* Restore task snapshot via a worker-level dependency
`get_access_token()` returned `None` inside background tasks whenever
`FASTMCP_DOCKET_URL` pointed at a `redis+cluster://` URL. The write side
was fine — it went through `docket.redis()`, which is cluster-aware —
but fastmcp kept a parallel sync Redis client just to read the snapshot
back, and `Redis.from_url()` rejects the cluster scheme.
Docket 0.19.1 ships worker-level dependencies that resolve per task in
the same asyncio.Task as user code, so ContextVars propagate cleanly.
That lets us load the snapshot once via `restore_task_snapshot` and
drop the sync Redis path entirely. Sync helpers like
`get_access_token()` and `get_http_request()` now just read a
ContextVar; Docket is the sole Redis consumer.
Closes#3897
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Revert TaskKey stub to a plain return
NotImplementedError would fire at module import if anything evaluated
the default; a no-op stub keeps the module usable without the
fastmcp[tasks] extra, which is what we want.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* Scope tasks to authorization context, not session
Tasks were keyed by the transport-layer Mcp-Session-Id, which is
server-assigned and changes on reconnect — so clients lost access to
their running tasks after any connection interruption.
The MCP spec says tasks should be bound to authorization context, not
session. This replaces session_id with task_scope (derived from
AccessToken.client_id, URL-encoded) in all task data Redis keys and
Docket task keys. When no auth is configured, a "_" sentinel is used
and security comes from UUID task ID entropy per the spec.
Session ID is still used for transport-level concerns (notification
queues, subscriber registration) and is now stored in the
TaskContextSnapshot payload so background workers can still deliver
notifications.
Also extracts all the task context infrastructure (TaskContextInfo,
TaskContextSnapshot, snapshot loading, session/server registries) from
server/dependencies.py into a new server/tasks/context.py to keep the
DI module from sprawling further.
Closes#3758🤖 Generated with Claude Code
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Rename _redis_key to _snapshot_redis_key
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Document in-process session registry as an optimization
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Tidy imports and docstrings
Hoist imports where safe, keep subscriptions/notifications deferred in
handlers.py since they pull in docket at module level. Sharpen docstrings
on keys.py and context.py so each module owns its lane. Clean up the
re-export block in dependencies.py.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Fix misleading comment on re-export block
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Tighten task scope: include sub claim, partition keyspaces
Addresses review feedback on #3800:
- Compose task scope from client_id and the JWT sub claim (when present)
so fixed-OAuth deployments isolate per user, not just per client.
- Replace the "_" anonymous sentinel with a tagged keyspace partition.
Docket keys are now auth:{enc_scope}:... or anon:..., and Redis keys
use fastmcp:task:auth:{enc_scope}:... or fastmcp:task:anon:...,
routed through a single task_redis_prefix() helper.
- get_task_scope() returns the raw scope (or None); encoding happens
once at the keys.py boundary, collapsing the previous double-quote
invariant.
- Drop the dormant fallback in notifications.py that routed
input_required relays into the anon keyspace when task_scope was
missing -- log and skip instead.
- Add comprehensive parser/encoder tests in test_task_keys.py covering
round-trips, malformed keys, and adversarial scopes ("anon", "_", and
scopes containing : / | %).
- Add cross-scope rejection tests: distinct client_ids, distinct sub
claims under a shared client_id, and authenticated vs anonymous.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Retry when LLM returns text instead of calling final_response tool
Instead of raising RuntimeError immediately when the LLM returns a text
response instead of calling the `final_response` tool for structured
output, retry up to 3 times with an explicit nudge message asking the
model to use the tool. This mirrors the existing retry behavior for
validation errors but with a separate, smaller cap.
Fixes#3847
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Add tests for text response retry logic
Tests cover:
- Text response followed by successful final_response (retry works)
- Text response exceeding max retries (raises RuntimeError)
- Nudge message appended to history on retry
- No retry when result_type is None (text is valid)
Addresses review feedback from PR review tool (v1 flagged missing tests as high severity).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Slim down text response retry tests
Remove test_nudge_message_in_history (implementation detail).
Reduce boilerplate in remaining 3 tests.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Fix ruff format
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Detect async/sync generators after tool execution and materialize
into lists before the result conversion pipeline processes them
- Generator materialization runs inside timeout scope so slow generators
respect the configured timeout
- Handle bytes return types: UTF-8 bytes as text, non-UTF-8 as base64
- Suppress output_schema for bytes return types (can't be structured JSON)
- Catch UnicodeDecodeError alongside PydanticSerializationError in
convert_result for robustness
Fixes#3829
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Unify background task context forwarding and fix concurrent dependency bugs
We've been getting a steady trickle of edge-case reports around background tasks
and contextual dependencies over the last few months (#3654, #3656, #3569). Each
one pointed at a different symptom, but they all traced back to the same area:
the way context is negotiated between the "frontend" server and Docket workers
was grown piecemeal, with each new piece of context (access tokens, HTTP headers,
origin request IDs) getting its own Redis key, its own restore function, and its
own ContextVar. This made it hard to reason about what state was available where,
and the shared-instance Dependency pattern made concurrent tasks stomp on each
other's cleanup state.
This takes a step back and reworks the whole thing as a single unified system:
- Dependency subclasses (_CurrentContext, Progress, _CurrentAccessToken, etc.)
are now stateless factories — __aenter__ returns a fresh per-invocation
object, so concurrent tasks never share mutable state. Fixes#3654, #3656.
- The three individual context-snapshot Redis keys (access_token, http_headers,
origin_request_id) are collapsed into a single TaskContextSnapshot stored as
one JSON key per task. The three _restore_task_* functions and two ContextVars
they populated are gone.
- Sync functions like get_http_request() and get_access_token() now find the
snapshot transparently in background tasks via a 3-tier sync fallback:
ContextVar (set by _CurrentContext for functions with deps) → in-memory dict
(same-process workers) → sync Redis GET (out-of-process workers). No function
wrapping needed.
- The _wrap_for_task_http_headers hack is deleted. FunctionTool registers its
raw function with Docket so Docket sees and resolves ALL dependencies,
including Docket-native ones like Retry and Timeout.
- ProxyTool.from_mcp_tool() now propagates execution.taskSupport metadata from
remote tools. Fixes#3569.
- Removed redundant _current_docket/_current_worker ContextVar management from
Context.__aenter__/__aexit__ (they're only set in the lifespan now).
Closes#3654Closes#3656Closes#3569🤖 Generated with Claude Code
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Address code review feedback
- _OptionalCurrentContext: guard __aexit__ against cleaning up contexts it
didn't create (check is_background_task before delegating)
- Narrow except clauses in snapshot loading (OSError, JSONDecodeError, etc.
instead of bare Exception)
- Fix docstrings on register_with_docket for resources/prompts/templates
- Simplify Progress: read ExecutionProgress directly from current_execution
instead of creating and manually entering a DocketProgress wrapper
🤖 Generated with Claude Code
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Use pop-on-access transfer buffer instead of bounded LRU cache for snapshots
The in-memory snapshot dict is a transfer mechanism, not a cache. Entries go
in at submission and come out at the worker's first access. Using pop instead
of get means the dict only holds entries during the brief submission-to-execution
window, bounded by task concurrency (~10) rather than a 10,000-entry LRU limit.
🤖 Generated with Claude Code
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Drop in-memory transfer buffer, use sync Redis for all backends
Instead of maintaining an in-memory dict to bridge the async/sync gap, use
a sync Redis client directly. For memory:// backends (fakeredis), shares the
same FakeServer instance via docket._redis.get_memory_server() so data written
by the async Docket client is visible to sync reads. For real Redis, creates a
standard sync connection. No in-process state to manage at all.
🤖 Generated with Claude Code
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Move snapshot operations to TaskContextSnapshot methods
capture(), from_json(), to_json(), save() are now classmethod/instance methods
on the dataclass instead of free functions. Deduplicates JSON parsing that was
copy-pasted between the async and sync load paths.
🤖 Generated with Claude Code
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Trim implementation details from register_with_docket docstrings
🤖 Generated with Claude Code
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Clarify docket lookup comment in submit_to_docket
🤖 Generated with Claude Code
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Restore docket/worker ContextVar bridge in Context.__aenter__
Servers that own the Docket (the parent) re-set _current_docket/_current_worker
from their instance attributes when entering a Context. Mounted children skip
this (their _docket is None), so they inherit the parent's value. This is needed
for ASGI deployments where ContextVars set during the lifespan don't propagate
to request handlers.
🤖 Generated with Claude Code
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Key snapshot cache by task_id to prevent cross-task context leakage
Docket workers may reuse the same asyncio context for sequential tasks.
The ContextVar cache now stores (task_id, snapshot) tuples so stale entries
from previous tasks are automatically ignored.
🤖 Generated with Claude Code
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* pin pydantic-monty to 0.0.8
* rename tool/prompt/resource base modules to avoid decorator name shadow
* add sys.modules shims for old submodule import paths
* preserve original module paths in deprecation warnings
* clarify when sys.modules shims can be removed
* perf: expose minimum_check_interval, reduce task pickup latency
The Docket Worker polls for new tasks every minimum_check_interval
(previously hardcoded to 250ms in pydocket). Expose this setting so
users can tune it, default to 50ms, and override to 10ms in tests.
This cuts average task pickup latency from ~125ms to ~5ms per task.
* perf: reduce task test overhead and eliminate cross-test contamination
- Expose minimum_check_interval setting (default 50ms, 10ms in tests)
to reduce Docket Worker task pickup latency
- Isolate fakeredis per test via unique memory:// URLs to prevent
stale _async_blocking tasks from contaminating subsequent tests
- Make client disconnect timeout configurable (default 5s, 1s in tests)
- Add --durations=50 to CI for passive performance regression detection
- Remove 15s timeout band-aids from task test conftest files
- Add explicit @pytest.mark.timeout(10) to cancellation tests
- Fix deprecated FastMCP.as_proxy() usage in test_task_proxy.py
* fix: shield lifespan teardown from cancellation
Co-authored-by: Claude <noreply@anthropic.com>
* fix: stabilize flaky task and timeout tests under parallel execution
Co-authored-by: Claude <noreply@anthropic.com>
---------
Co-authored-by: Claude <noreply@anthropic.com>
* Replace vendored DI with uncalled-for
FastMCP vendored a minimal DI engine extracted from Docket (~164 lines)
with try/except fallback patterns everywhere. The `uncalled-for` package
is a clean, typed extraction of this same system, and since Docket will
also depend on it (chrisguidry/docket#353), `uncalled_for.Dependency`
becomes the single canonical base class.
This deletes the `_vendor/docket_di/` directory, replaces all the
try/except import patterns with direct `uncalled_for` imports, and
updates the `Dependency.execution` → `current_execution` ContextVar
references to match the Docket branch. The `Progress` class now
delegates to an internal impl and returns `self` from `__aenter__`
(matching Docket's pattern) so that ty's generic resolution works
without `type: ignore` suppressions.
Temporarily points pydocket at the `use-uncalled-for` branch so both
sides can be validated together in CI.
🤖 Generated with Claude Code
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Re-export Dependency from fastmcp.dependencies
Internal code like azure.py should import from the fastmcp namespace
rather than reaching into uncalled_for directly.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Import Dependency from fastmcp namespace in tests
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Add generic type parameters to Dependency subclasses
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Mention uncalled-for in DI docs
The DI engine now comes from uncalled-for, so the docs should credit
it alongside Docket. Also updates the Docket docs link to docket.lol.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Point docket dependency at main
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Bump uncalled-for pin to >=0.2.0
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Fix uncalled-for imports for 0.2.0 API changes
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Support Shared() dependencies without docket
Enters a SharedContext at server lifetime so that Shared() dependencies
from uncalled-for resolve once and are cached across tool/resource/prompt
calls. When running with docket, the Worker already handles this; this
covers the non-docket path and direct call_tool() usage.
Also re-exports Shared from fastmcp.dependencies.
Closes#3251
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Bump docket lockfile to latest main
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Remove duplicate test classes from rebase conflict resolution
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Point docket dependency at pydocket>=0.18.0 release
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Pair SharedContext __aenter__ with __aexit__ in Context lifecycle
The old `_ensure_shared_context` on the server called `__aenter__()` on a
lazy `SharedContext` but never `__aexit__()`, leaking the exit stack and
its resources. Moved the SharedContext management into Context's own
enter/exit so it's properly paired: when docket is available the lifespan
handles it, otherwise Context creates and cleans up a per-request one.
Updated Shared() tests to use Client (which runs the lifespan) rather
than calling server methods directly, since cross-request sharing
requires a lifespan.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Hoist SharedContext import to module level
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* Fix background Context request correlation
* Make OptionalCurrentContext type-safe
Refactor OptionalCurrentContext to wrap CurrentContext instead of overriding __aenter__ with a wider return type. Adds a background-task origin_request_id round-trip test and applies ruff formatting.
* Update repository references from jlowin/fastmcp to prefecthq/fastmcp
* Retrigger CI after repo transfer
* chore: Update SDK documentation
* Only run deep triage on bug issues for jlowin
---------
Co-authored-by: marvin-context-protocol[bot] <225465937+marvin-context-protocol[bot]@users.noreply.github.com>
Moves relay_elicitation() into elicitation.py so it can reuse
handle_task_input() for the Redis push instead of duplicating that logic.
notifications.py just detects the trigger and calls it.
Also fixes the related-task metadata key from modelcontextprotocol.io/ to
io.modelcontextprotocol/ to match the current spec:
https://modelcontextprotocol.io/specification/2025-11-25/basic/utilities/tasks
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When a background task calls ctx.elicit(), the notification subscriber now
detects the input_required notification and sends a standard elicitation/create
request to the client via session.elicit(). The client's elicitation_handler
fires, and the relay pushes the response to Redis for the blocked worker.
This means clients can respond to background task elicitation using the same
elicitation_handler they'd use for any other elicitation — no need to interact
with Redis or call handle_task_input() directly.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace 1300+ lines of mock-heavy unit tests with 391 lines of integration
tests using real Client(mcp) connections and memory:// Docket backend.
- test_context_background_task.py: 17 tests covering report_progress delta
tracking, elicitation flow, edge cases, and fail-fast on push failure
- test_notifications.py: 2 E2E tests for notification queue lifecycle
Tasks belong in capabilities.tasks (first-class field) per SEP-1686,
not capabilities.experimental.tasks. This fixes VS Code Copilot 1.107+
integration which checks capabilities.tasks?.requests?.tools?.call.
Changes:
- Update get_task_capabilities() to return ServerTasksCapability types
- Override get_capabilities() in LowLevelServer to set tasks field
- Remove experimental_capabilities parameter usage
- Update test to verify correct location
Fixes#2870
Adds a test that verifies task cancellation actually interrupts running
coroutines (they receive CancelledError) rather than just marking the task
as cancelled in Redis while the coroutine continues to completion.
This requires pydocket >= 0.16.2 which added best-effort cancellation via
Redis pub/sub signaling to workers.
Closes#2679🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Refactor FastMCPProxy into ProxyProvider
Move proxy functionality from custom manager classes to the Provider pattern:
- Create ProxyProvider that implements the Provider interface
- Move all proxy code to src/fastmcp/server/providers/proxy.py
- Keep FastMCPProxy as a convenience wrapper using ProxyProvider
- Add deprecation warning when importing from old location
- Convert handler classmethods to module-level functions
- Remove redundant get_* methods (base class defaults work)
* Remove unused Components class, simplify TaskComponents()
* Add poll_interval to TaskConfig
Allow users to configure polling interval per component via
TaskConfig(poll_interval=timedelta(...)). Default is 5 seconds.
* Update snapshots for poll_interval field
* Add version badge to poll_interval docs
* Add defensive handling for Redis data and align default poll intervals
* Replace type: ignore[attr-defined] with isinstance assertions in tests
* Fix isinstance assertions in failing tests
- Fix enum test to check for ResponseEnum instead of str
- Fix binary resource test to check for BlobResourceContents instead of TextResourceContents
- Fix Root type tests to check attributes directly instead of isinstance checks
* Fix type errors without using type: ignore
- Remove execution methods from TransformingProvider (only handles transformations)
- Add execution methods to base Provider class with default implementations
- Fix type narrowing in tests using cast() instead of type: ignore
- Fix PromptResult type handling in prompt render tests
- Fix type narrowing in middleware test for arguments and structured_content
* Add supports_tasks() method to replace string mode checks
Consolidates task config mode checks into a readable method on TaskConfig.
Instead of `task_config.mode == "forbidden"` or `task_config.mode != "forbidden"`,
code now uses `task_config.supports_tasks()` for clearer intent.
Updated 20 instances across the codebase and added type assertions in tests
to resolve type checker warnings.
* Update test to match new error message
* Add test_custom_subclass_tasks.py
* Refactor provider execution: delegate to middleware via wrapper components
- Remove execution methods (call_tool, read_resource, etc.) from Provider base
- Add FastMCPProvider* wrapper classes that delegate to child server middleware
- Move task routing to Tool._run() using contextvars (_task_metadata, _tool_call_key)
- Add convert_to_tool_result(result, output_schema) utility for Docket results
- Add convert_to_prompt_result() utility for prompt task results
- Pass namespaced key via add_to_docket(name=) for mounted tool lookup
* Standardize add_to_docket() with fn_key/task_key parameters
All components now use explicit fn_key (function lookup) and task_key
(result storage) parameters instead of relying on implicit key handling.
This fixes mounted component task execution where the MCP-visible key
differs from the Docket-registered function name.
* Add middleware chain tests for three-level mount hierarchy
Tests verify middleware runs at parent, child, and grandchild levels
for tools, resources, prompts, and resource templates.
* WIP: Provider refactor - unified submit_to_docket, template _read() in progress
Work in progress on refactoring execution to use component _read()/_run()/_render() methods.
Template background tasks not yet working - needs fix for Docket key lookup.
* Fix conversion functions to take full component for attribute access
Pass Tool/Prompt/Resource/Template to conversion functions instead of
individual attributes, ensuring access to serializer, output_schema,
mime_type, etc. Also fixes mixed-content output schema validation.
* Refactor: unified convert_result() methods and check_background_task helper
- Add convert_result() instance methods to all component types (Tool, Prompt, Resource, ResourceTemplate)
- Extract duplicated task routing logic into check_background_task() helper
- Fix type annotations on FastMCPProviderResource.read() and FastMCPProviderPrompt.render()
- Update protocol.py to use component.convert_result() uniformly
* Update tests to use namespace= instead of deprecated prefix= parameter
* Use CreateTaskResult for background task creation
Move result conversion logic to components (convert_result methods) and
return proper CreateTaskResult SDK type from task handlers. Consolidates
MCP protocol handler overrides into server.py with documentation.
* Address CodeRabbit nitpicks
- Add type annotation for resource parameter in handle_resource_as_task
- Move RootModel import to module level in client.py
* Document intentionally unused task_meta parameters
Prefix with underscore to suppress lint warnings. Client TTL will be
configurable via TaskConfig in the future; keeping parameter for API stability.
* Simplify .key as computed property
Keep .key as the standard lookup interface for all components but
implement it as a computed property instead of a stored field.
- Remove _key private attribute and custom model_copy(key=...)
- .key returns .name for tools/prompts, str(.uri) for resources,
.uri_template for templates
- Use .key universally for component lookups in managers
- MountedProvider: prefix URIs only for resources/templates, not names
- Docket registration: tools/prompts use .key, resources use .name
(matches fn.__name__ for function lookup)
* Simplify .key as computed property
Keep .key as the standard lookup interface for all components but
implement it as a computed property instead of a stored field.
- Remove _key private attribute and custom model_copy(key=...)
- .key returns .name for tools/prompts, str(.uri) for resources,
.uri_template for templates
- Use .key universally for component lookups and Docket registration
- MountedProvider: prefix URIs only for resources/templates, not names
- Add _backend_* fields to proxy classes to preserve original identifiers
for backend calls when prefixed via import_server
* Standardize .key as computed property
- .key is now a read-only computed property:
- Tools/Prompts: returns .name
- Resources: returns str(.uri)
- Templates: returns .uri_template
- Prefixing uses model_copy(update={...}) to change underlying field
- Resource/template names are NOT prefixed, only URIs
- Move import_server tests to tests/deprecated/
When a tool was registered with a custom name different from its function
name, task execution would fail because Docket registered the function by
its `__name__` but the handler looked it up by the tool's configured name.
This switches to using Docket's new `names=` parameter (pydocket 0.16.0)
to register functions with their proper lookup keys, and removes the
`_create_named_fn_wrapper` hack that was used for mounted servers.
Closes#2642🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
