Reject incomplete upstream SSE instead of emitting message_stop

Overview

Impact: High. Priority: P1. Estimated effort: Small/Medium.

An upstream HTTP200 event stream that ends cleanly after partial output, without a finish reason or [DONE], becomes a successful Anthropic end_turn / message_stop. The editor can accept a truncated answer or partial tool arguments as a normal completion.

Current state and reproduction

Audited main 093cd44806ec64b2c79b42b557b4856259d2ea58:

• [DONE] is ignored instead of recorded.
• Any successful HTTP completion invokes StreamClose.
• Missing finish reason defaults to successful end_turn.

Hermetic loopback test compiled unchanged messages.cpp, translate.cpp, and loopback_auth.cpp with repo-pinned cpp-httplib0.46.1/nlohmann3.11.3; only output/state-directory functions stubbed. Mock upstream returned HTTP200, content-type text/event-stream, exactly this body and a clean HTTP end:

data: {"id":"x","choices":[{"delta":{"content":"partial"},"finish_reason":null}]}

Authenticated downstream /v1/messages requested stream:true. Actual output assertion:

TRUNCATED_WITHOUT_DONE status=200 success_stop=1 error=0

success_stop checks event: message_stop; error checks event: error. This is clean HTTP EOF with an incomplete inference protocol, not a TCP read failure. No real API/production traffic used.

Proposed solution and implementation

• Track protocol terminal state independently of HTTP transport success.
• Require the documented upstream terminal condition before emitting a successful Anthropic close.
• Surface incomplete streams as typed errors, retaining partial output only as partial output.
• Do not flush incomplete buffered tool arguments as executable tool use.
• Add tests for empty EOF, partial text EOF, partial tool JSON EOF, missing/malformed terminal frame, and normal finish/usage/DONE ordering.

Success criteria

A clean transport EOF before valid inference termination never creates a successful message_stop or executable incomplete tool call. Valid completed streams still produce exactly one message_stop and final usage. Public API fault injection remains deferred until API access is supplied.

Related: #82 modifies this stream transport. This issue adds no release-blocker label; release triage is deferred.

Implementation handoff (2026-09-13)

PR #104 implements this issue; it remains open for acceptance. Merge order: #82 -> #103 -> #104. Current PR104 base is codex/issue83-stream-overload; after PR103 merges, rebase only PR104's commits and retarget to main. Siddhesh owns the merge. Full validation, rationale, limitations and remaining packaged/live checks are in PR104 and the implementation report.

Implementation and merge status

Implemented in PR #104 at ea4eb4ac2bec2f8eb596e760fd1df9a300c73417; final independent published-head review found no unresolved blocking findings. See the final review comment for exact tests and remaining platform/live acceptance. Keep this issue open.

Merge owner: @Siddhesh2377. Sequence: #82 → #103 → #104, ultimately main. Rebase and retarget dependent PRs after prerequisites merge, then rerun checks. Nothing has been deployed by this implementation work.