresh root/background-agent tuple
-or reports missing required inputs. `robosprawl` is an alias for this class.
+The recipe still requires migration to the concrete endpoint contract in this
+branch; importing it currently fails on the removed lazy-reference API. The
+standalone `orchestrator` and `librarian` constructors and their capabilities are
+migrated. The description below records the recipe behavior to preserve.
+
+`roboshed.deployments.robosprawl.robosprawl` is the concrete lazy persistent
+orchestrator and Librarian recipe. Call it with an already-scoped sandbox,
+`endpoint_getter`, `memory_endpoint`, `additional_capabilities`, `specialists`,
+`interaction_mode`, and optional `event_sinks`. It returns a fresh root and
+background-agent tuple.
 
 The recipe loads project memory and supplies project locations through initial
 messages. Its root follows the selected model getter; the Librarian uses its
-separate memory endpoint. Configuration and building start no agents or provider
-clients and create no project directories. Use a fresh recipe per new run;
-setters snapshot the sandbox and sequence containers, while supplied endpoints
-and capability/child objects remain caller-owned. The application owns scope
-selection and runtime lifecycle. See the [migration example](../../docs/agent-factories.md#fixed-robosprawl-recipe).
+separate memory endpoint. Construction starts no agents and creates no
+directories before the build requires its persistence sinks. The application
+owns scope selection and runtime lifecycle.
 
 ## Conversation compaction
 
@@ -92,8 +93,9 @@ Include this tool in an agent's `default_tools` and pass that agent's owning
 `system_prompt` and `skill_message` override the full continuation instructions.
 The tool preserves the contiguous bootstrap prefix and folds the remaining
 history, including previous summaries, into a new continuation message. Its
-status also carries the summary for event persistence. Each constructed tool
-owns its compaction count; constructing one per agent keeps counters independent.
+status also carries the summary for event persistence. Each capability build
+creates a fresh compaction context; tools copied or rebound to that context share
+its count, while separate builds keep counters independent.
 
 Successful status reports describe the compacted history's current usage and
 headroom. Summaries are budgeted below the configured threshold and endpoint
@@ -116,12 +118,12 @@ memory pipeline; core provides the mechanisms they use.
 
 ## Context API migration
 
-Low-level tool factories now use `roboz.Ctx(**values)` directly, with service,
-endpoint, and executable objects supplied without wrappers. The specialized
-context classes have been removed. Existing tool builders retain their keyword
-arguments, defaults, permission checks, cancellation, and timeout behavior.
-See the [migration guide](https://github.com/Tachion-Oy/roboz/blob/main/docs/context-migration.md)
-for low-level context fields and state ownership.
+Low-level tool factories use concrete context classes from `roboshed.tools`.
+Their typed constructors own required fields, defaults, validation, and fresh
+state. Direct resources such as endpoints may also be factory contexts. Existing
+tool builders retain their keyword arguments, permission checks, cancellation,
+and timeout behavior. See the
+[context guide](../../docs/shed-tool-contexts.md) for the complete mapping.
 
 
 ## Deployable agent graphs
@@ -161,9 +163,10 @@ capabilities, and sinks remain caller-owned. No execution state is retained on
 the definition. The old deployment wrapper, factories, host protocols, and
 result bundles are removed.
 Use core's `roboz.llm.ModelSelector` for lazy model selection.
-`roboz.dependencies` supplies exact dependency registration and binding;
-`roboshed.dependency_health` supplies isolated `inspect_dependencies`, probes,
-and monitoring without a web framework or provider SDK.
+`roboz.dependencies` supplies the resource contract and ordered deduplication.
+`DeployableAgent.external_dependencies()` builds an unstarted graph and inspects
+its resources. `roboshed.dependency_health` monitors resource-owned checks without a
+web framework or provider SDK.
 Permission policies treat configured folder names literally. Health scheduling
 retries observation failures; timed-out workers retain their concurrency slots
 until completion.
@@ -174,3 +177,88 @@ HUD file-link/markdown contract. The external RoboSprawl application may select
 it through `Capability(auto_loaded_skills=(robosprawl,))`. Concrete paths,
 endpoints, extra capabilities, and specialist definitions remain application
 choices. The `robosprawl` recipe assembles and builds a fresh agent graph.
+
+
+## Dependency health
+
+Call `definition.external_dependencies()` on the configured `DeployableAgent`.
+It constructs fresh agents without event sinks and delegates to their existing
+tool dependency inspection, including foreground and background descendants and
+unloaded skills. It returns a deduplicated `tuple[ExternalDependency, ...]` and
+does not invoke agents or request endpoint initialization or availability checks.
+Capability builders run normally, including any construction effects they own.
+
+The former `inspect_dependencies` callback helper and its temporary sandbox are
+removed. The optional definition method requires configuration sufficient for
+normal construction; the monitor never calls it automatically. Agents do not need
+to be running. Existing runtime agents can still be inspected directly. If inputs
+are substituted for discovery, they must produce the resource declarations used
+by the actual deployment.
+
+The monitor accepts resources independently of agents. Combine agent resources
+with other resources explicitly, for example:
+
+```python
+monitor = DependencyHealthMonitor((
+    *definition.external_dependencies(),
+    *selectable_models,
+    transcription_endpoint,
+))
+```
+
+Here `selectable_models` is a sequence of concrete `LLMEndpoint` objects; they need
+not be attached to an agent or selected yet. The monitor keeps the first resource
+for each dependency ID across the combined sequence. Resources are captured when
+the monitor is constructed; create a new monitor if the resource set changes.
+
+Pass those resources directly to `DependencyHealthMonitor(resources)`. Its
+constructor creates pending records without performing checks. Explicit
+`run_once()` or scheduled observation calls each resource's synchronous
+`check() -> bool` in a worker thread. Checks own their service-specific behavior
+and any required client initialization; the monitor supplies bounded concurrency,
+timeouts, scheduling, and cached observations.
+
+```python
+import asyncio
+import sys
+
+from roboz.dependencies import ExecutableDependency
+from roboshed.dependency_health import DependencyHealthMonitor, DependencyStatus
+
+program = ExecutableDependency(sys.executable)
+monitor = DependencyHealthMonitor((program,))
+assert monitor.records()[0].status is DependencyStatus.PENDING
+asyncio.run(monitor.run_once())
+assert monitor.records()[0].status is DependencyStatus.AVAILABLE
+```
+
+`check_dependency(resource)` performs one synchronous check and returns a
+`DependencyCheckResult`: `True` means available, `False` becomes `model_unavailable`
+for models or `not_found` for other resources, and exceptions become sanitized
+reason codes. Other return values produce `protocol_error`. Provider exception
+payloads are not exposed through records. Record schemas and metadata filtering
+are unchanged. Async checker callbacks and registration records are removed;
+implement the synchronous method on the resource instead.
+
+Replace `check_executable`, `check_openai_compatible_endpoint`, and
+`check_network_service` with `check_dependency` when a sanitized health result
+is needed, or use `resource.check()` for the primitive boolean/exception contract.
+The old helper names have no compatibility aliases. Capability bindings and the
+RoboSprawl deployment recipe now use the concrete context and endpoint contracts.
+
+## Concrete tool contexts
+
+File-command, guard, editing, maintenance, and email factories now use concrete context
+classes exported from `roboshed.tools`. Existing `get_run_file_command`,
+`get_apply_patch`, and `get_compactify_messages_when_needed_tool` keyword arguments
+are retained. Direct factory users should follow the
+[context migration guide](../../docs/shed-tool-contexts.md), including the context
+ownership rules for compaction counters. Command and summary resources are
+reported through `tool.external_dependencies()` without running external work.
+
+
+Email contexts live in the same module. `get_work_with_email` keeps its existing
+arguments; direct email execution uses `EmailContext`, and attachment resolvers
+accept `Path`. `EmailService` defines every provider operation and the resource
+identity/metadata contract. Its availability check calls the existing read-only
+probe. See the [email context contract](../../docs/shed-tool-contexts.md#email-services-and-contexts).

{roboshed-0.1.1.dev2 → roboshed-0.1.1.dev3}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "roboshed"
-version = "0.1.1.dev2"
+version = "0.1.1.dev3"
 description = "roboshed: optional building blocks for Roboz"
 readme = "README.md"
 requires-python = ">=3.13"
@@ -9,7 +9,7 @@ license = "Apache-2.0"
 license-files = ["LICENSE"]
 import-names = ["roboshed"]
 classifiers = ["Development Status :: 2 - Pre-Alpha", "Typing :: Typed"]
-dependencies = ["roboz>=0.1.2.dev4,<0.2.0","pydantic>=2.12.4"]
+dependencies = ["roboz>=0.1.2.dev5,<0.2.0","pydantic>=2.12.4"]
 
 [build-system]
 requires = ["hatchling>=1.32"]

{roboshed-0.1.1.dev2 → roboshed-0.1.1.dev3}/src/roboshed/capabilities.py

@@ -25,11 +25,16 @@ from roboshed.tools.compactification import (
     DEFAULT_THRESHOLD_PERCENT,
 )
 from roboshed.tools.consolidate_memory import consolidate_memory
+from roboshed.tools.contexts import (
+    ConsolidateMemoryContext,
+    PurgeFilesContext,
+    SleepBetweenRunsContext,
+    SnapshotConversationsContext,
+)
 from roboshed.tools.purge_files import purge_files
 from roboshed.tools.sleep_between_runs import sleep_between_runs
 from roboshed.tools.snapshot_conversations import snapshot_conversations
 from roboshed.sandbox import PermissionPolicy, Sandbox
-from roboz.dependencies import ExternalDependencyReference
 from roboz.deployment import (
     AgentCapability,
     Capability,
@@ -37,12 +42,11 @@ from roboz.deployment import (
     RequiredAttributeType,
     RequiredAttributes,
 )
-from roboz.llm import EndpointLike, LLMEndpoint, MockLLMEndpoint
+from roboz.llm import EndpointLike, LLMEndpoint, LLMEndpointRoute, MockLLMEndpoint
 from roboz.runtime import EventPipe
-from roboz.tooling.context import Ctx
 
 
-_ENDPOINT_TYPES = (LLMEndpoint, MockLLMEndpoint, ExternalDependencyReference)
+_ENDPOINT_TYPES = (LLMEndpoint, MockLLMEndpoint, LLMEndpointRoute)
 _AGENT_ENDPOINT_REQUIRED: RequiredAttributes = {
     "agent_endpoint": _ENDPOINT_TYPES,
 }
@@ -174,7 +178,7 @@ class ConversationSnapshots(AgentCapability):
         return Capability(
             default_tools=(
                 snapshot_conversations(
-                    Ctx(
+                    SnapshotConversationsContext(
                         endpoint=endpoint,
                         conversation_root=sandbox.project_logs_dir(),
                         snapshot_root=sandbox.project_snapshots_dir(),
@@ -241,7 +245,7 @@ class MemoryConsolidation(AgentCapability):
         return Capability(
             default_tools=(
                 consolidate_memory(
-                    Ctx(
+                    ConsolidateMemoryContext(
                         endpoint=endpoint,
                         snapshot_root=sandbox.project_snapshots_dir(),
                         memory_root=sandbox.project_memory_dir(),
@@ -278,14 +282,14 @@ class ArtifactRetention(AgentCapability):
         return Capability(
             default_tools=(
                 purge_files(
-                    Ctx(
+                    PurgeFilesContext(
                         folders=[sandbox.project_logs_dir()],
                         pattern="*.json",
                         max_files=self.max_log_files,
                     )
                 ).copy(name=PURGE_LOGS_TOOL_NAME),
                 purge_files(
-                    Ctx(
+                    PurgeFilesContext(
                         folders=[sandbox.project_snapshots_dir()],
                         pattern="*.md",
                         max_files=self.max_snapshot_files,
@@ -293,7 +297,7 @@ class ArtifactRetention(AgentCapability):
                     )
                 ).copy(name=PURGE_SNAPSHOTS_TOOL_NAME),
                 purge_files(
-                    Ctx(
+                    PurgeFilesContext(
                         folders=[sandbox.project_memory_dir()],
                         pattern="*.md",
                         max_files=self.max_memory_files,
@@ -328,7 +332,7 @@ class MaintenanceCadence(AgentCapability):
         return Capability(
             default_tools=(
                 sleep_between_runs(
-                    Ctx(
+                    SleepBetweenRunsContext(
                         seconds=self.seconds,
                         is_cancelled=lambda: pipe.cancelled,
                         conversation_root=sandbox.project_logs_dir(),