command-code 1.37.0 → 1.38.0

package/CHANGELOG.md

@@ -1,5 +1,18 @@
 # command-code
 
+## 1.38.0
+
+### Patch Changes
+
+- feat: add reasoning effort support for Tencent Hy4 Preview
+- feat: add reasoning effort support for Tencent Hy4 Preview
+
+## Unreleased
+
+### Patch Changes
+
+- feat: Allow custom agents to set `reasoningEffort`
+
 ## 1.37.0
 
 ### Minor Changes

package/dist/bundled/command-code-knowledge/reference/custom-agents.md

@@ -238,6 +238,7 @@ Only these keys are read; anything else is ignored. All except `name` are option
     Tools the agent may use. `"*"` grants all; otherwise a comma- or space-separated list (or a YAML array). Omit it and the agent gets no tools.
     Deny list applied after `tools`. Deny beats allow, so pair it with `tools: "*"` to mean "everything except these". Same format as `tools`.
     Model this agent always runs on: any `/model` id. See [Available models](https://commandcode.ai/docs/reference/cli/models) for the full list, or run `cmd --list-models`. Omit it (or set `model: inherit`) to follow the session model.
+    Reasoning effort this agent runs at. Reasoning levels are model-specific. Pin/select a level that agent's model supports; an unsupported level (or a model with no effort levels) falls back to the model's default reasoning effort. Omit to use the model default.
     Caps the agent's loop. Default `100`.
     Overrides the session's permission mode for this agent. One of `default`, `auto-accept`, `bypass`, `plan`, `dont-ask`. A session already in `plan` or `bypass` wins. Defaults to inheriting the session.
     When `true`, every run detaches: the `agent` tool returns an `agent_id` and results come from `agent_output`.
@@ -275,6 +276,18 @@ Use ids exactly as `--list-models` prints them. Examples: `claude-opus-4-8`, `cl
 
 Omitting `model` is the same as `model: inherit` - the agent follows the session's `/model`. Pinning a model gives the agent its own prompt cache, so a slow planner and a fast implementer don't thrash each other's.
 
+### Reasoning effort
+
+`reasoningEffort` pins how hard the agent thinks, independent of the session's `/effort`. Pin/select a reasoning level the agent's model supports. To see a model's exact set, open the `/model` picker and select the model - its reasoning-effort step lists precisely the levels that model accepts.
+
+The value is validated in two stages so a bad one never reaches the provider mid-delegation:
+
+- At load time an unknown level (a typo like `meduim`) is dropped with a warning, and the agent falls back to the model default - the file still loads.
+- At runtime, the level is checked against the model actually chosen (an `inherit` model isn't known until launch). 
+- A level the resolved model doesn't support falls back to the model default rather than being clamped to a different level.
+
+Omit it to run at the model's default effort.
+
 ### Full example
 
 Every field, with valid values. Copy it and delete what you don't need.
@@ -287,6 +300,7 @@ description: "Writes a structured research report from the current repo. Delegat
 tools: "*"                                  # grant everything…
 disallowedTools: shell_command, write_file  # …except shells and file writes
 model: claude-opus-4-8                       # any /model id, or omit to inherit
+reasoningEffort: high                        # a level THIS model supports (varies by model); omit for default
 maxTurns: 40                                 # cap the loop (default 100)
 permissionMode: plan                         # default|auto-accept|bypass|plan|dont-ask
 background: true                             # detach; collect via agent_output
@@ -310,6 +324,7 @@ Quick lookup:
 | `tools` | string \| list | `"*"`, or comma/space-separated ids, or a YAML list | none |
 | `disallowedTools` | string \| list | same format as `tools`; applied after it | - |
 | `model` | string | any `/model` id, or `inherit` | inherit |
+| `reasoningEffort` | string | reasoning level the pinned model supports | model default |
 | `maxTurns` | integer | positive integer | `100` |
 | `permissionMode` | string | `default`, `auto-accept`, `bypass`, `plan`, `dont-ask` | inherit |
 | `background` | boolean | `true` / `false` | `false` |

package/dist/bundled/command-code-knowledge/reference/models.md

@@ -40,7 +40,7 @@ The model catalog — every id `/model`, `--model`, `model:effort` shorthand, an
 | `stepfun/Step-3.7-Flash` | Step 3.7 Flash | 256K | — | $0.2/$1.15 · cache $0.04 | Go and above | multimodal sparse-MoE reasoning |
 | `stepfun/Step-3.5-Flash` | Step 3.5 Flash | 1M | — | $0.1/$0.3 · cache $0.02 | Go and above | fast sparse-MoE agentic reasoning |
 | `tencent/hy3-paid` | Tencent Hy3 | 262K | — | $0.14/$0.58 · cache $0.035 | Go and above | sparse-MoE reasoning & agentic tool use |
-| `tencent/hy4-preview` | Tencent Hy4 Preview | 1.05M | — | $0.834/$2.501 · cache $0.042 | Go and above | agentic coding & sustained multi-step tool use |
+| `tencent/hy4-preview` | Tencent Hy4 Preview | 1.05M | low, medium, high | $0.834/$2.501 · cache $0.042 | Go and above | agentic coding & sustained multi-step tool use |
 | `nvidia/nemotron-3-ultra-550b-a55b` | Nemotron 3 Ultra | 1M | — | $0.6/$2.4 · cache $0.12 | Go and above | open reasoning model for long-horizon autonomous agents |
 | `thinkingmachines/inkling` | Inkling | 256K | — | $1/$4.05 · cache $0.17 | Go and above | multimodal MoE reasoning |
 | `thinkingmachines/inkling-small` | Inkling Small | 1M | — | $0.5/$1.2 · cache $0.1 | Go and above | lightweight MoE reasoning at lower cost and latency |