Replace the M3 stub in `agents session tell` with real orchestrator actor
invocation via `SessionWorkflow.tell()`. The stub echoed a canned
`Acknowledged: ...` response without calling any LLM or actor; this
commit wires up the full pipeline:
Architecture:
- New `SessionWorkflow` (Application layer) orchestrates LLM invocation
for session tell. Accepts a `ProviderRegistry` and `ToolRegistry`;
falls back to `FakeListLLM` when no provider is configured.
- `LangChainSessionCaller` implements the `LLMCaller` protocol, building
history-aware LangChain message lists from the session conversation and
invoking the LLM via `ToolCallingRuntime.run_tool_loop()`.
- `TellResult` (Pydantic BaseModel) carries the assistant response plus
token usage (input_tokens, output_tokens, cost_usd, duration_ms).
Domain / service layer:
- `SessionActorNotConfiguredError` added to the session domain model;
raised when tell is invoked with no actor on the session and no
`--actor` override. Clear message; CLI exits with code 1.
- `SessionService.get_messages()` abstract method added (+ implementation
in `PersistentSessionService`) to load ordered message history.
A2A facade:
- `A2aLocalFacade` gains `message/send` and `message/stream` standard
A2A operation handlers that route to `SessionWorkflow.tell()`.
Total supported operations count: 42 → 44.
CLI:
- `session tell` command delegates to `_build_session_workflow()`
(patchable factory) instead of directly calling `SessionService`.
- Non-streaming: `SessionWorkflow.tell()` returns `TellResult`; output
includes a Usage panel (Rich/Plain) or a `usage` object (JSON/YAML).
- Streaming: `SessionWorkflow.tell_stream()` yields tokens; CLI prints
them via `console.print` (not raw `sys.stdout.write`).
- Token usage recorded via `SessionService.update_token_usage()` in
both paths.
Tests:
- New Behave feature `session_tell_llm.feature` (4 scenarios): real LLM
response persisted; streaming yields tokens; no-actor exits code 1;
`--actor` override resolves correctly.
- New Robot suite `session_tell_llm.robot` (4 tests): end-to-end with
stub LLM injected via monkey-patching `_resolve_llm`.
- Updated all existing `session tell` test steps to patch
`_build_session_workflow` returning a mock `TellResult` so tests
remain isolated from the LLM layer.
- Updated operation-count assertions (42 → 44) in
`a2a_cli_facade_integration`, `consolidated_misc`,
`m6_autonomy_acceptance` feature files and steps.
Cycle 7 (Review ID 8088) fixes:
- Blocker 4: Split `session_workflow.py` (was 801 lines) into two files:
`session_workflow.py` (467 lines) and `session_caller.py` (318 lines).
Extracted: `LangChainSessionCaller`, `extract_content`,
`extract_token_usage`, `estimate_cost`, `estimate_tokens`,
`history_to_langchain_messages`, and stub classes.
- Blocker 5: Route CLI streaming path through
`_facade_dispatch("message/stream", ...)` instead of directly
calling `workflow.tell_stream()`. The facade's
`_handle_message_stream` falls back to non-streaming with
`streamed: false` (acceptable for this milestone per the spec).
- Updated streaming test assertion to validate full response presence
(no longer checks token-by-token word positions, since the facade
fallback returns a complete message).
ISSUES CLOSED: #5784
