Why: The r0.4 trial could prove that the parser missed [DONE] but could not distinguish clean EOF, malformed frames, read failure, or provider error without retaining raw SSE.
Evidence: RED receipts fast-slow-u3-r05-task1-terminal-red.json and task1-http-error-red.json; GREEN targeted receipt and full vet/test/race receipt fast-slow-u3-r05-task1-agent-native-full.json.
Impact: Add bounded scalar diagnostics and successful HTTP status to terminal events while preserving the existing incomplete-tool rejection, returned Usage, and retry behavior.
Refs: U3 r0.5 approved design and plan; Agent-Native PR #46
Why: The parser also recognizes function_call as a tool finish, but cannot execute a legacy call when no deltas were collected. It must fail closed like tool_calls.
Evidence: The function_call variant failed before this guard and passed afterward; full Agent-Native vet/test passed.
Impact: Inconsistent legacy tool terminals fail with returned Usage preserved; valid streams are unchanged.
Refs: U3 r0.4 independent review Important 2 closure
Why: A provider can report finish_reason=tool_calls without emitting any tool call. Treating it as a successful empty answer hides an incomplete operation.
Evidence: New SSE regression failed before the guard and passed afterward; full go vet and go test -count=1 ./... passed.
Impact: Inconsistent tool terminals fail closed while preserving returned Usage; valid text and tool streams are unchanged.
Refs: U3 r0.4 independent review issue 2
Why: a model tool finish could execute partial JSON or an unterminated SSE stream, while post-Usage errors could lose their charge.
Evidence: missing-DONE, malformed-JSON, identity, non-object and post-Usage RED→GREEN tests; GOWORK=off go vet ./... and go test -count=1 ./... passed.
Impact: invalid tool terminals fail without execution or automatic retry and retain returned Usage; valid text and multi-part tool streams keep working.
Refs: runtime/llm/sse.go; runtime/loop.go
Why: Interleaved-thinking models (MiniMax, DeepSeek R1) require the
assistant's thinking to be passed back verbatim in message history; dropping
it degrades later turns. The product's P2-I1 replayed thinking only into the
request-side rebuild; with the session transcript becoming the authoritative
context (quality P2-I2), capture must live in the kernel so thinking persists
in the transcript and replays on every request.
Evidence: SELF - TestReasoningCaptureAndReplay covers capture (SSE
reasoning_content into Message.Thinking, content unpolluted) and replay
(<think> prefix on thinking assistants, untouched non-thinking messages);
go vet + go test ./... pass.
Impact: sse.go accumulates reasoning deltas into the terminal message's
Thinking; client.go toProviderMessages renders captured thinking as a
<think> block prefix for assistant messages. Anthropic path unchanged;
messages without thinking serialize exactly as before.
Why: Buffering a whole attempt hid live output, overflow detection missed common provider errors, and a cut-off tool call could still run.
Evidence: GOWORK=off go test -race -count=1 ./runtime/llm/ ./runtime/ passed.
Impact: WithRetry forwards events immediately and retries only when nothing has been forwarded. Length or an unclean end drops unfinished tool calls. Loop usage keeps a billing sum and the last request separately.
Refs: feature/llm-robustness
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: the race detector can deliver message_end when the context is already cancelled, because select picks either ready case.
Evidence: GOWORK=off go test -count=50 -race ./runtime/llm/ passed, including TestParseSSECancel.
Impact: a cancelled parse ends with the aborted event.
Refs: D1
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: token budget let the current tool batch run, compaction could split a tool pair or race an append, writes were not crash-atomic, and several MCP and session edges only existed as comments.
Evidence: GOWORK=off go test -count=1 -race ./runtime/ ./runtime/llm/ ./runtime/session/ ./runtime/tools/ ./runtime/mcp/ ./adapters/native/ ./adapters/intentgate/ passed, including the TestR2* cases for budget, replacement, cancel, compaction, SSE, UTF-8, schema, MCP retry, session terminate, poison reload, and registry reload.
Impact: over-budget tool calls are skipped and paired; Stop.Replacement is appended to the session; the intent gate can follow Registry.Reload.
Refs: N5, N6, N7, N8, N9, N11, N12, N15, N16, N17, N18, N19, N21, I9
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: the direct chat path needs one shared loop for deliverable checks, tool pairing, reminders, and truncated replies, while squad callers keep today's behavior when the new fields stay unset.
Evidence: go test ./... in agent-native passed, including Stop cap, BeforeTool pairing, reminders absent from the session, repeat-guard cap, length continuation, and zero-value loop tests.
Impact: runtime.Config and llm.Event grow optional fields; AgentSession.Prompt copies them through. Squad sessions that leave them zero are unchanged.
Refs: feature/agent-harness-p1
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: The runtime could only speak OpenAI chat completions, so thinking levels, images, and the other vendor wires had nowhere to go.
Evidence: go test -count=1 ./runtime/llm/
Impact: Existing chat-completions calls stay the same unless a thinking level or image is set. Overflow errors are not retried.
Refs: feature/pi-headless N2
Co-authored-by: Cursor <cursoragent@cursor.com>
agent-exec-transport
weichen