Addresses recommendations from the UX developer's agent experience report.
Adds provenance predicates, DAG structure changes, error resilience, and
a published OWL ontology.
Explainability additions:
- Tool candidates: tg:toolCandidate on Analysis events lists the tools
visible to the LLM for each iteration (names only, descriptions in config)
- Termination reason: tg:terminationReason on Conclusion/Synthesis events
(final-answer, plan-complete, subagents-complete)
- Step counter: tg:stepNumber on iteration events
- Pattern decision: new tg:PatternDecision entity in the DAG between
session and first iteration, carrying tg:pattern and tg:taskType
- Latency: tg:llmDurationMs on Analysis events, tg:toolDurationMs on
Observation events
- Token counts on events: tg:inToken/tg:outToken/tg:llmModel on
Grounding, Focus, Synthesis, and Analysis events
- Tool/parse errors: tg:toolError on Observation events with tg:Error
mixin type. Parse failures return as error observations instead of
crashing the agent, giving it a chance to retry.
Envelope unification:
- Rename chunk_type to message_type across AgentResponse schema,
translator, SDK types, socket clients, CLI, and all tests.
Agent and RAG services now both use message_type on the wire.
Ontology:
- specs/ontology/trustgraph.ttl — OWL vocabulary covering all 26 classes,
7 object properties, and 36+ datatype properties including new predicates.
DAG structure tests:
- tests/unit/test_provenance/test_dag_structure.py verifies the
wasDerivedFrom chain for GraphRAG, DocumentRAG, and all three agent
patterns (react, plan, supervisor) including the pattern-decision link.
Expose LLM token usage (in_token, out_token, model) across all
service layers
Propagate token counts from LLM services through the prompt,
text-completion, graph-RAG, document-RAG, and agent orchestrator
pipelines to the API gateway and Python SDK. All fields are Optional
— None means "not available", distinguishing from a real zero count.
Key changes:
- Schema: Add in_token/out_token/model to TextCompletionResponse,
PromptResponse, GraphRagResponse, DocumentRagResponse,
AgentResponse
- TextCompletionClient: New TextCompletionResult return type. Split
into text_completion() (non-streaming) and
text_completion_stream() (streaming with per-chunk handler
callback)
- PromptClient: New PromptResult with response_type
(text/json/jsonl), typed fields (text/object/objects), and token
usage. All callers updated.
- RAG services: Accumulate token usage across all prompt calls
(extract-concepts, edge-scoring, edge-reasoning,
synthesis). Non-streaming path sends single combined response
instead of chunk + end_of_session.
- Agent orchestrator: UsageTracker accumulates tokens across
meta-router, pattern prompt calls, and react reasoning. Attached
to end_of_dialog.
- Translators: Encode token fields when not None (is not None, not truthy)
- Python SDK: RAG and text-completion methods return
TextCompletionResult (non-streaming) or RAGChunk/AgentAnswer with
token fields (streaming)
- CLI: --show-usage flag on tg-invoke-llm, tg-invoke-prompt,
tg-invoke-graph-rag, tg-invoke-document-rag, tg-invoke-agent
Pipeline fixes:
- Fix agent getting empty response from graph-rag by combining answer +
explain data in single message (RequestResponse returns first msg)
- Fix Doc RAG pipeline: add content field to Qdrant doc payload, seed 10
document chunks, fix type mismatches across base/flow/client
- Forward explainability events from agent's KnowledgeQuery to client
- Add "agent" to TERM_BEARING_RESPONSE_SERVICES for triple translation
- Fix embeddings env var (OLLAMA_URL), user/collection threading, edge
scoring threshold, and various protocol mismatches
Branding:
- Rename TrustGraph → Beep Graph (title, sidebar, settings, about)
- Custom lambda + ThugLife pixel glasses SVG logo component
- Forest green color palette (brand-50 through brand-900)
- SVG favicon + PNG icons (16/32/180/192/512)
- PWA manifest with service worker for offline shell caching
- Splash screen with animated logo pulse on app load
- Ambient glow background with drifting green radial blobs
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add chat UX improvements: message actions toolbar (copy/delete/regenerate)
on hover, inline explainability subgraph visualization from RAG/agent
queries, and token metadata for all chat modes. Enhance graph page with
SPO query filters, configurable triple limit, and type legend overlay.
Extract shared graph utilities for reuse across components.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add dedicated /mcp-tools page for managing MCP servers and tools from the
workbench. Includes CRUD dialogs, config API integration, and feature flag
gating via mcpTools switch. QA pass also fixes accessibility across existing
pages: aria-expanded on chat phase blocks, tabpanel tabindex on prompts,
toggle contrast ratio (WCAG 2.1 SC 1.4.11) on settings.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add global focus-visible outline for buttons, switches, selects, and
inputs so all interactive elements show a visible brand-500 ring on
keyboard focus (not just NavLinks and dialog close)
- Darken light-mode --color-border from #e4e4e7 to #d4d4d8 so input
borders, dividers, and mode selector outlines are visible on white
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Fix Metadata/EntityEmbeddings schema migration tail and add regression tests (#776)
The Metadata dataclass dropped its `metadata: list[Triple]` field
and EntityEmbeddings/ChunkEmbeddings settled on a singular
`vector: list[float]` field, but several call sites kept passing
`Metadata(metadata=...)` and `EntityEmbeddings(vectors=...)`. The
bugs were latent until a websocket client first hit
`/api/v1/flow/default/import/entity-contexts`, at which point the
dispatcher TypeError'd on construction.
Production fixes (5 call sites on the same migration tail):
* trustgraph-flow gateway dispatchers entity_contexts_import.py
and graph_embeddings_import.py — drop the stale
Metadata(metadata=...) kwarg; switch graph_embeddings_import
to the singular `vector` wire key.
* trustgraph-base messaging translators knowledge.py and
document_loading.py — fix decode side to read the singular
`"vector"` key, matching what their own encode sides have
always written.
* trustgraph-flow tables/knowledge.py — fix Cassandra row
deserialiser to construct EntityEmbeddings(vector=...)
instead of vectors=.
* trustgraph-flow gateway core_import/core_export — switch the
kg-core msgpack wire format to the singular `"v"`/`"vector"`
key and drop the dead `m["m"]` envelope field that referenced
the removed Metadata.metadata triples list (it was a
guaranteed KeyError on the export side).
Defense-in-depth regression coverage (32 new tests across 7 files):
* tests/contract/test_schema_field_contracts.py — pin the field
set of Metadata, EntityEmbeddings, ChunkEmbeddings,
EntityContext so any future schema rename fails CI loudly
with a clear diff.
* tests/unit/test_translators/test_knowledge_translator_roundtrip.py
and test_document_embeddings_translator_roundtrip.py -
encode→decode round-trip the affected translators end to end,
locking in the singular `"vector"` wire key.
* tests/unit/test_gateway/test_entity_contexts_import_dispatcher.py
and test_graph_embeddings_import_dispatcher.py — exercise the
websocket dispatchers' receive() path with realistic
payloads, the direct regression test for the original
production crash.
* tests/unit/test_gateway/test_core_import_export_roundtrip.py
— pack/unpack the kg-core msgpack format through the real
dispatcher classes (with KnowledgeRequestor mocked),
including a full export→import round-trip.
* tests/unit/test_tables/test_knowledge_table_store.py —
exercise the Cassandra row → schema conversion via __new__ to
bypass the live cluster connection.
Also fixes an unrelated leaked-coroutine RuntimeWarning in
test_gateway/test_service.py::test_run_method_calls_web_run_app: the
mocked aiohttp.web.run_app now closes the coroutine that Api.run() hands
it, mirroring what the real run_app would do, instead of leaving it for
the GC to complain about.
* Fix RabbitMQ request/response race and chunker Flow API drift
Two unrelated regressions surfaced after the v2.2 queue class
refactor. Bundled here because both are small and both block
production.
1. Request/response race against ephemeral RabbitMQ response
queues
Commit feeb92b3 switched response/notify queues to per-subscriber
auto-delete exclusive queues. That fixed orphaned-queue
accumulation but introduced a setup race: Subscriber.start()
created the run() task and returned immediately, while the
underlying RabbitMQ consumer only declared and bound its queue
lazily on the first receive() call. RequestResponse.request()
therefore published the request before any queue was bound to the
matching routing key, and the broker dropped the reply. Symptoms:
"Failed to fetch config on notify" / "Request timeout exception"
repeating roughly every 10s in api-gateway, document-embeddings
and any other service exercising the config notify path.
Fix:
* Add ensure_connected() to the BackendConsumer protocol;
implement it on RabbitMQBackendConsumer (calls _connect
synchronously, declaring and binding the queue) and as a
no-op on PulsarBackendConsumer (Pulsar's client.subscribe is
already synchronous at construction).
* Convert Subscriber's readiness signal from a non-existent
Event to an asyncio.Future created in start(). run() calls
consumer.ensure_connected() immediately after
create_consumer() and sets _ready.set_result(None) on first
successful bind. start() awaits the future via asyncio.wait
so it returns only once the consumer is fully bound. Any
reply published after start() returns is therefore guaranteed
to land in a bound queue.
* First-attempt connection failures call
_ready.set_exception(e) and exit run() so start() unblocks
with the error rather than hanging forever — the existing
higher-level retry pattern in fetch_and_apply_config takes
over from there. Runtime failures after a successful start
still go through the existing retry-with-backoff path.
* Update the two existing graceful-shutdown tests that
monkey-patch Subscriber.run with a custom coroutine to honor
the new contract by signalling _ready themselves.
* Add tests/unit/test_base/test_subscriber_readiness.py with
five regression tests pinning the readiness contract:
ensure_connected must be called before start() returns;
start() must block while ensure_connected runs
(race-condition guard with a threading.Event gate);
first-attempt create_consumer and ensure_connected failures
must propagate to start() instead of hanging;
ensure_connected must run before any receive() call.
2. Chunker Flow parameter lookup using the wrong attribute
trustgraph-base/trustgraph/base/chunking_service.py was reading
flow.parameters.get("chunk-size") and chunk-overlap, but the Flow
class has no `parameters` attribute — parameter lookup is exposed
through Flow.__call__ (flow("chunk-size") returns the resolved
value or None). The exception was caught and logged as a
WARNING, so chunking continued with the default sizes and any
configured chunk-size / chunk-overlap was silently ignored:
chunker - WARNING - Could not parse chunk-size parameter:
'Flow' object has no attribute 'parameters'
The chunker tests didn't catch this because they constructed
mock_flow = MagicMock() and configured
mock_flow.parameters.get.side_effect = ..., which is the same
phantom attribute MagicMock auto-creates on demand. Tests and
production agreed on the wrong API.
Fix: switch chunking_service.py to flow("chunk-size") /
flow("chunk-overlap"). Update both chunker test files to mock the
__call__ side_effect instead of the phantom parameters.get,
merging parameter values into the existing flow() lookup the
on_message tests already used for producer resolution.
The Metadata dataclass dropped its `metadata: list[Triple]` field
and EntityEmbeddings/ChunkEmbeddings settled on a singular
`vector: list[float]` field, but several call sites kept passing
`Metadata(metadata=...)` and `EntityEmbeddings(vectors=...)`. The
bugs were latent until a websocket client first hit
`/api/v1/flow/default/import/entity-contexts`, at which point the
dispatcher TypeError'd on construction.
Production fixes (5 call sites on the same migration tail):
* trustgraph-flow gateway dispatchers entity_contexts_import.py
and graph_embeddings_import.py — drop the stale
Metadata(metadata=...) kwarg; switch graph_embeddings_import
to the singular `vector` wire key.
* trustgraph-base messaging translators knowledge.py and
document_loading.py — fix decode side to read the singular
`"vector"` key, matching what their own encode sides have
always written.
* trustgraph-flow tables/knowledge.py — fix Cassandra row
deserialiser to construct EntityEmbeddings(vector=...)
instead of vectors=.
* trustgraph-flow gateway core_import/core_export — switch the
kg-core msgpack wire format to the singular `"v"`/`"vector"`
key and drop the dead `m["m"]` envelope field that referenced
the removed Metadata.metadata triples list (it was a
guaranteed KeyError on the export side).
Defense-in-depth regression coverage (32 new tests across 7 files):
* tests/contract/test_schema_field_contracts.py — pin the field
set of Metadata, EntityEmbeddings, ChunkEmbeddings,
EntityContext so any future schema rename fails CI loudly
with a clear diff.
* tests/unit/test_translators/test_knowledge_translator_roundtrip.py
and test_document_embeddings_translator_roundtrip.py -
encode→decode round-trip the affected translators end to end,
locking in the singular `"vector"` wire key.
* tests/unit/test_gateway/test_entity_contexts_import_dispatcher.py
and test_graph_embeddings_import_dispatcher.py — exercise the
websocket dispatchers' receive() path with realistic
payloads, the direct regression test for the original
production crash.
* tests/unit/test_gateway/test_core_import_export_roundtrip.py
— pack/unpack the kg-core msgpack format through the real
dispatcher classes (with KnowledgeRequestor mocked),
including a full export→import round-trip.
* tests/unit/test_tables/test_knowledge_table_store.py —
exercise the Cassandra row → schema conversion via __new__ to
bypass the live cluster connection.
Also fixes an unrelated leaked-coroutine RuntimeWarning in
test_gateway/test_service.py::test_run_method_calls_web_run_app: the
mocked aiohttp.web.run_app now closes the coroutine that Api.run() hands
it, mirroring what the real run_app would do, instead of leaving it for
the GC to complain about.
Add the full MCP tool pipeline enabling agents to invoke external tools
(like Brave Search) via MCP servers:
- Add ToolRequest/ToolResponse types and mcp-tool topics to @trustgraph/base
- Create McpToolService (FlowProcessor) that connects to external MCP servers
via @modelcontextprotocol/sdk StreamableHTTP transport
- Add createMcpTool() to wire MCP tools into the agent's ReAct loop
- Implement config-driven tool registration in AgentService with backward-
compatible fallback to hardcoded tools
- Add tool filtering by group and state (port of Python tool_filter.py)
- Register mcp-tool in gateway dispatcher and export from @trustgraph/flow
- Fix flow restart race condition: skip restart when flow definitions unchanged
- Update seed config with MCP server config and tool definitions
- Add run scripts for MCP tool service and Brave Search MCP server
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Fix FalkorDB triples query: client v5 returns objects not arrays, use named field access
- Fix embeddings service: align spec names to "embeddings-request"/"embeddings-response"
- Fix client triplesQuery: read `triples` field instead of `response` from backend
- Fix graph page crash: guard against non-array triples, accept literals as entity nodes
- Add seed:demo script for AI industry knowledge graph (254 triples, 64 entities)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Derive consumer behaviour from queue class, remove
consumer_type parameter
The queue class prefix (flow, request, response, notify) now
fully determines consumer behaviour in both RabbitMQ and Pulsar
backends. Added 'notify' class for ephemeral broadcast (config
push notifications). Response and notify classes always create
per-subscriber auto-delete queues, eliminating orphaned queues
that accumulated on service restarts.
Change init-trustgraph to set up the 'notify' namespace in
Pulsar instead of old hangover 'state'.
Fixes 'stuck backlog' on RabbitMQ config notification queue.
The triples client returns Uri/Literal (str subclasses), not Term
objects. _quoted_triple() treated all values as IRIs, so literal
objects like skos:definition values were mistyped in focus
provenance events, and trace_source_documents could not match
them in the store.
Added to_term() to convert Uri/Literal back to Term, threaded a
term_map from follow_edges_batch through
get_subgraph/get_labelgraph into uri_map, and updated
_quoted_triple to accept Term objects directly.
Store the initialization Promise in the requestors map synchronously
before yielding, so concurrent callers for the same key await the same
instance — prevents orphaned RequestResponse objects and duplicate NATS
subscriptions. Mirrors upstream fix 8f18ba02.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Subscriber resilience: recreate consumer after connection failure
- Move consumer creation from Subscriber.start() into the run() loop,
matching the pattern used by Consumer. If the connection drops and the
consumer is closed in the finally block, the loop now recreates it on
the next iteration instead of spinning forever on a None consumer.
Consumer thread safety:
- Dedicated ThreadPoolExecutor per consumer so all pika operations
(create, receive, acknowledge, negative_acknowledge) run on the
same thread — pika BlockingConnection is not thread-safe
- Applies to both Consumer and Subscriber classes
Config handler type audit — fix four mismatched type registrations:
- librarian: was ["librarian"] (non-existent type), now ["flow",
"active-flow"] (matches config["flow"] that the handler reads)
- cores/service: was ["kg-core"], now ["flow"] (reads
config["flow"])
- metering/counter: was ["token-costs"], now ["token-cost"]
(singular)
- agent/mcp_tool: was ["mcp-tool"], now ["mcp"] (reads
config["mcp"])
Update tests
Three QA iterations to convergence (zero issues remaining):
Workbench UI:
- Connection badge: amber "Connected (no auth)" for unauthenticated state
- Theme persistence: restore script in index.html + localStorage sync
- Settings About section: add bottom padding so content isn't clipped
- Clear messages: cancel in-flight requests when clearing chat
- Feature switch labels: proper casing + acronym handling (MCP, LLM)
- Token Cost badge: hidden during loading state
- ARIA: role="switch", aria-checked on toggles, aria-labels on buttons
- ConfigApi: null-safe chaining for getPrompts/getSystemPrompt
Grafana dashboards:
- Auto-refresh 30s on all 3 dashboards
- Panel heights reduced to fit viewport without scrolling
- Anonymous role upgraded to Editor for Explore access
Infrastructure:
- Nginx: DNS resolver with variable-based upstream (prevents crash loop)
- Workbench port set to 3002 in .env
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Provenance triples are now included directly in explain messages from
GraphRAG, DocumentRAG, and Agent services, eliminating the need for
follow-up knowledge graph queries to retrieve explainability details.
Each explain message in the response stream now carries:
- explain_id: root URI for this provenance step (unchanged)
- explain_graph: named graph where triples are stored (unchanged)
- explain_triples: the actual provenance triples for this step (new)
Changes across the stack:
- Schema: added explain_triples field to GraphRagResponse,
DocumentRagResponse, and AgentResponse
- Services: all explain message call sites pass triples through
(graph_rag, document_rag, agent react, agent orchestrator)
- Translators: encode explain_triples via TripleTranslator for
gateway wire format
- Python SDK: ProvenanceEvent now includes parsed ExplainEntity
and raw triples; expanded event_type detection
- CLI: invoke_graph_rag, invoke_agent, invoke_document_rag use
inline entity when available, fall back to graph query
- Tech specs updated
Additional explainability test
- Fix agent react and orchestrator services appending bare methods
to config_handlers instead of using register_config_handler() —
caused 'method object is not subscriptable' on config notify
- Add exc_info to config fetch retry logging for proper tracebacks
- Remove debug print statements from collection management
dispatcher and translator
- Disable RabbitMQ heartbeats (heartbeat=0) to prevent broker
closing idle producer connections that can't process heartbeat
frames from BlockingConnection
Add full pipeline test that generates a real PDF, processes it through
the entire pipeline, and verifies knowledge lands in FalkorDB:
- Create test PDF generator using pdf-lib (2-page doc about Acme Corp)
- Add testFullPipeline() to integration tests with store verification
- Fix FalkorDB client connect() — createClient returns unconnected client
in both TriplesStore and TriplesQuery classes
Results: PDF decoded (2 pages) → chunked (2 chunks) → extracted
(4 relationships) → 16 triples stored in FalkorDB including:
alice-johnson → is-a-senior-engineer → acme-corporation
cloudsync → uses-aws-for-hosting → amazon-web-services
provenance: pages → prov:wasDerivedFrom → source document
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Two bugs found during end-to-end testing:
1. FlowProcessor never restarted flows when config changed — it only
started them once. Stale NATS JetStream data from previous sessions
caused services to bind to wrong topics. Fix: stop and restart flows
on every config push that includes flow definitions.
2. Gateway publishToTopic sent messages without an id property. Pipeline
FlowProcessor handlers check properties.id and silently return if
missing. Fix: auto-generate a message id when publishing to topics.
Both fixes validated: 13/13 integration tests passing, PDF decoder
correctly receives and processes document messages through the pipeline.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Wire up the query and retrieval side of the pipeline so the agent can
answer questions from stored knowledge:
- Triples query service (FalkorDB) — all SPO pattern queries via NATS
- Graph embeddings query service (Qdrant) — entity vector similarity
- Document embeddings query service (Qdrant) — chunk vector similarity
- Graph RAG service — full concept→entity→traverse→score→synthesize pipeline
- Document RAG service — embed→find chunks→synthesize pipeline
- Runner scripts for chunker, extractor, embeddings (missing from Phase 5)
- Add DocumentEmbeddingsRequest/Response schema types
- Add RAG prompt templates (extract-concepts, edge-scoring, synthesize)
- Add graph/doc embeddings query topics to seed config + flow manager
- Add all pipeline/query/retrieval services to docker-compose
- 8 new runner scripts, 8 new pnpm script aliases
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add end-to-end document processing pipeline:
- PDF decoder service (pdfjs-dist) extracts text per page from librarian docs
- Ollama native LLM service for local model inference
- FalkorDB triples store FlowProcessor consumer
- Qdrant graph embeddings store FlowProcessor consumer
- Fix spec name collisions in chunker/extractor (input→chunk-input, etc.)
- Gateway /load endpoint to trigger document processing
- Align flow manager blueprint and seed config with full pipeline topics
- Add runner scripts and test coverage for document load
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Marks FlowProcessor and EmbeddingsService constructors as protected
since these classes should only be instantiated via subclasses.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace the config push mechanism that broadcast the full config
blob on a 'state' class pub/sub queue with a lightweight notify
signal containing only the version number and affected config
types. Processors fetch the full config via request/response from
the config service when notified.
This eliminates the need for the pub/sub 'state' queue class and
stateful pub/sub services entirely. The config push queue moves
from 'state' to 'flow' class — a simple transient signal rather
than a retained message. This solves the RabbitMQ
late-subscriber problem where restarting processes never received
the current config because their fresh queue had no historical
messages.
Key changes:
- ConfigPush schema: config dict replaced with types list
- Subscribe-then-fetch startup with retry: processors subscribe
to notify queue, fetch config via request/response, then
process buffered notifies with version comparison to avoid race
conditions
- register_config_handler() accepts optional types parameter so
handlers only fire when their config types change
- Short-lived config request/response clients to avoid subscriber
contention on non-persistent response topics
- Config service passes affected types through put/delete/flow
operations
- Gateway ConfigReceiver rewritten with same notify pattern and
retry loop
Tests updated
New tests:
- register_config_handler: without types, with types, multiple
types, multiple handlers
- on_config_notify: old/same version skipped, irrelevant types
skipped (version still updated), relevant type triggers fetch,
handler without types always called, mixed handler filtering,
empty types invokes all, fetch failure handled gracefully
- fetch_config: returns config+version, raises on error response,
stops client even on exception
- fetch_and_apply_config: applies to all handlers on startup,
retries on failure
* fix: prevent duplicate dispatcher creation race condition in invoke_global_service
Concurrent coroutines could all pass the `if key in self.dispatchers` check
before any of them wrote the result back, because `await dispatcher.start()`
yields to the event loop. This caused multiple Pulsar consumers to be created
on the same shared subscription, distributing responses round-robin and
dropping ~2/3 of them — manifesting as a permanent spinner in the Workbench UI.
Apply a double-checked asyncio.Lock in both `invoke_global_service` and
`invoke_flow_service` so only one dispatcher is ever created per service key.
* test: add concurrent-dispatch tests for race condition fix
Add asyncio.gather-based tests that verify invoke_global_service and
invoke_flow_service create exactly one dispatcher under concurrent calls,
preventing the duplicate Pulsar consumer bug.
* fix: prevent duplicate dispatcher creation race condition in invoke_global_service
Concurrent coroutines could all pass the `if key in self.dispatchers` check
before any of them wrote the result back, because `await dispatcher.start()`
yields to the event loop. This caused multiple Pulsar consumers to be created
on the same shared subscription, distributing responses round-robin and
dropping ~2/3 of them — manifesting as a permanent spinner in the Workbench UI.
Apply a double-checked asyncio.Lock in both `invoke_global_service` and
`invoke_flow_service` so only one dispatcher is ever created per service key.
* test: add concurrent-dispatch tests for race condition fix
Add asyncio.gather-based tests that verify invoke_global_service and
invoke_flow_service create exactly one dispatcher under concurrent calls,
preventing the duplicate Pulsar consumer bug.
