command-code 1.80.0 → 1.82.0

package/CHANGELOG.md

@@ -1,5 +1,19 @@
 # command-code
 
+## 1.82.0
+
+### Minor Changes
+
+- feat: gemini argon
+- fix: No longer re-send the same prompt automatically on model refusal
+- fix: commands run with `!` bash mode and their output reach the agent.
+
+## 1.81.0
+
+### Minor Changes
+
+- feat: add `codemode` tool that lets the model write scripts to orchestrate tool calls, MCP only by default, change via /config
+
 ## 1.80.0
 
 ### Minor Changes

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

@@ -96,6 +96,7 @@ The model catalog — every id `/model`, `--model`, `model:effort` shorthand, an
 
 | Id (use EXACTLY this) | Name | Context | Efforts | $/1M in/out · cache read | Min plan | Best for |
 |---|---|---|---|---|---|---|
+| `google/gemini-4-argon` | Gemini 4 Argon | 1M | low, medium, high | $2/$10 · cache $0.1 | Pro and above | multimodal reasoning for documents and agent workflows |
 | `google/gemini-3.8-flash` | Gemini 3.8 Flash | 1M | low, medium, high | $1.5/$7.5 · cache $0.15 | GOAT and above | newest Gemini Flash, improved core reasoning |
 | `google/gemini-3.7-flash` | Gemini 3.7 Flash | 1.05M | low, medium, high | $1.5/$7.5 · cache $0.15 (write $0.08334) | GOAT and above | higher-quality coding & agentic workflows, fewer tokens |
 | `google/gemini-3.6-flash` | Gemini 3.6 Flash | 1M | low, medium, high | $1.5/$7.5 · cache $0.15 | Pro and above | previous Gemini Flash, still fast & capable |

package/dist/bundled/command-code-knowledge/reference/product-help.md

@@ -321,6 +321,7 @@ Valid model ids (the /model catalog — use these EXACT ids):
 - stepfun/Step-3.5-Flash
 - tencent/hy4-preview
 - tencent/hy3-paid
+- google/gemini-4-argon
 - google/gemini-3.8-flash
 - google/gemini-3.7-flash
 - google/gemini-3.6-flash

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

@@ -439,6 +439,58 @@ MCP tools are hidden entirely in plan mode - Command Code can't verify a remote
 
 ---
 
+## Codemode
+
+The `codemode` tool lets the model write a short JavaScript script that calls other tools. Instead of one tool call per round trip, the model writes a script that does several steps and runs it in one go: it loops, fans out with `Promise.all`, and filters, and only what it prints or returns reaches the context. A chain of MCP calls with a filter at the end becomes one call:
+
+```js
+const issues = await tools.mcp__linear__list_issues({team: "Core"});
+return issues.filter(issue => issue.open).map(issue => issue.id);
+```
+
+| Parameter | Type | Required | Description |
+| --- | --- | --- | --- |
+| `code` | string | yes | The body of an async function. Top-level `await` and `return` work. |
+| `timeout_ms` | number | no | Deadline for the whole script. Default 120,000, between 1,000 and 600,000. |
+
+### Configure codemode
+
+Codemode has three modes, set in `/config`. The mode decides when the tool is offered to the model and which tools a script may call.
+
+- **MCP tools.** The default. The tool is offered whenever an MCP server is connected, and scripts can call the MCP tools. Everything else works as before: MCP tools can still be loaded and called directly.
+- **Read-only and MCP tools.** The tool is offered in every session. Scripts can call the MCP tools and the read-only built-in tools: reading files and directories, glob, grep, web search, and the todo and task tools. These tools are made script-first and are removed from context, so that the model uses them through codemode.
+- **Off.** Disables codemode completely.
+
+In every mode, a script can never edit or write files, run commands, ask the user a question, review a plan, switch plan mode, load tools, or start another script.
+
+### Globals
+
+| Global | What it does |
+| --- | --- |
+| `tools.<name>(args)` | Calls a tool. One object argument, shaped like the tool's parameters. |
+| `ALL_TOOLS` | The tools this script may call, with a one-line description each. |
+| `searchTools("query")` | Finds callable tools by description. |
+| `describeTool("name")` | Returns a tool's full schema. |
+| `CWD` | The working directory. |
+| `text(...)`, `console.log(...)` | Print a line. |
+| `return`, `exit()` | Send back a value; end the script early. |
+| `store("key", value)`, `load("key")` | Keep a JSON value across scripts in the session. |
+
+### Permissions
+
+Every nested call is a real tool call: same mode filter, permission rules, input repair, and output truncation as a direct call, with prompts one at a time. `codemode` itself is read-only and never prompts, since it can do no more than the calls it makes. It is hidden in any mode that leaves a script nothing to call, such as plan mode under the default setting.
+
+### Limits
+
+Scripts run in a sandbox with no file system, network, timers, or process access; the tool bridge is the only way out.
+
+- 128 MB of memory, one sandbox per call
+- 8 nested calls in flight at once
+- 1 MB of printed output, after which the script fails; 25,000 tokens reach the model
+- 256 KB per stored value, 1 MB per session store
+
+---
+
 ## Permission modes & tool visibility
 
 Which tools the model even *sees* depends on the mode:

package/dist/bundled/config/references/settings.md

@@ -15,6 +15,7 @@ The executable registry is authoritative. Run `cmd config list --scope effective
 | `peek-recap-diagram` | user | boolean | `true`, `false` | When a task that changed files or ran commands finishes, draw one short text diagram recapping it (one extra model call per task, same as /peek). |
 | `peek-recap-delay` | user | enum | `0`, `1`, `2`, `5`, `10`, `15`, `30`, `60` | Recap a finished task only when this long passed since your last message, so a quick back-and-forth is not recapped every time (same as /peek delay). |
 | `image-vision` | user | enum | `ask`, `enabled`, `disabled` | Let a text-only model read attached images by describing them with the vision model ("Ask on first use" prompts the first time you attach an image). |
+| `codemode` | user | enum | `mcp`, `tools`, `off` | The codemode tool lets the model write a short code script that can call several tools and run it in one go: fewer round trips, less raw tool output in context. Scripts never edit, write, or run commands. |
 | `branch-summary-skip-prompt` | user | boolean | `true`, `false` | Never ask about summarizing when switching branches in /tree. |
 | `default-export-format` | user | enum | `html`, `jsonl`, `md` | Format a bare /export (no format or path) writes — an explicit /export html\|jsonl\|md\|<path> always overrides. |
 | `default-share-gist-format` | user | enum | `html`, `jsonl`, `md` | Format a bare /share gist (no format argument) posts — an explicit /share gist html\|jsonl\|md always overrides. |