diff --git a/packages/ai/CHANGELOG.md b/packages/ai/CHANGELOG.md
index 5927e59e15..7e1b0bdba0 100644
--- a/packages/ai/CHANGELOG.md
+++ b/packages/ai/CHANGELOG.md
@@ -10,6 +10,8 @@
 
 ### Fixed
 
+- Cursor suffix families the 2026-08-18 alias snapshot does not list (`grok-4.7`, `claude-opus-5-5`, `claude-fable-5-1` and its thinking variants, `gemini-3.8-flash`, `muse-spark-1.3`) group into one selectable identity with reasoning levels, derived from the live `GetUsableModels` batch, instead of arriving as flat singletons with reasoning off. The identity offers exactly the levels the server listed (an unlisted level is unsupported and clamps to a listed one), and an explicit level resolves to the exact server-listed variant id instead of silently downgrading to the representative. A derived family never takes an id that is already a static identity, a static alias key, or a raw catalog id; those members stay flat. Stored flat variants regroup on restore, and a stored group, static or derived, keeps its levels and representative whichever side of its flat aliases it was stored on. The context window is unchanged: a family without capability data or an observed server limit still uses the 200k fallback. ([#2038](https://github.com/code-yeongyu/senpi/issues/2038))
+
 ### Removed
 
 ## [2026.9.23-4] - 2026-09-23
@@ -25,7 +27,6 @@
 ### Fixed
 
 - The `pi-messages` provider projects tool-result text fields onto the wire instead of forwarding display audience metadata. ([#2041](https://github.com/code-yeongyu/senpi/issues/2041))
-- Cursor suffix families the 2026-08-18 alias snapshot does not list (`grok-4.7`, `claude-opus-5-5`, `claude-fable-5-1` and its thinking variants, `gemini-3.8-flash`, `muse-spark-1.3`) group into one selectable identity with reasoning levels, derived from the live `GetUsableModels` batch, instead of arriving as flat singletons with reasoning off. The identity offers exactly the levels the server listed (an unlisted level is unsupported and clamps to a listed one), and an explicit level resolves to the exact server-listed variant id instead of silently downgrading to the representative. A derived family never takes an id that is already a static identity, a static alias key, or a raw catalog id; those members stay flat. Stored flat variants regroup on restore, and a stored group, static or derived, keeps its levels and representative whichever side of its flat aliases it was stored on. The context window is unchanged: a family without capability data or an observed server limit still uses the 200k fallback. ([#2038](https://github.com/code-yeongyu/senpi/issues/2038))
 
 ### Removed
 
diff --git a/packages/coding-agent/CHANGELOG.md b/packages/coding-agent/CHANGELOG.md
index c39c4a34ee..b56bf75056 100644
--- a/packages/coding-agent/CHANGELOG.md
+++ b/packages/coding-agent/CHANGELOG.md
@@ -11,6 +11,7 @@
 ### Fixed
 
 - Truncated `bash` output no longer ends with `[Showing lines … ; earlier output dropped]` in the TUI. The marker still reaches the model unchanged as a model-only text part, and `bash_output` treats its `[N earlier chars dropped]` notice the same way. ([#2063](https://github.com/code-yeongyu/senpi/issues/2063))
+- Cursor legacy variant references for families the static alias table does not list (grok-4.7, claude-opus-5-5, claude-fable-5-1, gemini-3.8-flash, muse-spark-1.3) now resolve onto the runtime-derived identity with their thinking level instead of fuzzy-matching a flat `-fast` model: `--model cursor/grok-4.7:low` selects grok-4.7 at low (previously grok-4.7-xhigh-fast with thinking off), `cursor/grok-4.7-xhigh` (in any letter case) and stored variant ids restore their level, `cursor/grok-4.7-*` globs project the derived identity alongside the fast models, and both Cursor lanes keep the derived level-to-variant-id map across CLI catalog cache reloads. Exact raw models still win over derived aliases. The CLI catalog cache rebuilds its models from the stored `cursor-agent models` listing, so a recorded context limit no longer forces a re-probe, and a cache written before this change probes once; if that probe fails, the cached models are still served instead of the built-in offline list. ([#2038](https://github.com/code-yeongyu/senpi/issues/2038))
 - A tool call whose name senpi auto-corrects (a gateway-namespaced or recased name such as `mcp__<id>__Edit`) now looks like a direct call to the resolved tool: its card uses that tool's renderer live and on resume, and the `[auto-corrected]` notice reaches only the model. ([#2064](https://github.com/code-yeongyu/senpi/issues/2064))
 
 ### Removed
@@ -56,7 +57,6 @@
 - `--no-model-fallback`, `SENPI_NO_FALLBACK=1`, `--no-ask-user` and `--theme` now hold for the whole session. They were silently dropped the first time the session saved any setting, so a run that opted out of model fallback could still answer from a different model. Changing the same setting yourself during the session (for example with `/fallback`) still takes effect. ([#2052](https://github.com/code-yeongyu/senpi/issues/2052))
 - When a Claude subscription model rejects the local Claude Code as too old, the message now names the binary that ran and where it came from (`CLAUDE_CODE_EXECUTABLE`, the copy senpi ships, or `claude` on your PATH) and gives the fix that works for that binary, instead of always telling you to update senpi. On Windows, a Claude Code installed with `npm i -g @anthropic-ai/claude-code` is now found and used when it is newer than the bundled copy. ([#2053](https://github.com/code-yeongyu/senpi/issues/2053))
 - `/mcp auth <server>` opens the system browser and keeps the complete authorization URL in the transcript instead of replacing it with a progress notice. Browser launch failures leave manual authorization usable, and the MCP documentation now names the supported `auth`, `auth-start`, and `auth-complete` commands. ([oh-my-openagent#6724](https://github.com/code-yeongyu/oh-my-openagent/issues/6724))
-- Cursor legacy variant references for families the static alias table does not list (grok-4.7, claude-opus-5-5, claude-fable-5-1, gemini-3.8-flash, muse-spark-1.3) now resolve onto the runtime-derived identity with their thinking level instead of fuzzy-matching a flat `-fast` model: `--model cursor/grok-4.7:low` selects grok-4.7 at low (previously grok-4.7-xhigh-fast with thinking off), `cursor/grok-4.7-xhigh` (in any letter case) and stored variant ids restore their level, `cursor/grok-4.7-*` globs project the derived identity alongside the fast models, and both Cursor lanes keep the derived level-to-variant-id map across CLI catalog cache reloads. Exact raw models still win over derived aliases. The CLI catalog cache rebuilds its models from the stored `cursor-agent models` listing, so a recorded context limit no longer forces a re-probe, and a cache written before this change probes once; if that probe fails, the cached models are still served instead of the built-in offline list. ([#2038](https://github.com/code-yeongyu/senpi/issues/2038))
 
 ### Removed