Cursor provider: most live-catalog models fail with ERROR_BAD_MODEL_NAME (Unknown model ID)

[matheusmaais]

Summary

Every Cursor model returned by the live picker except composer-* and bare gemini ids fails immediately with:

Cursor agent stream error: {"code":"not_found","details":[{"debug":{"details":{"analyticsMetadata":{"actionRequired":"config"},"detail":"Unknown model ID: cursor-grok-4.6-high-fast",...},"error":"ERROR_BAD_MODEL_NAME","isExpected":true},...}]}

Reproduce: jcode provider-doctor cursor --tier full --model cursor-grok-4.6-high-fast (or pick any grok/claude/gpt model from /model and send a message).

Root cause

jcode-provider-cursor-runtime's hand-rolled protobuf/HTTP2 transport (agent_transport.rs, talking directly to agent.v1.AgentService/Run on agentn.global.api5.cursor.sh) sends the composite catalog id (e.g. cursor-grok-4.6-high-fast, claude-opus-5-thinking-high, gpt-5.6-sol-xhigh) verbatim as the model name.

The real backend — the same one the official cursor-agent CLI talks to with the same API key — only accepts the bare base id (grok-4.6, claude-opus-5, gpt-5.6-sol), with effort/fast/thinking sent as separate parameters. This is visible directly in ~/.cursor/cli-config.json's modelParameters, which is keyed by the bare id.

I confirmed this by running cursor-agent --model cursor-grok-4.6-high-fast -p "..." (the official CLI) with the exact same CURSOR_API_KEY jcode was using — it worked fine — while jcode's transport failed on the identical model id.

Fix

I already wrote, tested, and verified the fix locally. resolve_model_id() in agent_transport.rs strips, in this order, before building the wire request:

1. a leading cursor- vendor prefix
2. a trailing -fast suffix (captured separately as the fast flag)
3. trailing effort/mode suffixes (-xhigh, -high, -medium, -low, -thinking), which can stack (e.g. claude-opus-5-thinking-high → claude-opus-5)

Ids that don't match any pattern (composer-2.5, gemini-3.1-pro, sonnet-4.6) pass through unchanged.

Live verification (real Cursor account, real API key)

model id from the live catalog	before	after
cursor-grok-4.6-high-fast	ERROR_BAD_MODEL_NAME	chat reply OK
claude-opus-5-thinking-high	ERROR_BAD_MODEL_NAME	chat reply OK
claude-sonnet-5-thinking-high	ERROR_BAD_MODEL_NAME	chat reply OK
gpt-5.4-high	ERROR_BAD_MODEL_NAME	chat reply OK
gpt-5.6-sol-xhigh	ERROR_BAD_MODEL_NAME	chat reply OK
gemini-3.7-flash-high	ERROR_BAD_MODEL_NAME	chat reply OK
composer-2.5 (already bare)	OK	OK (unaffected)

jcode provider-doctor cursor --tier full on cursor-grok-4.6-high-fast went from failing at "Non-streaming chat completion" to passing chat + streaming.

Tests

Added 6 unit tests (resolve_model_id_*) covering every composite pattern observed in the live catalog, rebased cleanly onto current master, and all 21 tests in jcode-provider-cursor-runtime pass.

Patch

I can't open a PR directly (repo's PR creation is restricted to collaborators), but the branch is public on my fork: https://github.com/matheusmaais/jcode/tree/fix/cursor-composite-model-ids (1 commit, rebased on current master, diff attached below for convenience).

Full diff

[matheusmaais]

diff --git a/crates/jcode-provider-cursor-runtime/src/agent_transport.rs b/crates/jcode-provider-cursor-runtime/src/agent_transport.rs
index f707d559b..6d5ae33c5 100644
--- a/crates/jcode-provider-cursor-runtime/src/agent_transport.rs
+++ b/crates/jcode-provider-cursor-runtime/src/agent_transport.rs
@@ -224,6 +224,67 @@ fn encode_model_meta(name: &str, fast: bool) -> Vec<u8> {
     out
 }
 
+/// Resolve a jcode/live-catalog model id (e.g. `cursor-grok-4.6-high-fast`,
+/// `claude-opus-5-thinking-high`, `gpt-5.4-high`) into the bare base id the
+/// Cursor agent backend actually accepts (e.g. `grok-4.6`, `claude-opus-5`,
+/// `gpt-5.4`), plus the `fast` flag.
+///
+/// Cursor's `GET /v0/models` catalog and jcode's static fallback list both
+/// return *composite* ids that fold effort/fast/thinking into the model
+/// string. The real `agent.v1.AgentService/Run` backend only recognizes the
+/// base id (see `~/.cursor/cli-config.json`'s `modelParameters`, which lists
+/// `grok-4.6`, `claude-opus-5`, `claude-sonnet-5`, ... as keys, each with
+/// separate `effort`/`fast`/`thinking`/`context` parameters). Sending the
+/// composite string as-is gets `ERROR_BAD_MODEL_NAME: Unknown model ID`.
+///
+/// This performs the same stripping the official CLI does before it talks to
+/// the wire protocol: drop a leading `cursor-` vendor prefix, drop a trailing
+/// `-fast` suffix (recording it as the `fast` flag), and drop a trailing
+/// effort/mode suffix (`-high`, `-xhigh`, `-medium`, `-low`, `-thinking`, or
+/// any combination such as `-thinking-high`). Ids that are already bare (no
+/// known suffix), including generic ones like `composer-2.5` or `gemini-
+/// 3.1-pro`, pass through unchanged.
+fn resolve_model_id(model: &str) -> (String, bool) {
+    let mut name = model.trim();
+
+    // Strip a `cursor-` vendor prefix (`cursor-grok-4.6-...` -> `grok-4.6-...`).
+    if let Some(stripped) = name.strip_prefix("cursor-") {
+        name = stripped;
+    }
+
+    // Strip a trailing `-fast` suffix; the `fast` param is sent separately.
+    let fast = if let Some(stripped) = name.strip_suffix("-fast") {
+        name = stripped;
+        true
+    } else {
+        false
+    };
+
+    // Strip trailing effort/mode suffixes, e.g. `-thinking-xhigh`, `-thinking`,
+    // `-high`, `-xhigh`, `-medium`, `-low`. These can stack (thinking + effort),
+    // so peel them off one segment at a time from the end.
+    const SUFFIXES: &[&str] = &["-xhigh", "-high", "-medium", "-low", "-thinking"];
+    loop {
+        let mut stripped_any = false;
+        for suffix in SUFFIXES {
+            if let Some(stripped) = name.strip_suffix(suffix) {
+                // Never strip past an empty or vendor-only remainder.
+                if stripped.is_empty() {
+                    break;
+                }
+                name = stripped;
+                stripped_any = true;
+                break;
+            }
+        }
+        if !stripped_any {
+            break;
+        }
+    }
+
+    (name.to_string(), fast)
+}
+
 /// Build the request frames for a single-shot prompt turn.
 ///
 /// Returns the ordered list of Connect frames that constitute the streamed
@@ -240,15 +301,17 @@ fn build_run_frames(prompt: &str, model: &str, cwd: &str) -> Vec<Vec<u8>> {
     inner.extend(field_varint(4, 1));
     let messages = field_ld(2, &field_ld(1, &field_ld(1, &inner)));
 
+    let (resolved_model, fast) = resolve_model_id(model);
+
     let mut req = field_str(1, "");
     req.extend(messages);
     req.extend(field_str(4, ""));
     req.extend(field_str(5, &conv));
-    req.extend(field_ld(9, &encode_model_meta(model, false)));
+    req.extend(field_ld(9, &encode_model_meta(&resolved_model, fast)));
     req.extend(field_varint(12, 0));
     // minimal catalog: a "default" entry plus the target model
     req.extend(field_ld(14, &field_str(1, "default")));
-    req.extend(field_ld(14, &encode_model_meta(model, false)));
+    req.extend(field_ld(14, &encode_model_meta(&resolved_model, fast)));
     req.extend(field_str(16, &conv));
     let frame0 = connect_frame(&field_ld(1, &req));
 
@@ -716,6 +779,90 @@ mod tests {
         assert!(hay.contains("composer-2.5"));
     }
 
+    /// Live-verified against `agentn.global.api5.cursor.sh`: each of these
+    /// composite catalog ids previously failed with `ERROR_BAD_MODEL_NAME`
+    /// and the resolved base id on the right succeeded (2026-09-11).
+    #[test]
+    fn resolve_model_id_strips_vendor_prefix_and_fast_suffix() {
+        assert_eq!(
+            resolve_model_id("cursor-grok-4.6-high-fast"),
+            ("grok-4.6".to_string(), true)
+        );
+        assert_eq!(
+            resolve_model_id("cursor-grok-4.5-high"),
+            ("grok-4.5".to_string(), false)
+        );
+    }
+
+    #[test]
+    fn resolve_model_id_strips_thinking_and_effort_suffixes() {
+        assert_eq!(
+            resolve_model_id("claude-opus-5-thinking-high"),
+            ("claude-opus-5".to_string(), false)
+        );
+        assert_eq!(
+            resolve_model_id("claude-opus-5-thinking-high-fast"),
+            ("claude-opus-5".to_string(), true)
+        );
+        assert_eq!(
+            resolve_model_id("claude-sonnet-5-thinking-xhigh"),
+            ("claude-sonnet-5".to_string(), false)
+        );
+        assert_eq!(
+            resolve_model_id("claude-fable-5-thinking-high"),
+            ("claude-fable-5".to_string(), false)
+        );
+    }
+
+    #[test]
+    fn resolve_model_id_strips_plain_effort_suffix() {
+        assert_eq!(
+            resolve_model_id("gpt-5.4-high"),
+            ("gpt-5.4".to_string(), false)
+        );
+        assert_eq!(
+            resolve_model_id("gpt-5.4-medium"),
+            ("gpt-5.4".to_string(), false)
+        );
+        assert_eq!(
+            resolve_model_id("gpt-5.6-sol-high-fast"),
+            ("gpt-5.6-sol".to_string(), true)
+        );
+        assert_eq!(
+            resolve_model_id("gpt-5.6-sol-xhigh"),
+            ("gpt-5.6-sol".to_string(), false)
+        );
+        assert_eq!(
+            resolve_model_id("gemini-3.7-flash-high"),
+            ("gemini-3.7-flash".to_string(), false)
+        );
+    }
+
+    #[test]
+    fn resolve_model_id_leaves_bare_ids_unchanged() {
+        // These are already bare base ids and must pass through untouched.
+        assert_eq!(
+            resolve_model_id("composer-2.5"),
+            ("composer-2.5".to_string(), false)
+        );
+        assert_eq!(
+            resolve_model_id("gemini-3.1-pro"),
+            ("gemini-3.1-pro".to_string(), false)
+        );
+        assert_eq!(
+            resolve_model_id("sonnet-4.6"),
+            ("sonnet-4.6".to_string(), false)
+        );
+    }
+
+    #[test]
+    fn resolve_model_id_strips_fast_from_bare_composer() {
+        assert_eq!(
+            resolve_model_id("composer-2-fast"),
+            ("composer-2".to_string(), true)
+        );
+    }
+
     #[test]
     fn extract_answer_text_reads_nested_chunk() {
         // f1 { f1 { f1: "AUTH" } }

[dajiaohuang]

I prepared a focused branch for this fix: https://github.com/dajiaohuang/jcode/tree/fix/1226-cursor-model-id . It normalizes catalog vendor/effort/thinking suffixes to the bare Cursor model id and sends ast separately; cargo test -p jcode-provider-cursor-runtime passes (18 tests) and git diff --check is clean. I could not open an upstream PR because my account has read-only permission; if this direction fits, please open the PR from the branch or grant the required contribution access.

[1jehuang]

Confirmed that both Cursor request descriptors currently send the composite model ID with fast=false. The proposed patch addresses the rejected ID, but removes effort/thinking suffixes without transmitting those modes separately. I am leaving integration for focused protocol review so a successful response does not silently change the requested mode. Recommendation: retain base ID, fast, effort, and thinking semantics in both descriptors, with decoded-frame tests. The original contributor patch is welcome for review with credit preserved; no maintainer-authored rewrite is required.

---

— Jcode agent (automated triage), on behalf of @1jehuang

[yansigit]

I prepared a reviewable fork branch based on the latest upstream master:

• Branch: https://github.com/yansigit/jcode/tree/feat/provider-pool-account-management-upstream
• Compare: https://github.com/1jehuang/jcode/compare/master...yansigit:feat/provider-pool-account-management-upstream?expand=1

The branch includes the Cursor AgentService model-id normalization and model discovery fallback described here, including support for catalog entries such as grok-4.6 and kimi-k3. It also carries the dependent Cursor account-pool and request-scoping work required by the runtime integration.

Validation: jcode-provider-cursor-runtime tests passed 16/16, public Cursor model-list/account/usage checks passed, and real Cursor inference succeeded for grok-4.6 and kimi-k3.

Because the branch contains dependent account-pool work, it is offered as an integration reference first. I can provide a narrower stacked branch if the maintainer wants the model transport change reviewed separately.

[dajiaohuang]

Understood. The branch currently preserves fast but not effort or thinking; I will leave it as a reference pending descriptor semantics review.

[yansigit]

A focused Cursor native AgentService branch is available for protocol review:

https://github.com/yansigit/jcode/tree/feat/cursor-native-tools-mcp

It covers native HTTP/2 framing, current dated CLI metadata, Connect gzip/KV/context handling, streamed text and reasoning extraction, MCP native tool round-trips, sanitized collision-safe tool names, bounded transport behavior, and stream cleanup. The branch will be rebased and split if needed before PR submission. It should be reviewed together with the model descriptor requirements raised here, especially preserving base model, fast, effort, and thinking semantics rather than silently dropping modes.

[yansigit]

The native Cursor transport/tool bridge is submitted as draft PR #1261. This remains complementary to this issue, whose model descriptor semantics still need independent review.

[1jehuang]

Thanks for separating the work. PR #1261 is now open and ready for review as the native transport/tool bridge, but it does not by itself settle this issue's base-model, fast, effort, and thinking contract. Those modes still need explicit representation and decoded-frame coverage rather than silent suffix removal. PR #1270 also overlaps the account/model integration work, so the reviews should coordinate without conflating the issues.

---

Jcode agent (automated triage), on behalf of @1jehuang