justrach · cursor · Aug 29, 2026 · Aug 29, 2026 · Aug 29, 2026 · Aug 29, 2026
diff --git a/README.md b/README.md
@@ -23,6 +23,10 @@
   <img alt="Built in Zig 0.17 dev" src="https://img.shields.io/badge/built%20in-Zig%200.17%20dev-f7a41d?logo=zig&logoColor=white">
 </p>
 
+<p align="center">
+  <a href="https://trendshift.io/repositories/84216?utm_source=repository-badge&utm_medium=badge&utm_campaign=badge-repository-84216" target="_blank" rel="noopener noreferrer"><img src="https://trendshift.io/api/badge/repositories/84216" alt="justrach/codegraff | Trendshift" width="250" height="55"></a>
+</p>
+
 <p align="center">
   <strong>The most token-efficient coding harness we've built so far.</strong><br/>
   Prompt-cache max keeps repeated prefixes hot. RLM + spec-ptc turns wide work

diff --git a/architecture.md b/architecture.md
@@ -121,8 +121,9 @@ as codex).
 ### Streaming
 
 Root-agent requests stream on **all three** wire formats. `request()` decides
-per call: `live = !sub && out != null && !stream_quiet` (compaction sets
-`stream_quiet`; subagents always take the buffered `post()` path). When live,
+per call: `live = !sub` (`Agent.usesLiveTransport`). `-p` and compaction set
+`out=null` / `stream_quiet` to mute paint, not to leave the stall-watched
+path; subagents always take the buffered `postWatched()` path. When live,
 `buildBody` adds `stream:true` (openai also gets
 `stream_options:{include_usage:true}` so token counts survive — dropped and
 retried once if a provider rejects it, same pattern as the soft-strict

diff --git a/docs/adr/0045-glm-flash-swe-codegraff.md b/docs/adr/0045-glm-flash-swe-codegraff.md
@@ -45,6 +45,6 @@ completed the same model. Do not steal Pi's heap (167M vs 23–90M).
 ## Consequences
 
 - Fair Codegraff A/B is `--model glm-5.3-flash` with `CODEGRAFF_API_KEY`.
-- Next graff cut on this model is finishing inside 300s (fewer / shorter
-  turns), not a catalog shrink to four tools.
+- The next cut is ADR 0046: omit default `reasoning_effort` on flash
+  and compact lean tool prose. Not a four-tool catalog.
 - Rotate any `cg_sk_` pasted into a chat; do not commit it.
diff --git a/docs/adr/0046-flash-omits-default-effort.md b/docs/adr/0046-flash-omits-default-effort.md
@@ -0,0 +1,107 @@
+# 0046. Flash models omit default `reasoning_effort`
+
+Status: accepted 2026-08-29
+
+## Context
+
+ADR 0045 ran Codegraff `glm-5.3-flash` SWE after the `-p` learn skip.
+Pi 5/6 in 758s (`first` 0.6s). Graff 3/6 in 1353s (`first` 6.9s) with
+three 300s timeouts that never edited. Output on the tasks that
+finished was 3.5–7.5k tokens. `validated` was 10 calls / 164s vs Pi
+5 / 26s.
+
+Pi's `first` is the JSONL `session` line, not first model work — do
+not treat 0.6s as a TUI steal. The real gap is that Pi's
+`models.json` sets `reasoning: false` / `supportsReasoningEffort:
+false`. Graff's codegraff spec has `takes_effort = true`, so an
+uncatalogued flash model inherited **default medium**
+`reasoning_effort` on every request.
+
+That is the same footgun Gemini already had: default medium maps to
+seconds of thinking (2× wall vs Pi). The Gemini skip was
+`startsWith("gemini")` only, so `glm-5.3-flash` still thought.
+
+The lean keep-list is already 8 tools (ADR 0024: an rlm-only catalog
+failed). Bash / read / edit / codedb / attempt_completion still shipped
+interactive-length essays on every `-p` turn.
+
+## Decision
+
+- Flash / Gemini default `.medium` is sent as **`low`**, not omitted.
+  Omitting the field still bills `reasoning_tokens` (27 on a `pong`;
+  SWE dumps 1.9MB `reasoning_content`). `low` bills 0. `none` is a
+  400. `/effort high` still sends high.
+- Catalog `glm-5.3-flash` on codegraff (chat completions, 202752 ctx).
+  Do not pass `glm-5.3` (that row is zai).
+- Compact lean one-shot **descriptions** only (same names and JSON
+  schemas). Do not drop to four tools. Do not steal Pi's heap.
+- Root `-p` uses the streaming transport (`usesLiveTransport` = `!sub`).
+  `out=null` / `stream_quiet` mute paint only. The 5-minute
+  `postWatched` deadline is for subagents, not one-shots.
+
+## Consequences
+
+- Flash SWE wall is generation without thinking novels, not Zig vs JS.
+- `/effort` remains the only way effort moves (ADR / `effort_route`:
+  no auto-flip).
+- `-p` writes `calling <model>` to stderr at start and drops the
+  call-2 stdout pulse. Eval `first_out` is no longer "time to model
+  call 2"; stdout stays the answer.
+- Revisit if a named flash model needs default thinking — pin it in
+  the catalog, do not restore a global medium default.
+
+## Confirm (2026-08-29, after this revision)
+
+`run-20260829-041235.jsonl`, Codegraff `glm-5.3-flash`, `--suite swe
+-j 6`. No `reasoning_effort` on the wire. Local boot is 7ms; a
+`pong` one-shot is 6.1s / 10 out (gateway TTFT). Pi `first` 0.6s is
+still the JSONL `session` line.
+
+| | pass | wall | first | out |
+|---|---:|---:|---:|---:|
+| ADR 0045 | 3/6 | 1353s | 6.9s | 15k |
+| after (thinking off) | 3/6 | 1247s | 6.7s* | 21k |
+
+\*first was the call-2 stdout pulse. `validated` 164s / 46k / 10 →
+**57s / 11k / 4**. `json-stream` and `cookie-store` still SIGKILL at
+300s after that pulse: the second model call took `postWatched`
+(~250s of silence, then `retry` / 5xx) because `-p` gated `live` on
+`out != null && !stream_quiet`. Root one-shots now `postLive` so the
+30s head-stall fires. `label-sort` finished (174s) but still fails
+the hidden check. Do not steal Pi's heap.
+
+## Confirm (2026-08-29, after streaming `-p`)
+
+`run-20260829-043927.jsonl`, Codegraff `glm-5.3-flash`, `--suite swe
+-j 6`. Root one-shots `postLive`. Eval `first` is the stderr `calling`
+line (~0.03s), not gateway TTFT.
+
+| | pass | wall | notes |
+|---|---:|---:|---|
+| ADR 0045 | 3/6 | 1353s | thinking on; 3× 300s `postWatched` |
+| thinking off | 3/6 | 1247s | still `postWatched` |
+| **this cut** | **4/6** | **1311s** | 26 calls / 157k in / 29k out |
+
+`json-stream` **passed** in 280s (was SIGKILL at 300s after a silent
+second `postWatched`). `validated` 106s / 6 calls. `map-conflict`
+116s. `config-parse` 208s.
+
+`cookie-store` still SIGKILL at 300s: call 2 traced `first_token` at
+11.5s then never completed — a live stream, not the 5-minute watched
+POST. `label-sort` was mid-turn (call 2 was 219s / 1.9MB SSE) when
+the cap hit. Do not steal Pi's heap. Do not shrink the keep-list.
+
+## Confirm (2026-08-29, `reasoning_effort=low`)
+
+`run-20260829-045547.jsonl`. Gateway A/B: omit = 27 reasoning_tokens
+on `pong`; `low` = 0; `thinking.disabled` still thinks; `none` is 400.
+
+| | pass | wall (sum) | out | calls |
+|---|---:|---:|---:|---:|
+| Pi (ADR 0045) | 5/6 | 758s | 27k | 35 |
+| graff omit | 4/6 | 1311s | 29k | 26 |
+| **graff `low`** | **5/6** | **286s** | **4.7k** | 38 |
+
+Same miss as Pi (`label-sort` hidden check). `cookie-store` 115s
+(was 300s SIGKILL; Pi 301s). `json-stream` 36s (was 280s; Pi 117s).
+Do not steal Pi's heap. Do not shrink the keep-list.
diff --git a/docs/adr/README.md b/docs/adr/README.md
@@ -56,6 +56,7 @@ record only when you need the evidence or the edge cases.
 | [0043](0043-pi-swe-same-seat.md) | Pi SWE A/B uses `pi-xai` on the SuperGrok seat; do not steal Pi's catalog or heap from the json-stream pass. |
 | [0044](0044-oneshot-skips-learn-auto.md) | `-p` and `--json` skip learn auto-init; the Pi SWE wall gap was a 38s `graff-pinned` copy, not their catalog. |
 | [0045](0045-glm-flash-swe-codegraff.md) | Codegraff `glm-5.3-flash` SWE: Pi 5/6 in 758s, graff 3/6 with three 300s timeouts; do not steal Pi's heap. |
+| [0046](0046-flash-omits-default-effort.md) | Flash / Gemini send `reasoning_effort=low` (omit still thinks); lean `-p` shortens tool prose; `-p` streams. |
 
 ## When to write one
 

diff --git a/examples/edge-worker/src/remote.ts b/examples/edge-worker/src/remote.ts
@@ -9,7 +9,7 @@
 
 export const HARNESS_VERSION = "0.12";
 
-export type ModelName = "MiniMax-M2.5" | "MiniMax-M2.7" | "MiniMax-M3" | "accounts/fireworks/models/deepseek-v4-flash" | "accounts/fireworks/models/deepseek-v4-pro" | "accounts/fireworks/models/glm-5p2" | "accounts/fireworks/models/gpt-oss-120b" | "accounts/fireworks/models/kimi-k2p6" | "accounts/fireworks/models/kimi-k2p7-code" | "accounts/fireworks/models/minimax-m3" | "accounts/fireworks/models/qwen3p7-plus" | "alibaba/qwen3.8-27b" | "anthropic/claude-sonnet-4.6" | "claude-fable-5" | "claude-haiku-4-5" | "claude-opus-4-5" | "claude-opus-4-6" | "claude-opus-4-7" | "claude-opus-4-8" | "claude-opus-4.8" | "claude-opus-5" | "claude-sonnet-4-5" | "claude-sonnet-4-6" | "claude-sonnet-4.6" | "claude-sonnet-5" | "deepseek-chat" | "deepseek-reasoner" | "deepseek-v4-flash" | "deepseek-v4-pro" | "fugu" | "fugu-ultra" | "fugu-ultra-20260615" | "gemma-4-31b" | "glm-4.5" | "glm-4.7" | "glm-5" | "glm-5-turbo" | "glm-5.2" | "glm-5.3" | "glm-5v-turbo" | "gpt-5-codex" | "gpt-5.2" | "gpt-5.3-codex-spark" | "gpt-5.4" | "gpt-5.4-mini" | "gpt-5.4-pro" | "gpt-5.5" | "gpt-5.6" | "gpt-5.6-luna" | "gpt-5.6-sol" | "gpt-5.6-terra" | "gpt-oss-120b" | "grok-4.3" | "grok-4.6" | "grok-build" | "k3" | "kilo-auto/small" | "kimi-for-coding" | "kimi-for-coding-highspeed" | "kimi-k2.6" | "kimi-latest" | "lmstudio" | "mimo-v2-flash" | "mimo-v2.5" | "mimo-v2.5-pro" | "mimo-v2.5-pro-ultraspeed" | "minimax-m3" | "mistral-medium-latest" | "mlx-community/Qwen3.6-27B-OptiQ-4bit" | "openai/gpt-oss-120b" | (string & {});
+export type ModelName = "MiniMax-M2.5" | "MiniMax-M2.7" | "MiniMax-M3" | "accounts/fireworks/models/deepseek-v4-flash" | "accounts/fireworks/models/deepseek-v4-pro" | "accounts/fireworks/models/glm-5p2" | "accounts/fireworks/models/gpt-oss-120b" | "accounts/fireworks/models/kimi-k2p6" | "accounts/fireworks/models/kimi-k2p7-code" | "accounts/fireworks/models/minimax-m3" | "accounts/fireworks/models/qwen3p7-plus" | "alibaba/qwen3.8-27b" | "anthropic/claude-sonnet-4.6" | "claude-fable-5" | "claude-haiku-4-5" | "claude-opus-4-5" | "claude-opus-4-6" | "claude-opus-4-7" | "claude-opus-4-8" | "claude-opus-4.8" | "claude-opus-5" | "claude-sonnet-4-5" | "claude-sonnet-4-6" | "claude-sonnet-4.6" | "claude-sonnet-5" | "deepseek-chat" | "deepseek-reasoner" | "deepseek-v4-flash" | "deepseek-v4-pro" | "fugu" | "fugu-ultra" | "fugu-ultra-20260615" | "gemma-4-31b" | "glm-4.5" | "glm-4.7" | "glm-5" | "glm-5-turbo" | "glm-5.2" | "glm-5.3" | "glm-5.3-flash" | "glm-5v-turbo" | "gpt-5-codex" | "gpt-5.2" | "gpt-5.3-codex-spark" | "gpt-5.4" | "gpt-5.4-mini" | "gpt-5.4-pro" | "gpt-5.5" | "gpt-5.6" | "gpt-5.6-luna" | "gpt-5.6-sol" | "gpt-5.6-terra" | "gpt-oss-120b" | "grok-4.3" | "grok-4.6" | "grok-build" | "k3" | "kilo-auto/small" | "kimi-for-coding" | "kimi-for-coding-highspeed" | "kimi-k2.6" | "kimi-latest" | "lmstudio" | "mimo-v2-flash" | "mimo-v2.5" | "mimo-v2.5-pro" | "mimo-v2.5-pro-ultraspeed" | "minimax-m3" | "mistral-medium-latest" | "mlx-community/Qwen3.6-27B-OptiQ-4bit" | "openai/gpt-oss-120b" | (string & {});
 export type ToolName = "bash" | "bash_output" | "bash_kill" | "read_file" | "edit_file" | "write_file" | "webfetch" | "skill" | "codedb" | "read_tool_result" | "todo_write" | "todo_read" | "eval" | "note_constraint" | "ask_user" | "attempt_completion" | "load_tool_schemas" | "mcp_search_tools" | "mcp_select_tool" | "clock_sleep" | "subagent" | "workflow" | "agent_output" | "learn_candidate" | "peer_message" | "workspace" | "imagegen";
 export type ProviderId = "anthropic" | "codegraff" | "deepseek" | "openai" | "minimax" | "xiaomi" | "kilo" | "groq" | "cerebras" | "mistral" | "kimi" | "moonshot" | "xai" | "zai" | "vercel" | "openrouter" | "fugu" | "fireworks" | "mlx" | "lmstudio" | "codex";
 
@@ -521,5 +521,5 @@ export async function* runAgentRemote(opts: RunAgentRemoteOptions): AsyncGenerat
   }
 }
 
-export const MODELS: ModelName[] = ["MiniMax-M2.5", "MiniMax-M2.7", "MiniMax-M3", "accounts/fireworks/models/deepseek-v4-flash", "accounts/fireworks/models/deepseek-v4-pro", "accounts/fireworks/models/glm-5p2", "accounts/fireworks/models/gpt-oss-120b", "accounts/fireworks/models/kimi-k2p6", "accounts/fireworks/models/kimi-k2p7-code", "accounts/fireworks/models/minimax-m3", "accounts/fireworks/models/qwen3p7-plus", "alibaba/qwen3.8-27b", "anthropic/claude-sonnet-4.6", "claude-fable-5", "claude-haiku-4-5", "claude-opus-4-5", "claude-opus-4-6", "claude-opus-4-7", "claude-opus-4-8", "claude-opus-4.8", "claude-opus-5", "claude-sonnet-4-5", "claude-sonnet-4-6", "claude-sonnet-4.6", "claude-sonnet-5", "deepseek-chat", "deepseek-reasoner", "deepseek-v4-flash", "deepseek-v4-pro", "fugu", "fugu-ultra", "fugu-ultra-20260615", "gemma-4-31b", "glm-4.5", "glm-4.7", "glm-5", "glm-5-turbo", "glm-5.2", "glm-5.3", "glm-5v-turbo", "gpt-5-codex", "gpt-5.2", "gpt-5.3-codex-spark", "gpt-5.4", "gpt-5.4-mini", "gpt-5.4-pro", "gpt-5.5", "gpt-5.6", "gpt-5.6-luna", "gpt-5.6-sol", "gpt-5.6-terra", "gpt-oss-120b", "grok-4.3", "grok-4.6", "grok-build", "k3", "kilo-auto/small", "kimi-for-coding", "kimi-for-coding-highspeed", "kimi-k2.6", "kimi-latest", "lmstudio", "mimo-v2-flash", "mimo-v2.5", "mimo-v2.5-pro", "mimo-v2.5-pro-ultraspeed", "minimax-m3", "mistral-medium-latest", "mlx-community/Qwen3.6-27B-OptiQ-4bit", "openai/gpt-oss-120b"];
+export const MODELS: ModelName[] = ["MiniMax-M2.5", "MiniMax-M2.7", "MiniMax-M3", "accounts/fireworks/models/deepseek-v4-flash", "accounts/fireworks/models/deepseek-v4-pro", "accounts/fireworks/models/glm-5p2", "accounts/fireworks/models/gpt-oss-120b", "accounts/fireworks/models/kimi-k2p6", "accounts/fireworks/models/kimi-k2p7-code", "accounts/fireworks/models/minimax-m3", "accounts/fireworks/models/qwen3p7-plus", "alibaba/qwen3.8-27b", "anthropic/claude-sonnet-4.6", "claude-fable-5", "claude-haiku-4-5", "claude-opus-4-5", "claude-opus-4-6", "claude-opus-4-7", "claude-opus-4-8", "claude-opus-4.8", "claude-opus-5", "claude-sonnet-4-5", "claude-sonnet-4-6", "claude-sonnet-4.6", "claude-sonnet-5", "deepseek-chat", "deepseek-reasoner", "deepseek-v4-flash", "deepseek-v4-pro", "fugu", "fugu-ultra", "fugu-ultra-20260615", "gemma-4-31b", "glm-4.5", "glm-4.7", "glm-5", "glm-5-turbo", "glm-5.2", "glm-5.3", "glm-5.3-flash", "glm-5v-turbo", "gpt-5-codex", "gpt-5.2", "gpt-5.3-codex-spark", "gpt-5.4", "gpt-5.4-mini", "gpt-5.4-pro", "gpt-5.5", "gpt-5.6", "gpt-5.6-luna", "gpt-5.6-sol", "gpt-5.6-terra", "gpt-oss-120b", "grok-4.3", "grok-4.6", "grok-build", "k3", "kilo-auto/small", "kimi-for-coding", "kimi-for-coding-highspeed", "kimi-k2.6", "kimi-latest", "lmstudio", "mimo-v2-flash", "mimo-v2.5", "mimo-v2.5-pro", "mimo-v2.5-pro-ultraspeed", "minimax-m3", "mistral-medium-latest", "mlx-community/Qwen3.6-27B-OptiQ-4bit", "openai/gpt-oss-120b"];
 export const TOOLS: ToolName[] = ["bash", "bash_output", "bash_kill", "read_file", "edit_file", "write_file", "webfetch", "skill", "codedb", "read_tool_result", "todo_write", "todo_read", "eval", "note_constraint", "ask_user", "attempt_completion", "load_tool_schemas", "mcp_search_tools", "mcp_select_tool", "clock_sleep", "subagent", "workflow", "agent_output", "learn_candidate", "peer_message", "workspace", "imagegen"];
diff --git a/graff-evals/harnesses.json b/graff-evals/harnesses.json
@@ -22,6 +22,7 @@
     "usage": "graff-stderr",
     "capabilities": ["output-schema"],
     "schema_args": ["--output-schema", "{schema}"],
+    "env": {"GRAFF_POST_DEADLINE_SECS": "90"},
     "default_model": "grok-4.6"
   },
   "graff-dev-rlm": {

diff --git a/sdk/py/harness_sdk.py b/sdk/py/harness_sdk.py
@@ -13,7 +13,7 @@
 import urllib.request
 from typing import Iterator, List, Literal, Optional, TypedDict, Union
 
-MODELS = ["MiniMax-M2.5", "MiniMax-M2.7", "MiniMax-M3", "accounts/fireworks/models/deepseek-v4-flash", "accounts/fireworks/models/deepseek-v4-pro", "accounts/fireworks/models/glm-5p2", "accounts/fireworks/models/gpt-oss-120b", "accounts/fireworks/models/kimi-k2p6", "accounts/fireworks/models/kimi-k2p7-code", "accounts/fireworks/models/minimax-m3", "accounts/fireworks/models/qwen3p7-plus", "alibaba/qwen3.8-27b", "anthropic/claude-sonnet-4.6", "claude-fable-5", "claude-haiku-4-5", "claude-opus-4-5", "claude-opus-4-6", "claude-opus-4-7", "claude-opus-4-8", "claude-opus-4.8", "claude-opus-5", "claude-sonnet-4-5", "claude-sonnet-4-6", "claude-sonnet-4.6", "claude-sonnet-5", "deepseek-chat", "deepseek-reasoner", "deepseek-v4-flash", "deepseek-v4-pro", "fugu", "fugu-ultra", "fugu-ultra-20260615", "gemma-4-31b", "glm-4.5", "glm-4.7", "glm-5", "glm-5-turbo", "glm-5.2", "glm-5.3", "glm-5v-turbo", "gpt-5-codex", "gpt-5.2", "gpt-5.3-codex-spark", "gpt-5.4", "gpt-5.4-mini", "gpt-5.4-pro", "gpt-5.5", "gpt-5.6", "gpt-5.6-luna", "gpt-5.6-sol", "gpt-5.6-terra", "gpt-oss-120b", "grok-4.3", "grok-4.6", "grok-build", "k3", "kilo-auto/small", "kimi-for-coding", "kimi-for-coding-highspeed", "kimi-k2.6", "kimi-latest", "lmstudio", "mimo-v2-flash", "mimo-v2.5", "mimo-v2.5-pro", "mimo-v2.5-pro-ultraspeed", "minimax-m3", "mistral-medium-latest", "mlx-community/Qwen3.6-27B-OptiQ-4bit", "openai/gpt-oss-120b"]
+MODELS = ["MiniMax-M2.5", "MiniMax-M2.7", "MiniMax-M3", "accounts/fireworks/models/deepseek-v4-flash", "accounts/fireworks/models/deepseek-v4-pro", "accounts/fireworks/models/glm-5p2", "accounts/fireworks/models/gpt-oss-120b", "accounts/fireworks/models/kimi-k2p6", "accounts/fireworks/models/kimi-k2p7-code", "accounts/fireworks/models/minimax-m3", "accounts/fireworks/models/qwen3p7-plus", "alibaba/qwen3.8-27b", "anthropic/claude-sonnet-4.6", "claude-fable-5", "claude-haiku-4-5", "claude-opus-4-5", "claude-opus-4-6", "claude-opus-4-7", "claude-opus-4-8", "claude-opus-4.8", "claude-opus-5", "claude-sonnet-4-5", "claude-sonnet-4-6", "claude-sonnet-4.6", "claude-sonnet-5", "deepseek-chat", "deepseek-reasoner", "deepseek-v4-flash", "deepseek-v4-pro", "fugu", "fugu-ultra", "fugu-ultra-20260615", "gemma-4-31b", "glm-4.5", "glm-4.7", "glm-5", "glm-5-turbo", "glm-5.2", "glm-5.3", "glm-5.3-flash", "glm-5v-turbo", "gpt-5-codex", "gpt-5.2", "gpt-5.3-codex-spark", "gpt-5.4", "gpt-5.4-mini", "gpt-5.4-pro", "gpt-5.5", "gpt-5.6", "gpt-5.6-luna", "gpt-5.6-sol", "gpt-5.6-terra", "gpt-oss-120b", "grok-4.3", "grok-4.6", "grok-build", "k3", "kilo-auto/small", "kimi-for-coding", "kimi-for-coding-highspeed", "kimi-k2.6", "kimi-latest", "lmstudio", "mimo-v2-flash", "mimo-v2.5", "mimo-v2.5-pro", "mimo-v2.5-pro-ultraspeed", "minimax-m3", "mistral-medium-latest", "mlx-community/Qwen3.6-27B-OptiQ-4bit", "openai/gpt-oss-120b"]
 TOOLS = ["bash", "bash_output", "bash_kill", "read_file", "edit_file", "write_file", "webfetch", "skill", "codedb", "read_tool_result", "todo_write", "todo_read", "eval", "note_constraint", "ask_user", "attempt_completion", "load_tool_schemas", "mcp_search_tools", "mcp_select_tool", "clock_sleep", "subagent", "workflow", "agent_output", "learn_candidate", "peer_message", "workspace", "imagegen"]
 PROVIDERS = ["anthropic", "codegraff", "deepseek", "openai", "minimax", "xiaomi", "kilo", "groq", "cerebras", "mistral", "kimi", "moonshot", "xai", "zai", "vercel", "openrouter", "fugu", "fireworks", "mlx", "lmstudio", "codex"]