[Bug] Zed Claude models 400: missing field `is_error` on tool_result after first tool round-trip

[AMSeify]

Summary

When routing Zed IDE → 9router → Zed Hosted AI with Claude models (e.g. zd/claude-sonnet-5-5), the first turn succeeds, but the next turn that includes tool results fails:

[400]: failed to parse Anthropic request: missing field `is_error`
URL: https://cloud.zed.dev/completions

Repro

1. Connect a Zed OAuth account in 9router.
2. Point Zed IDE (or any OpenAI-format client) at 9router with a Claude model via the Zed provider.
3. Ask something that triggers tools (e.g. list files).
4. First completion succeeds; the follow-up with role: tool messages returns 400.

Log pattern:

▶ POST zd/claude-sonnet-5-5 → zed/claude-sonnet-5-5 · FMT: openai→openai · … · 4 MSG · 12 TOOL
📊 DONE …
▶ POST … · 6 MSG · 12 TOOL
✗ ERROR 400 · … missing field `is_error`

Root cause

open-sse/executors/zed.js converts OpenAI Chat Completions → Anthropic for Claude models via openaiToClaudeRequest, then posts that as provider_request to cloud.zed.dev/completions.

Zed’s Anthropic wire types require is_error: bool on every tool_result (non-optional). Anthropic’s public API treats it as optional (default false), so our translator omitted it:

• OpenAI role: "tool" → Claude tool_result never set is_error
• Claude-shaped passthrough only spread is_error when truthy (...(part.is_error && { is_error }))

Fix

Always emit is_error (boolean) on every tool_result produced by openai-to-claude.js. Harmless for Anthropic’s public API; required for Zed’s strict serde parser.

Related

No existing issue/PR specifically for this Zed is_error 400. Closest known translator gap: OpenAI-bridge loss of tool_result.is_error documented in tests/translator/bugs-openai-bridge.test.js / tests/translator/AGENTS.md (Claude→OpenAI direction).