fix(llm): drop deepseek v4 pro from all harness fallback chains

MODELS.md

@@ -0,0 +1,188 @@
+# Models
+
+Default and fallback model assignments for every harness in this repo.
+
+## Source of truth
+
+[models.json](models.json) maps a stable alias to a concrete model ID. Templates
+(`*.tpl.*`) reference the alias as an `__UPPER_SNAKE__` placeholder;
+[scripts/llm-update.sh](scripts/llm-update.sh) sed-substitutes them into the
+generated configs.
+
+```bash
+make llm-update
+```
+
+Never hand-edit a generated config. Edit the `.tpl.*` file, regenerate, and commit both.
+
+### Aliases
+
+| Placeholder | Model ID |
+| --- | --- |
+| `__CLAUDE_OPUS__` | `claude-opus-5` |
+| `__CLAUDE_SONNET__` | `claude-sonnet-5` |
+| `__CLAUDE_HAIKU__` | `claude-haiku-4-5-20251001` |
+| `__GPT__` | `gpt-5.6-sol` |
+| `__GPT_LUNA__` | `gpt-5.6-luna` |
+| `__GPT_CODEX__` | `gpt-5.3-codex` |
+| `__GPT_IMAGE__` | `gpt-image-2` |
+| `__GEMINI_PRO__` | `gemini-3.1-pro-preview` |
+| `__GEMINI_FLASH__` | `gemini-3.6-flash` |
+| `__DEEPSEEK_FLASH__` | `deepseek-v4-flash` |
+| `__DEEPSEEK_PRO__` | `deepseek-v4-pro` |
+| `__GLM__` | `glm-4.7` |
+| `__GEMMA__` | `gemma-4-31b-it` |
+| `__GEMMA_LOCAL__` | `gemma3:4b` |
+| `__MINIMAX__` | `minimax-m3` |
+| `__KIMI__` | `kimi-k3` |
+| `__GROK__` | `grok-4.5` |
+| `__QWEN__` | `qwen3.6-plus` |
+| `__QWEN_LOCAL__` | `qwen3.5-0.8b-optiq` |
+
+This table is the complete set of keys in [models.json](models.json). Every key
+also gets two derived forms: `__<KEY>_PRETTY__` for the display name
+("Deepseek V4 Flash") and `__<KEY>_NONDOT__` for the dot-stripped ID, used in
+OpenRouter `@preset/` names.
+
+Two provider-specific overrides are declared in
+[scripts/llm-update.sh](scripts/llm-update.sh) rather than in `models.json`,
+because the upstream ID differs from the canonical one:
+
+| Placeholder | Value | Used by |
+| --- | --- | --- |
+| `__GPT_IMAGE_OPENROUTER__` | `openai/gpt-5.4-image-2` | OpenRouter, aliased back to `gpt-image-2` |
+| `__DEEPSEEK_FLASH_0731__` | `deepseek-v4-flash-0731` | Aliyun, aliased back to `deepseek-v4-flash` |
+
+## The shared fallback chain
+
+Harnesses that support runtime fallback use one chain, in this order:
+
+```
+deepseek-v4-flash  (primary)
+  -> gemma-4-31b-it
+  -> glm-4.7
+  -> free           (OpenRouter free router, last resort)
+```
+
+Rules:
+
+- `deepseek-v4-flash` is the only DeepSeek model on any automatic path.
+  `deepseek-v4-pro` stays in the CLIProxy catalog and is addressable by explicit
+  request, but it is never a fallback hop and never a default.
+- Every hop resolves through CLIProxy, so provider-level rotation (OpenCode ->
+  Aliyun -> OpenRouter) already happens inside a single hop. Do not add a hop
+  that repeats the primary model.
+- Keep the chain model-diverse. Each hop should be a different vendor family so
+  a vendor-wide outage cannot exhaust the chain.
+
+## Per-harness assignments
+
+### Fallback-capable
+
+| Harness | Default | Fallback chain | Config |
+| --- | --- | --- | --- |
+| OpenCode | `shunkakinoki/deepseek-v4-flash` | shared chain, `shunkakinoki/` prefix | [opencode-fallback.tpl.jsonc](config/opencode/opencode-fallback.tpl.jsonc) |
+| OpenClaw | `cliproxy/deepseek-v4-flash` | shared chain, `cliproxy/` prefix | [openclaw.tpl.json](config/openclaw/openclaw.tpl.json) |
+| Hermes | `cliproxy/deepseek-v4-flash` | shared chain via `fallback_providers` | [config.tpl.yaml](config/hermes/config.tpl.yaml) |
+
+OpenCode fallback is driven by the `opencode-runtime-fallback@0.2.3` plugin:
+
+| Setting | Value |
+| --- | --- |
+| `retry_on_errors` | `401, 404, 429, 500, 502, 503, 504` |
+| `retryable_error_patterns` | `unknown provider for model` |
+| `max_fallback_attempts` | `5` |
+| `cooldown_seconds` | `60` |
+| `timeout_seconds` | `30` |
+
+OpenClaw and Hermes have no equivalent error-pattern matcher, so they hard-fail
+on the `unknown provider for model <prefixed-name>` 400 that OpenCode absorbs.
+
+Hermes also runs a Mixture-of-Agents preset: reference models
+`deepseek-v4-flash` + `minimax-m3`, aggregator `deepseek-v4-flash`.
+
+### No fallback chain
+
+| Harness | Role | Model | Config |
+| --- | --- | --- | --- |
+| OpenCode | `small_model` | `shunkakinoki/glm-4.7` | [opencode.tpl.jsonc](config/opencode/opencode.tpl.jsonc) |
+| OpenCode | `code-reviewer` agent | `shunkakinoki/deepseek-v4-flash` | [opencode.tpl.jsonc](config/opencode/opencode.tpl.jsonc) |
+| OMP | `default` | `cliproxyapi/deepseek-v4-flash` | [config.tpl.yml](config/omp/config.tpl.yml) |
+| OMP | `smol` | `openai-codex/gpt-5.6-luna` | |
+| OMP | `slow`, `vision`, `plan` | `openai-codex/gpt-5.6-sol` | |
+| OMP | `commit` | `openai/gpt-5.6-luna` | |
+| OMP | `task` | `openai-codex/gpt-5.6-luna` | |
+| Codex | default | `gpt-5.6-sol` | [config.tpl.toml](config/codex/config.tpl.toml) |
+| Codex | subagents | `gpt-5.6-luna` | |
+| Codex | `qwen-local` profile | `qwen3.5-0.8b-optiq` (LM Studio) | |
+| Pi | `defaultModel` | `glm-4.7` (provider `cliproxyapi`) | [settings.tpl.json](config/pi/settings.tpl.json) |
+| Factory (droid) | session default | `deepseek-v4-flash` | [settings.tpl.json](config/factory/settings.tpl.json) |
+| Factory (droid) | custom local | `gemma3:4b` (Ollama) | |
+| aichat | default | `cliproxy:glm-4.7` | [config.tpl.yaml](config/aichat/config.tpl.yaml) |
+| llm | default | `glm-4.7` | [default_model.tpl.txt](config/llm/default_model.tpl.txt) |
+| Handy | transcript post-process | `@preset/glm-4.7` (OpenRouter) | [settings_store.tpl.json](config/handy/settings_store.tpl.json) |
+
+OMP subagent overrides: `code-explorer`, `comment-analyzer`, and
+`pr-test-analyzer` use `gpt-5.6-luna`; the rest (`code-architect`,
+`code-reviewer`, `code-simplifier`, `silent-failure-hunter`,
+`type-design-analyzer`) use `gpt-5.6-sol`.
+
+### CCS profiles
+
+Claude Code Switch swaps the whole Anthropic model triple per profile.
+
+| Profile | Opus + Sonnet slot | Haiku slot |
+| --- | --- | --- |
+| `agy` | `claude-opus-5` | `claude-sonnet-5` |
+| `codex` | `gpt-5.3-codex` | `gpt-5.3-codex` |
+| `gemini` | `gemini-3.1-pro-preview` | `gemini-3.6-flash` |
+| `glm` | `glm-4.7` | `glm-4.7` |
+
+### Fish shortcuts
+
+The `l` suffix means local (LM Studio), `h` means headless.
+
+| Function | Model |
+| --- | --- |
+| `ocxe`, `ocxeh` | `cliproxyapi/deepseek-v4-flash` |
+| `ocxel`, `ocxelh` | `lmstudio/qwen3.5-0.8b-optiq` |
+| `coxe`, `coxeh` | `gpt-5.6-sol` |
+| `coxel`, `coxelh` | `qwen3.5-0.8b-optiq` (`--oss --local-provider lmstudio`) |
+| `pixe`, `pixeh` | `cliproxyapi/glm-4.7` |
+| `pixel`, `pixelh` | `lmstudio/qwen3.5-0.8b-optiq` |
+
+## CLIProxy routing
+
+Three provider prefixes appear above. All three are CLIProxy. Each harness names
+its own, so the prefix alone does not tell you remote vs local. `cliproxyapi/`
+in particular means different endpoints in different harnesses.
+
+| Prefix | Harness | Endpoint |
+| --- | --- | --- |
+| `shunkakinoki/` | OpenCode | `https://cliproxy.shunkakinoki.com/v1` (remote) |
+| `cliproxy/` | OpenClaw, Hermes | `https://cliproxy.shunkakinoki.com/v1` (remote) |
+| `cliproxyapi/` | OpenCode | `http://localhost:8317/v1` (local) |
+| `cliproxyapi/` | OMP | `http://127.0.0.1:8317/v1` (local) |
+| `cliproxyapi/` | Pi | `https://cliproxy.shunkakinoki.com/v1` (remote) |
+
+Every one of them resolves through
+[config.tpl.yaml](config/cliproxyapi/config.tpl.yaml). Higher `priority` wins.
+
+| Provider | Priority | DeepSeek models served |
+| --- | --- | --- |
+| `opencode` | 300 | `deepseek-v4-pro`, `deepseek-v4-flash` |
+| `aliyun` | 200 | `deepseek-v4-pro`, `deepseek-v4-flash-0731` aliased to `deepseek-v4-flash` |
+| `openrouter` | 100 | `@preset/deepseek-v4-pro`, `@preset/deepseek-v4-flash` |
+
+So a single `deepseek-v4-flash` request tries OpenCode Zen, then Aliyun, then
+OpenRouter before the harness-level fallback chain sees a failure.
+
+## Changing a model
+
+1. To swap a model version everywhere: edit [models.json](models.json), run
+   `make llm-update`, commit the template and generated files together.
+2. To change a default or a fallback hop for one harness: edit that harness's
+   `.tpl.*` file, run `make llm-update`.
+3. Update the matching assertions in `spec/llm_update_spec.sh`,
+   `spec/openclaw_hydrate_spec.sh`, `spec/hermes_hydrate_spec.sh`, and
+   `spec/cliproxyapi_spec.sh`, then run `shellspec`.

README.md

@@ -14,6 +14,8 @@ To pin a named host (skips hostname auto-detection):
 
 For troubleshooting and frequently asked questions, see [FAQ.md](./FAQ.md).
 
+For default and fallback model assignments per harness, see [MODELS.md](./MODELS.md).
+
 ## Credits
 
 See [REFERENCES.md](./REFERENCES.md) for more information.

config/factory/settings.json

@@ -1,6 +1,6 @@
 {
   "sessionDefaultSettings": {
-    "model": "deepseek-v4-pro",
+    "model": "deepseek-v4-flash",
     "reasoningEffort": "high"
   },
   "customModels": [

config/factory/settings.tpl.json

@@ -1,6 +1,6 @@
 {
   "sessionDefaultSettings": {
-    "model": "__DEEPSEEK_PRO__",
+    "model": "__DEEPSEEK_FLASH__",
     "reasoningEffort": "high"
   },
   "customModels": [

config/hermes/config.template.yaml

@@ -10,14 +10,10 @@ providers:
     api_mode: chat_completions
     supports_prompt_cache_key: true
 fallback_providers:
-  - provider: cliproxy
-    model: deepseek-v4-pro
   - provider: cliproxy
     model: gemma-4-31b-it
   - provider: cliproxy
     model: glm-4.7
-  - provider: cliproxy
-    model: minimax-m3
   - provider: cliproxy
     model: free
 moa:

config/hermes/config.tpl.yaml

@@ -10,14 +10,10 @@ providers:
     api_mode: chat_completions
     supports_prompt_cache_key: true
 fallback_providers:
-  - provider: cliproxy
-    model: __DEEPSEEK_PRO__
   - provider: cliproxy
     model: __GEMMA__
   - provider: cliproxy
     model: __GLM__
-  - provider: cliproxy
-    model: __MINIMAX__
   - provider: cliproxy
     model: free
 moa:

config/omp/config.tpl.yml

@@ -76,7 +76,7 @@ modelRoles:
   # Commit message model.
   commit: "openai/__GPT_LUNA__"
   # Fast subagent path.
-  task: "openai-codex/__GPT_CODEX_SPARK__"
+  task: "openai-codex/__GPT_LUNA__"
 # =============================================================================
 # DISPLAY
 # =============================================================================
@@ -473,7 +473,7 @@ task:
   disabledAgents: []
   agentModelOverrides:
     code-architect: "__GPT__"
-    code-explorer: "__GPT_CODEX_SPARK__"
+    code-explorer: "__GPT_LUNA__"
     code-reviewer: "__GPT__"
     code-simplifier: "__GPT__"
     comment-analyzer: "__GPT_LUNA__"

config/omp/config.yml

@@ -76,7 +76,7 @@ modelRoles:
   # Commit message model.
   commit: "openai/gpt-5.6-luna"
   # Fast subagent path.
-  task: "openai-codex/gpt-5.3-codex-spark"
+  task: "openai-codex/gpt-5.6-luna"
 # =============================================================================
 # DISPLAY
 # =============================================================================
@@ -473,7 +473,7 @@ task:
   disabledAgents: []
   agentModelOverrides:
     code-architect: "gpt-5.6-sol"
-    code-explorer: "gpt-5.3-codex-spark"
+    code-explorer: "gpt-5.6-luna"
     code-reviewer: "gpt-5.6-sol"
     code-simplifier: "gpt-5.6-sol"
     comment-analyzer: "gpt-5.6-luna"

config/openclaw/openclaw.template.json

@@ -244,10 +244,8 @@
       "model": {
         "primary": "cliproxy/deepseek-v4-flash",
         "fallbacks": [
-          "cliproxy/deepseek-v4-pro",
           "cliproxy/gemma-4-31b-it",
           "cliproxy/glm-4.7",
-          "cliproxy/minimax-m3",
           "cliproxy/free"
         ]
       },

config/openclaw/openclaw.tpl.json

@@ -244,10 +244,8 @@
       "model": {
         "primary": "cliproxy/__DEEPSEEK_FLASH__",
         "fallbacks": [
-          "cliproxy/__DEEPSEEK_PRO__",
-          "cliproxy/gemma-4-31b-it",
-          "cliproxy/glm-4.7",
-          "cliproxy/__MINIMAX__",
+          "cliproxy/__GEMMA__",
+          "cliproxy/__GLM__",
           "cliproxy/free"
         ]
       },

config/opencode/opencode-fallback.jsonc

@@ -7,10 +7,8 @@
   "timeout_seconds": 30,
   "notify_on_fallback": true,
   "fallback_models": [
-    "shunkakinoki/deepseek-v4-pro",
     "shunkakinoki/gemma-4-31b-it",
     "shunkakinoki/glm-4.7",
-    "shunkakinoki/minimax-m3",
     "shunkakinoki/free"
   ]
 }

config/opencode/opencode-fallback.tpl.jsonc

@@ -7,10 +7,8 @@
   "timeout_seconds": 30,
   "notify_on_fallback": true,
   "fallback_models": [
-    "shunkakinoki/__DEEPSEEK_PRO__",
     "shunkakinoki/__GEMMA__",
     "shunkakinoki/__GLM__",
-    "shunkakinoki/__MINIMAX__",
     "shunkakinoki/free"
   ]
 }

models.json

@@ -4,7 +4,6 @@
   "claude-haiku": "claude-haiku-4-5-20251001",
   "gpt": "gpt-5.6-sol",
   "gpt-codex": "gpt-5.3-codex",
-  "gpt-codex-spark": "gpt-5.3-codex-spark",
   "gpt-image": "gpt-image-2",
   "gpt-luna": "gpt-5.6-luna",
   "gemini-pro": "gemini-3.1-pro-preview",

spec/hermes_hydrate_spec.sh

@@ -170,10 +170,8 @@ It 'uses the OpenClaw primary and CLIProxy-only fallback chain'
 When run bash -c "sed -n '1,22p' '$PWD/config/hermes/config.template.yaml'"
 The output should include 'default: cliproxy/deepseek-v4-flash'
 The output should include 'provider: cliproxy'
-The output should include 'model: deepseek-v4-pro'
 The output should include 'model: gemma-4-31b-it'
 The output should include 'model: glm-4.7'
-The output should include 'model: minimax-m3'
 The output should include 'model: free'
 End
 

spec/llm_update_spec.sh

@@ -124,7 +124,7 @@ The status should be success
 End
 
 It 'generates a separate OpenCode fallback chain'
-When run bash -c "jq -e '.retry_on_errors == [401,404,429,500,502,503,504] and .retryable_error_patterns == [\"unknown provider for model\"] and .max_fallback_attempts == 5 and .fallback_models == [\"shunkakinoki/deepseek-v4-pro\",\"shunkakinoki/gemma-4-31b-it\",\"shunkakinoki/glm-4.7\",\"shunkakinoki/minimax-m3\",\"shunkakinoki/free\"]' config/opencode/opencode-fallback.jsonc >/dev/null"
+When run bash -c "jq -e '.retry_on_errors == [401,404,429,500,502,503,504] and .retryable_error_patterns == [\"unknown provider for model\"] and .max_fallback_attempts == 5 and .fallback_models == [\"shunkakinoki/gemma-4-31b-it\",\"shunkakinoki/glm-4.7\",\"shunkakinoki/free\"]' config/opencode/opencode-fallback.jsonc >/dev/null"
 The status should be success
 End
 

spec/openclaw_hydrate_spec.sh

@@ -9,10 +9,8 @@ template_uses_cliproxy_flash_default() {
     .agents.defaults.model == {
       "primary": "cliproxy/deepseek-v4-flash",
       "fallbacks": [
-        "cliproxy/deepseek-v4-pro",
         "cliproxy/gemma-4-31b-it",
         "cliproxy/glm-4.7",
-        "cliproxy/minimax-m3",
         "cliproxy/free"
       ]
     } and