feat(llm): configurable first-token timeout

[lin-snow]

Summary

Adds an opt-in, per-node first_token_timeout and threads it through graphon's LLM invoke path. graphon carries the value from node config down to the LLMProtocol invoke calls — it does not enforce the timeout. Enforcement (closing the connection when the first token is late, then raising FirstTokenTimeoutError) is the host transport adapter's job (e.g. Dify's ModelInstance), delivered separately.

How it fits together

flowchart LR
    subgraph G["graphon — this PR (carrier only)"]
        direction TB
        CFG["LLM-family node data:<br/>invocation.first_token_timeout (ms, 0 = disabled)"]
        NODE["LLM / QuestionClassifier /<br/>ParameterExtractor node"]
        PROTO["LLMProtocol.invoke_llm(<br/>..., first_token_timeout=...)"]
        CFG --> NODE --> PROTO
    end
    subgraph H["host adapter — e.g. Dify (separate, later phase)"]
        direction TB
        ADP["ModelInstance / DifyPreparedLLM"]
        GATE{"first token<br/>within timeout?"}
        ADP --> GATE
        GATE -->|"yes"| OK["stream normally"]
        GATE -->|"no"| ERR["raise FirstTokenTimeoutError"]
    end
    PROTO -->|"protocol call — contract boundary"| ADP
    ERR -.->|"bubbles up as node failure"| STRAT["node error-strategy:<br/>retry / fail-branch / default"]

Loading

Changes

File	Change
model_runtime/errors/invoke.py	New FirstTokenTimeoutError(InvokeConnectionError) — canonical kernel type for hosts to raise/import. Defined here, not raised by graphon (carrier only).
nodes/llm/entities.py	New LLMInvocationConfig value object: first_token_timeout: int = Field(default=0, ge=0) (ms; 0 = disabled, negatives rejected at validation) + first_token_timeout_seconds property (maps 0 to None). Composed onto the LLM / question-classifier / parameter-extractor node data as an invocation field — same idiom as ModelConfig / VisionConfig. RetryConfig / base_node_data.py are untouched — the deadline is a client-side invocation concept, not a retry one, and must not sit on the shared base that every node inherits.
nodes/llm/runtime_protocols.py	Optional keyword-only first_token_timeout: float | None = None on LLMProtocol.invoke_llm and invoke_llm_with_structured_output.
nodes/llm/node.py	LLMNode.invoke_llm forwards the value; _invoke_llm_for_run reads self.node_data.invocation.first_token_timeout_seconds (non-polling path).
nodes/question_classifier/…, nodes/parameter_extractor/…	Node data composes LLMInvocationConfig; each node reads invocation.first_token_timeout_seconds and passes it through.
dsl/slim/llm.py	SlimLLM accepts the parameter for protocol conformance; does not enforce it (v1 standalone/DSL limitation).

Compatibility & release ordering ⚠️

Backward compatible at graphon's config / serialization layer: workflows saved before this field deserialize fine (invocation defaults to a config with first_token_timeout = 0 = disabled; no migration).

It is not transparent at the host call contract. LLMNode.invoke_llm passes first_token_timeout= unconditionally (even when unset, it passes None). Any LLMProtocol implementation that lacks the parameter and lacks **kwargs raises TypeError on every LLM call — not only when a timeout is configured.

graphon (caller)	host LLMProtocol impl	Result
old (never passes the kwarg)	old or new	✅ works
new (this PR)	old — no param, no **kwargs	❌ TypeError on every LLM / QC / PE invoke
new (this PR)	new — accepts the kwarg	✅ works

Ordering requirement: the host adapter must accept first_token_timeout (accept-and-ignore is enough) before the host bumps its graphon pin to a version containing this PR. graphon's own merge/release is independent — the only gate is the host's pin bump.

For Dify specifically: DifyPreparedLLM (api/core/workflow/node_runtime.py) currently has neither the parameter nor **kwargs and pins graphon==0.6.0, so its accept-and-ignore shim must land before Dify bumps the pin.

flowchart LR
    A["host: adapter accepts kwarg<br/>(accept-and-ignore shim)"] --> B["host: bump graphon pin<br/>to a version with this PR"]
    B --> C["host: implement real enforcement<br/>+ raise FirstTokenTimeoutError"]

Loading

Out of scope

Transport-level enforcement; the lower LLMModelRuntime protocol (not on the node invoke path); the LLM polling branch (start_llm_polling has no such parameter — intentionally not threaded, pinned by a test).

Tests

LLMInvocationConfig defaults / ms→s conversion / 0-disables / a negative value rejected at validation (ge=0), plus a guard asserting the three LLM-family node data classes compose an invocation field and that the deadline does not leak onto RetryConfig / BaseNodeData; forwarding assertions covering both a configured value and the default None for the LLM, question-classifier, and parameter-extractor nodes (question-classifier and parameter-extractor drive it end-to-end through model_validate({"invocation": {...}})); a test pinning that the polling path does not receive the parameter. just check (ruff + ty) and the full suite (618 tests) pass.