[wangyeyu1129]

Description:

A response could end successfully at the transport layer while generation was truncated. AgentSession previously continued only after tool calls, so a no-tool token limit or provider pause ended the user turn early. Carry a provider-neutral stop reason and continue within a bounded per-turn budget.

Fixes:

No linked issue. Implements the requested stop-reason continuation contract; the paired vdoc-server integration pins local SDK packages until an upstream release is available.

Behaviours Completed:

• Add optional stopReason and rawStopReason to response endings and cascade text finishes; export StopReason from core while preserving existing response statuses.
• Normalize OpenAI-compatible completion, tool calls, output limits, filtering, cancellation, failure and unknown outcomes. Keep raw reasons to short enum tokens; cancellation/failure status takes precedence.
• Continue max_tokens and pause even without tools. Wait for every issued tool result before deciding; preserve the existing validation, authorization, serial execution and interruption gates.
• Stop and report refusal, content filtering, failure, unknown reasons or tool_calls without actual calls. Providers that omit reasons retain the previous inference behavior.
• Default to 16 continuations per user turn, validate maxContinuations, reset the budget on new input, and report continuation_limit when exhausted.
• Ignore repeated response endings and stale epochs; discard unfinished tool argument streams without executing them. Publish one final turn hook with the standardized stop reason and continuation count.

Files Changed:

• packages/core/src/{events,agent,index,session}.ts: public contract, metrics, bounded scheduling and lifecycle guards.
• packages/openai-compatible/src/index.ts: provider mapping and short raw-reason preservation.
• packages/cascade/src/index.ts: finish propagation after synthesized audio completes; preserve conversation history for continuations.
• Core, Boson and cascade tests: retryable and terminal reasons, tool settlement, duplicate endings, budgets, interruption, history and audio ordering.

Brief Explanation:

Transport status remains independent of generation stop reason. The runtime makes one decision only after the response has ended and its issued calls have settled; subsequent generations stay in the same turn. Only real tool feedback uses autoReplyAfterResult. No fabricated user messages, write replay, provider network retries or browser wire-protocol changes are introduced. This branch is based on current upstream main, including bounded cascade history.

Testing:

• npm run typecheck: passed. npm test: 591 tests across 34 files passed on current upstream main.
• npm run pack, npm run check:budgets, npm run check:spec and git diff --check: passed.
• Regression coverage exercises max_tokens → pause → completed, terminal outcomes with and without tools, budget reset/default/zero, late interrupted responses, duplicate usage and preserved assistant/tool history.

Unsure About:

• Real Boson/Claude API behavior and speech continuity require a separate credentialed integration run; simulated tests do not claim those are verified.
• completed means generation ended, not that a business task or document write succeeded. Tool results and application verification retain that responsibility.
• Providers that omit stop reasons keep legacy behavior and cannot identify native truncation. Existing provider auto-reply capability contracts and timeout behavior remain in force.