Skip to content

Claude Code. Any Model. The most powerful AI coding agent now speaks every language.

fix(openai-codex): send an SSE error when a socket drops inside an exec

Jack Rudenkoa1b21969be68
diff --git a/ai-docs/architecture/code-mode.md b/ai-docs/architecture/code-mode.md
index 4733bbe2..1e0a59c7 100644
--- a/ai-docs/architecture/code-mode.md
+++ b/ai-docs/architecture/code-mode.md
@@ -147,6 +147,9 @@ downstream Claude message) — the same mechanism that retries a degenerate leg
`EXEC_DRAIN_MS = 10_000` (R1: the terminal event arrives milliseconds after the exec item), so the
outcome carries its usage. An exec cut by `response.incomplete` is never run: the model gets
`Script failed … exec input was truncated by max_output_tokens` and can send a shorter script.
+ An exec cut by a dropped socket (or a body that ends inside it) ends the message with an SSE
+ `error`, as a cut classic tool block does: the script never ran and nothing reached Claude Code,
+ so its retry is safe, where `end_turn` would stop the agent.
- The hooks (`code-mode/relay.ts`) create the record, pin the session and run the script:
- **blocked** on N nested calls → N whole `tool_use` blocks `toolu_cx_<hex>_<seq>` (Claude Code's
tool names; buffered tools repaired through Layer 4; `onToolCallObserved` once per nested
diff --git a/packages/cli/src/handlers/shared/stream-parsers/responses/responses-leg.ts b/packages/cli/src/handlers/shared/stream-parsers/responses/responses-leg.ts
index 9d63796d..785eca07 100644
--- a/packages/cli/src/handlers/shared/stream-parsers/responses/responses-leg.ts
+++ b/packages/cli/src/handlers/shared/stream-parsers/responses/responses-leg.ts
@@ -84,7 +84,14 @@ export type LegOutcome =
usage?: Usage;
}
| { kind: "api-error"; code: string; message: string; usage?: Usage }
- | { kind: "stream-error"; error: unknown; openToolBlock: boolean; usage?: Usage }
+ | {
+ kind: "stream-error";
+ error: unknown;
+ openToolBlock: boolean;
+ /** Code mode: the socket died while the exec input streamed (never a block, never run). */
+ openExec?: boolean;
+ usage?: Usage;
+ }
| { kind: "cancelled"; usage?: Usage };
export interface ResponsesLegContext {
@@ -377,7 +384,13 @@ class ResponsesLeg {
// what the client does with the fragment (see the driver).
const openToolBlock = this.openToolBlocks.size > 0;
this.closeTools();
- return { kind: "stream-error", error, openToolBlock, usage: this.usage };
+ return {
+ kind: "stream-error",
+ error,
+ openToolBlock,
+ ...(this.heldExec ? { openExec: true } : {}),
+ usage: this.usage,
+ };
} finally {
guard?.disarm();
signal.removeEventListener("abort", onAbort);
@@ -400,6 +413,7 @@ class ResponsesLeg {
kind: "stream-error",
error: new Error("the stream ended inside the exec call"),
openToolBlock: false,
+ openExec: true,
usage: this.usage,
};
}
diff --git a/packages/cli/src/handlers/shared/stream-parsers/responses/responses-turn.ts b/packages/cli/src/handlers/shared/stream-parsers/responses/responses-turn.ts
index e3d3c107..79353986 100644
--- a/packages/cli/src/handlers/shared/stream-parsers/responses/responses-turn.ts
+++ b/packages/cli/src/handlers/shared/stream-parsers/responses/responses-turn.ts
@@ -605,7 +605,10 @@ class ResponsesTurn {
// message with no completed tool_use, and it is the honest report — the
// turn did fail. devin-connect ends a mid-stream fault the same way.
// With no tool call in flight the inline text keeps partial prose visible.
- if (outcome.openToolBlock) {
+ // Code mode: an exec cut mid-input is a tool call in flight too. It was
+ // never a block and never ran, so Claude Code's retry is safe — and
+ // `end_turn` would stop the agent where classic mode retries.
+ if (outcome.openToolBlock || outcome.openExec) {
log("[ResponsesSSE] tool arguments cut off mid-stream → error event");
this.sink.endWithError(`Stream error: ${outcome.error}`);
} else {