Bug: LLM stream EOF without terminal finish is committed as a successful partial response

[andrew-stelmach-fleet]

Summary

An LLM stream that yields partial content and then reaches clean EOF without a terminal finish chunk is committed as a successful assistant response. The turn ends as completed, the partial response enters derived history, and request-error/retry policy is bypassed.

Tested on master at 47f943859bef60e4160492346772ded9b24f765a.

Minimal reproduction

I mounted the public minimal agent composition (LlmRuntime, SessionStore, SystemPrompt, ToolRuntime, AgentRegistry, and AgentLoop) with a custom LlmAdapter whose stream emits:

yield { type: 'block-start', index: 0, blockType: 'text' }
yield { type: 'text-delta', index: 0, text: 'partial response' }
return // no terminal finish chunk

Then I submitted one ordinary agent.followup() and waited for idle.

Observed result:

{
  "assistantMessage": "partial response",
  "turnEnd": { "kind": "completed" },
  "derivedHistoryContainsPartialReply": true
}

Repeating the same reproduction with InvariantRegistry and @deepseek-ai/dsh-llm/invariant explicitly mounted correctly produces no assistant message and ends the turn with an invariant error. This means the safety check exists only as optional diagnostics, not at the required runtime boundary.

Expected behavior

Every composition should fail closed when an adapter or llm/stream middleware reaches EOF without exactly one terminal finish chunk:

• do not append assistant/message;
• surface a stable structured request failure;
• allow the installed retry policy to handle the failure;
• do not retain truncated text or incomplete tool calls in model-visible history.

Root cause

• adapterStream() accepts iterator.next().done as a normal return.
• BlockAssembler.finish defaults an absent finish to { kind: 'stop' }.
• The agent loop therefore passes request-error handling and commits the assembled message.

This conflicts with the adapter contract that every exposed outcome ends in exactly one terminal finish chunk: packages/llm/llm/README.md and StreamChunk.

The shipped base composition mounts the LLM runtime and agent loop but not the invariant registry/companions, so ordinary product and custom compositions can take the fail-open path.

Impact

A buggy third-party adapter, clean-EOF transport bug, or stream middleware that accidentally suppresses the terminal chunk can durably turn truncated model output into a successful response. This bypasses retry behavior and contaminates later model context. An incomplete tool-call stream may also be materialized with defaulted fields and reach tool handling.

Verification

The disposable real-loop reproduction was removed after execution. Focused existing tests also demonstrate the split contract:

pnpm exec vitest run packages/llm/llm/tests/properties.spec.ts \
  -t "finish reflects the last finish chunk, or defaults to stop when none arrives"

pnpm exec vitest run packages/llm/llm/tests/invariant.spec.ts \
  -t "rejects malformed stream"

Both pass: assembler/property coverage currently codifies missing finish as stop, while the optional invariant rejects the same malformed stream.

Acceptance criteria

• Clean EOF without a terminal finish produces a structured failure at the normal runtime boundary.
• No partial assistant/message or tool call is committed.
• Request-error/retry policy receives the failure.
• The behavior is correct without mounting optional invariant diagnostics.
• Focused tests cover both adapter EOF and middleware-suppressed finish.

I searched all current Discussions before posting. Discussions #199, #231, and #232 cover different reasoning-content and iterator-cleanup failures; I found no report for clean EOF being accepted as a successful completion.