[Bug]: deepseek-v4-flash-free can emit incomplete tool calls that break streamed turns

[yxr1995-maker]

Current status

This issue is now narrowed to the remaining OpenCode Zen / DeepSeek streaming compatibility defect.

The startup syntax crash reported against 2.12.0 is already resolved on current dev: the invalid continue outside a loop is no longer present. That part should not be reimplemented.

Current dev also already accepts [DONE] without a finish_reason: the [DONE] branch flushes pending tool calls and emits done. A raw EOF with neither [DONE] nor finish_reason remains a truncation condition and should stay fail-closed unless provider evidence proves a safe alternative.

Remaining defect

openai-chat can still flush a pending tool call with an empty function name.

Current flushToolCalls() emits:

yield { type: "tool_call_start", id: call.id, name: call.name };

without first proving that call.name is non-empty.

The affected OpenCode Zen route can stream argument deltas while omitting the matching function.name. When that incomplete call is flushed, the downstream Codex tool-call contract receives an unusable call and the turn can fail.

This issue now tracks only the provider-compatibility question around incomplete streamed tool-call metadata.

Reproduction

1. Route Codex through opencode-free/deepseek-v4-flash-free using the openai-chat adapter.
2. Run a turn that causes the upstream to emit tool-call argument deltas without a usable function.name.
3. Allow the stream to reach a normal flush boundary such as [DONE].
4. Observe the pending call being emitted with an empty name and the downstream turn failing.

Expected behavior

OpenCodex must not emit a completed usable tool call with an empty name.

A fix must preserve the existing fail-closed rules for genuinely malformed or truncated tool calls. In particular:

• do not silently invent a function name;
• do not treat arbitrary raw EOF as successful completion;
• do not weaken the nested tool_calls validation from [Bug] Nested field shapes bypass the #1219 frame guard and crash both adapter paths #1325;
• do not silently drop a claimed tool call if doing so can orphan a matching result;
• provider-specific normalization is acceptable only when the upstream shape is proven semantically safe.

Resolved portion

The 2.12.0 startup failure:

Cannot use "continue" here

is no longer present on current dev and is closed as part of this issue's history.

Related

• [Bug] Nested field shapes bypass the #1219 frame guard and crash both adapter paths #1325 - generic nested tool_calls validation and fail-closed behavior.
• [Compatibility] Xiaomi MiMo v2.5 fails on 2.12.0 with invalid tool calls and an incorrect reasoning ladder #1483 - Xiaomi MiMo compatibility report. Similar symptom, different provider evidence and reasoning-effort scope.
• Bug: deepseek-v4-flash-free upstream returns malformed tool_calls causing "invalid tool calls" error #1472 / [Bug]: deepseek-v4-flash-free upstream returns malformed tool_calls causing "invalid tool calls" error #1473 - earlier auto-closed reports of the OpenCode Zen symptom.

Acceptance criteria

• A pending streamed tool call with no usable function name is never emitted as a valid completed tool call.
• Normal named tool calls remain unchanged.
• [DONE] without finish_reason remains accepted.
• Raw EOF without a terminal signal remains fail-closed unless separately proven safe.
• Malformed nested tool_calls shapes remain fail-closed.
• Focused regression coverage reproduces the OpenCode Zen incomplete-name shape without weakening generic tool-call safety.

[Wibias]

Triage update: I narrowed this issue to the one defect that still exists on current dev.

The startup parser crash is already fixed, and [DONE] without finish_reason is already accepted. The remaining actionable problem is that flushToolCalls() can emit a pending call with an empty function name.

I updated the title/body and labels accordingly. Please keep any fix focused on incomplete tool-call metadata and preserve the existing fail-closed behavior for genuinely malformed or truncated streams.

[github-actions]

Maintainer decision respected

A maintainer has reopened this issue. The automated closure has been deactivated.

[lidge-jun]

Fixed on dev by #1531, merged as 4e0ffb2cbef0d3724a3aa9ebef26279aed44b694.

The narrowed defect is closed: flushToolCalls() no longer emits a pending call whose function.name never arrived. An unnamed streamed call now terminates the turn through the adapter error channel, so the downstream tool-call contract never receives a call it cannot dispatch.

The constraints listed in this issue were kept:

• no invented function name — the id is still synthesized when absent because it is an opaque correlation handle, but a name is not;
• raw EOF stays fail-closed and still reports truncation rather than the new error;
• the [Bug] Nested field shapes bypass the #1219 frame guard and crash both adapter paths #1325 nested tool_calls validation is unchanged and still terminal;
• the call is not silently dropped, so a matching result cannot be orphaned.

Two things were found while implementing it and are included:

1. The buffered parseResponse validator checked the type of name but not its emptiness, so a blank name was emitted there too. It is now rejected the same way.
2. The delta.tool_calls cast is a TypeScript convenience over upstream JSON, so a truthy non-string name reached the accumulator unvalidated and threw during string handling. Present function, name, arguments, and id fields are now runtime-checked at ingest and routed to the existing invalid-tool-calls error.

Verification at the merged tree: bun x tsc --noEmit exit 0; focused suites 48/48; 11-file adapter sweep 161/161; full bun run test 11,310 pass / 0 fail across 698 files; all four Linux CI shards plus macOS green on the PR.