@iinm/plain-agent 1.10.15 → 1.10.17

package/README.md

@@ -1,26 +1,24 @@
 # Plain Agent
 
-A lightweight, capable coding agent for the terminal.
+A lightweight coding agent for the terminal.
 
-- **Multi-provider** — Use Claude, GPT, Gemini, or any OpenAI-compatible model via Bedrock, Vertex AI, or direct APIs.
-- **Fine-grained auto-approval** — Auto-approve tool calls by name and arguments using regex patterns, with path validation on tool arguments.
+- **Multi-provider** — Use Claude, GPT, Gemini, or any OpenAI-compatible model via direct APIs or through Bedrock, Vertex AI, or Azure.
+- **Fine-grained auto-approval** — Auto-approve tool calls by matching tool names and inputs against configurable patterns, while validating string inputs as paths for safety.
 - **Sandboxed execution** — Run commands in a Docker container with filesystem and network isolation.
-- **Claude Code compatible** — Use Claude Code plugins, commands, subagents, and skills from `.claude/`.
-- **Zero external dependencies** — Built with Node.js standard libraries only.
+- **Supports Claude Code resources** — Use Claude Code plugins, commands, subagents, and skills from `.claude/`.
+- **Zero external dependencies** — Built using only Node.js standard libraries.
 
 ## Limitations
 
-- **Path validation only covers tool arguments** — It blocks paths outside the working directory, directory traversal (`..`), symlinks escaping the project, and git-ignored files — but only for paths explicitly passed as tool-use arguments; it cannot control file access inside arbitrary scripts. Always use sandboxed execution when allowing arbitrary script execution.
-- **Sequential subagent execution** — Subagents run one at a time rather than
-  in parallel. The trade-off is full visibility: every step is streamed to
-  your terminal so you can follow exactly what each subagent is doing.
+- **Path validation only covers tool arguments** — It blocks paths outside the working directory, directory traversal (`..`), symlinks that escape the project, and git-ignored files. However, it only applies to paths explicitly passed as tool-use arguments, so it cannot control file access inside arbitrary scripts. Always use sandboxed execution when running arbitrary scripts.
+- **Sequential subagent execution** — Subagents run one at a time rather than in parallel. The trade-off is that every step is streamed to your terminal, so you can follow exactly what each subagent is doing.
 
 ## Requirements
 
 - Node.js 22 or later
-- LLM provider credentials
+- Credentials for your LLM provider
 - [ripgrep](https://github.com/burntsushi/ripgrep), [fd](https://github.com/sharkdp/fd)
-- Bash / Docker for sandboxed execution
+- Bash and Docker for sandboxed execution
 
 ## Quick Start
 
@@ -28,13 +26,13 @@ A lightweight, capable coding agent for the terminal.
 npm install -g @iinm/plain-agent
 ```
 
-List available models.
+List the available models.
 
 ```sh
 plain list-models
 ```
 
-Create the configuration.
+Create a configuration file.
 
 ```js
 // ~/.config/plain-agent/config.local.json
@@ -63,14 +61,14 @@ Create the configuration.
     }
   ],
 
-  // (Optional) Enable web tools
+  // Optional: enable web tools
   "tools": {
     "webSearch": {
       "provider": "gemini",
       "apiKey": "<GEMINI_API_KEY>",
       "model": "gemini-3.5-flash"
 
-      // Or use Vertex AI (Requires gcloud CLI to get authentication token)
+      // Or use Vertex AI (requires the gcloud CLI for authentication)
       // "provider": "gemini-vertex-ai",
       // "baseURL": "https://aiplatform.googleapis.com/v1beta1/projects/<project_id>/locations/<location>",
       // "model": "gemini-3.5-flash"
@@ -86,7 +84,7 @@ Create the configuration.
       "apiKey": "<GEMINI_API_KEY>",
       "model": "gemini-3.5-flash"
 
-      // Or use Vertex AI (Requires gcloud CLI to get authentication token)
+      // Or use Vertex AI (requires the gcloud CLI for authentication)
 
       // Or use a custom command
       // "provider": "command",
@@ -98,33 +96,43 @@ Create the configuration.
 ```
 
 <details>
-<summary><b>Azure / Bedrock / Vertex AI provider examples</b></summary>
+<summary><b>Bedrock / Vertex AI / Azure provider examples</b></summary>
 
 ```js
 {
   "platforms": [
+    // Bedrock: Requires the AWS CLI for authentication
     {
-      // Requires Azure CLI to get access token
-      "name": "azure",
-      "variant": "openai",
-      "baseURL": "https://<resource>.openai.azure.com/openai",
-      // Optional
-      "azureConfigDir": "/home/xxx/.azure-for-agent"
-    },
-    {
-      // Requires AWS CLI to get credentials
       "name": "bedrock",
       "variant": "default",
       "baseURL": "https://bedrock-runtime.<region>.amazonaws.com",
       "awsProfile": "<AWS_PROFILE>"
     },
+
+    // Vertex AI: Requires the gcloud CLI for authentication
     {
-      // Requires gcloud CLI to get authentication token
       "name": "vertex-ai",
       "variant": "default",
       "baseURL": "https://aiplatform.googleapis.com/v1beta1/projects/<project>/locations/<location>",
-      // (Optional) Impersonate this service account to obtain an auth token
+      // Optional: impersonate this service account to obtain an auth token
       "account": "<SERVICE_ACCOUNT_EMAIL>"
+    },
+
+    // Azure: Requires the Azure CLI for authentication
+    {
+      "name": "azure",
+      "variant": "default",
+      "baseURL": "https://<resource>.services.ai.azure.com",
+      // Optional
+      "azureConfigDir": "/home/xxx/.azure-for-agent"
+    },
+    // Azure OpenAI
+    {
+      "name": "azure",
+      "variant": "openai",
+      "baseURL": "https://<resource>.openai.azure.com/openai",
+      // Optional
+      "azureConfigDir": "/home/xxx/.azure-for-agent"
     }
   ]
 }
@@ -132,7 +140,7 @@ Create the configuration.
 </details>
 
 <details>
-<summary><b>OpenAI compatible provider examples</b></summary>
+<summary><b>OpenAI-compatible provider examples</b></summary>
 
 ```js
 {
@@ -185,7 +193,7 @@ Create the configuration.
 </details>
 
 <details>
-<summary><b>Bedrock example using Claude Japan inference profiles</b></summary>
+<summary><b>Bedrock example with Claude Geo Cross-Region inference</b></summary>
 
 ```js
 {
@@ -260,22 +268,21 @@ Run the agent.
 
 ```sh
 plain
-```
 
-```
+# Or
 plain -m <model+variant>
 ```
 
-Press **Ctrl-C** to pause auto-approve. The agent will finish the current tool call, then return to the prompt.
+Press **Ctrl-C** to pause auto-approval. The agent will finish the current tool call, then return to the prompt.
 
-Display the help message.
+Show the help message.
 
 ```
 /help
 ```
 
-Run in batch mode (non-interactive).
-In batch mode, config files are not loaded automatically. Only the files specified with `-c` are loaded.
+Run in non-interactive batch mode.
+In batch mode, configuration files are not loaded automatically. Only the files specified with `-c` are loaded.
 
 ```sh
 plain batch \
@@ -284,34 +291,32 @@ plain batch \
       "Add tests for ..."
 ```
 
-Show daily token costs across sessions. `plain cost` reads
+Show daily token cost. `plain cost` reads
 `~/.local/share/plain-agent/usage.jsonl`; use `--from` / `--to` to set the
-period. Currencies are shown separately.
+period. Costs are shown separately by currency.
 
 ```sh
 plain cost
-```
 
-```
+# Or
 plain cost --from 2026-04-01 --to 2026-04-30
 ```
 
 Resume a previously interrupted interactive session. Sessions are
-auto-saved to `.plain-agent/sessions/` and can be removed with `rm` when
-no longer needed. Without an argument, the most recently updated session
+automatically saved to `.plain-agent/sessions/` and can be deleted with `rm` when
+no longer needed. If no argument is provided, the most recently updated session
 is resumed. Use `--list` to see resumable sessions. Switching models is
-not supported (`-m` is rejected).
+not supported (`-m` is not allowed).
 
 ```sh
 plain resume
-```
 
-```
+# Or
 plain resume --list
 plain resume 2026-05-10-0803-a7k
 ```
 
-Configure plain-agent for your project.
+Set up Plain Agent for your project.
 
 ```
 /configure Auto-approve file writes and patches
@@ -343,7 +348,7 @@ Files are loaded in the following order. Settings in later files override earlie
 ### Example
 
 <details>
-<summary><b>YOLO mode example (requires sandbox for safety)</b></summary>
+<summary><b>YOLO mode example (requires a sandbox for safety)</b></summary>
 
 ```js
 {
@@ -363,7 +368,7 @@ Files are loaded in the following order. Settings in later files override earlie
         "toolName": { "$regex": "^(web_search|web_fetch)$" },
         "action": "allow"
       }
-      // ⚠️ Never do this. mcp run outside the sandbox, so they can send anything externally.
+      // ⚠️ Never do this. MCP runs outside the sandbox, so it can send anything externally.
       // {
       //   "toolName": { "$regex": "." },
       //   "action": "allow"
@@ -395,10 +400,10 @@ Files are loaded in the following order. Settings in later files override earlie
 ```js
 {
   "autoApproval": {
-    // Absolute paths outside the working directory to allow access to. Relative paths are ignored.
+    // Absolute paths outside the working directory that are allowed. Relative paths are ignored.
     "allowedPaths": ["/tmp"],
     "defaultAction": "ask",
-    // The maximum number of automatic approvals.
+    // Maximum number of automatic approvals.
     "maxApprovals": 50,
     // Patterns are evaluated in order. First match wins.
     "patterns": [
@@ -425,7 +430,7 @@ Files are loaded in the following order. Settings in later files override earlie
         "action": "allow"
       },
 
-      // MCP Tool naming convention: mcp__<serverName>__<toolName>
+      // MCP tool naming convention: mcp__<serverName>__<toolName>
       {
         "toolName": { "$regex": "mcp__slack__slack_(read|search)_.+" },
         "action": "allow"
@@ -467,7 +472,7 @@ Files are loaded in the following order. Settings in later files override earlie
       "command": "npx",
       "args": ["-y", "chrome-devtools-mcp@latest", "--isolated"]
     },
-    // ⚠️ Add to config.local.json to avoid committing secrets to Git
+    // ⚠️ Add this to config.local.json to avoid committing secrets to Git
     "slack": {
       "command": "npx",
       "args": ["-y", "mcp-remote", "https://mcp.slack.com/mcp", "--header", "Authorization:Bearer <SLACK_TOKEN>"],
@@ -476,7 +481,7 @@ Files are loaded in the following order. Settings in later files override earlie
       "command": "npx",
       "args": ["-y", "mcp-remote", "https://mcp.notion.com/mcp"],
       "options": {
-        // Enable only specific tools (optional - if not specified, all tools are enabled)
+        // Enable only specific tools. If not specified, all tools are enabled.
         "enabledTools": ["notion-search", "notion-fetch"]
       }
     },
@@ -484,18 +489,18 @@ Files are loaded in the following order. Settings in later files override earlie
       "command": "npx",
       "args": ["-y", "mcp-remote", "https://knowledge-mcp.global.api.aws"]
     },
-    // ⚠️ Add to config.local.json to avoid committing secrets to Git
+    // ⚠️ Add this to config.local.json to avoid committing secrets to Git
     "google_developer-knowledge": {
       "command": "npx",
       "args": ["-y", "mcp-remote", "https://developerknowledge.googleapis.com/mcp", "--header", "X-Goog-Api-Key:<GOOGLE_API_KEY>"]
     }
   },
 
-  // Override default notification command
+  // Override the default notification command
   "notifyCmd": { "command": "plain-notify-desktop", "args": [] }
 
   // Voice input. See "Voice Input" below.
-  // ⚠️ Add to config.local.json to avoid committing secrets to Git
+  // ⚠️ Add this to config.local.json to avoid committing secrets to Git
   "voiceInput": {
     "provider": "openai",
     "apiKey": "<OPENAI_API_KEY>"
@@ -506,7 +511,7 @@ Files are loaded in the following order. Settings in later files override earlie
 
 ## Available Tools
 
-The agent can use the following tools to assist with tasks:
+The agent can use the following tools:
 
 - **read_file**: Read a file with line numbers (1-indexed). Supports `offset` and `limit` to read a specific range.
 - **write_file**: Write a file.
@@ -517,7 +522,7 @@ The agent can use the following tools to assist with tasks:
 - **web_fetch**: Fetch the contents of a single URL and answer a question based on it (requires Google API key, Vertex AI configuration, or the `command` provider with a local fetch command such as `w3m`, `curl`, or `lynx`).
 - **switch_to_subagent**: Switch to a subagent role within the same conversation, focusing on the specified goal.
 - **switch_to_main_agent**: Switch back to the main agent role and report the result. After reporting, the subagent's conversation history is removed from the context.
-- **compact_context**: Compact the conversation context by discarding prior messages and reloading task state from a memory file. Use when the context has grown large but the task is not yet complete. Can also be invoked via the `/compact` slash command.
+- **compact_context**: Compact the conversation context by discarding earlier messages and reloading task state from a memory file. Use this when the context has grown large but the task is not yet complete. You can also invoke it with the `/compact` slash command.
 
 ## Prompts
 
@@ -551,7 +556,7 @@ Prompts located in a `shortcuts/` subdirectory (e.g., `.plain-agent/prompts/shor
 
 ## Subagents
 
-Subagents are specialized agents designed for specific tasks.
+Subagents are specialized helpers for specific tasks.
 
 ### Locations
 
@@ -577,12 +582,7 @@ You are a web content reader and analyzer. Given a URL and a question, you:
 
 ## Claude Code Plugin Support
 
-Plugins are installed under `.plain-agent/claude-code-plugins/` and must be
-installed per project by running `plain install-claude-code-plugins` from
-the project root. Global installation (e.g., under `~/.plain-agent`) is not
-supported, because plugins may include skills that the agent invokes
-autonomously, and scoping them to the project keeps approval rules and
-permission management straightforward.
+Plugins are installed under `.plain-agent/claude-code-plugins/` and must be installed per project by running `plain install-claude-code-plugins` from the project root. Global installation (e.g., under `~/.plain-agent`) is not supported because plugins may include skills the agent invokes autonomously. Keeping them scoped to the project keeps approval rules and permission management straightforward.
 
 Example:
 
@@ -613,15 +613,13 @@ plain install-claude-code-plugins
 
 ## Voice Input
 
-Press **Ctrl-O** to start recording, press it again to stop. Partial
-transcripts are inserted into the prompt as you speak so you can edit
-and send them like regular text.
+Press **Ctrl-O** to start recording, then press it again to stop. Partial transcripts are inserted into the prompt as you speak, so you can edit and send them like regular text.
 
 ### Requirements
 
 - A recording command on `PATH`: `arecord`, `sox`, or `ffmpeg`.
 - An API key for the chosen provider.
-- Your host must have microphone access. The sandbox does not need to.
+- Your host must have microphone access.
 
 ### Providers
 
@@ -655,17 +653,17 @@ and send them like regular text.
 
 ### Options
 
-- `toggleKey` — Rebind the toggle. Accepts `"ctrl-<char>"` where `<char>`
+- `toggleKey` — Rebind the toggle key. Accepts `"ctrl-<char>"` where `<char>`
   is a letter (a-z) or one of `[ \ ] ^ _`. Defaults to `"ctrl-o"`.
-- `recorder` — Override recorder auto-detection, e.g. `{ "command": "sox", "args": ["-q", "-d", "-b", "16", "-c", "1", "-r", "24000", "-e", "signed-integer", "-t", "raw", "-"] }`. Must write raw 16-bit little-endian mono PCM to stdout at 24 kHz (OpenAI) or 16 kHz (Gemini).
+- `recorder` — Override automatic recorder detection, e.g. `{ "command": "sox", "args": ["-q", "-d", "-b", "16", "-c", "1", "-r", "24000", "-e", "signed-integer", "-t", "raw", "-"] }`. It must write raw 16-bit little-endian mono PCM to stdout at 24 kHz (OpenAI) or 16 kHz (Gemini).
 
 ## Development
 
 ```sh
-# Run lint, typecheck, and test
+# Run lint, typecheck, and tests
 npm run check
 
-# Fix lint errors
+# Fix lint issues
 npm run fix
 # or
 npm run fix -- --unsafe
@@ -692,7 +690,7 @@ npm publish --access public
 <summary><b>Amazon Bedrock</b></summary>
 
 ```sh
-# IAM Identity Center 
+# IAM Identity Center
 identity_center_instance_arn="<IDENTITY_CENTER_INSTANCE_ARN>" # e.g., arn:aws:sso:::instance/ssoins-xxxxxxxxxxxxxxxx"
 identity_store_id=<IDENTITY_STORE_ID>
 aws_account_id=<AWS_ACCOUNT_ID>
@@ -700,7 +698,7 @@ aws_account_id=<AWS_ACCOUNT_ID>
 # Create a permission set
 permission_set_arn=$(aws sso-admin create-permission-set \
   --instance-arn "$identity_center_instance_arn" \
-  --name "BedrockForCodingAgent" \
+  --name "BedrockCodingAgent" \
   --description "Allows only Bedrock model invocation" \
   --query "PermissionSet.PermissionSetArn" --output text)
 
@@ -775,7 +773,7 @@ aws bedrock-runtime invoke-model \
 
 ```sh
 resource_group=<RESOURCE_GROUP>
-account_name=<ACCOUNT_NAME> # resource name
+account_name=<ACCOUNT_NAME> # Resource name
 
 # Create a service principal
 service_principal=$(az ad sp create-for-rbac --name "CodingAgentServicePrincipal" --skip-assignment)
@@ -798,7 +796,7 @@ az role assignment create \
 export app_secret=$(echo "$service_principal" | jq -r .password)
 export tenant_id=$(echo "$service_principal" | jq -r .tenant)
 
-export AZURE_CONFIG_DIR=$HOME/.azure-for-agent # Change the location to store credentials
+export AZURE_CONFIG_DIR=$HOME/.azure-for-agent # Change this to store credentials elsewhere
 az login --service-principal -u "$app_id" -p "$app_secret" --tenant "$tenant_id"
 ```
 </details>

package/config/config.predefined.json

@@ -916,7 +916,7 @@
     },
 
     {
-      "name": "gpt-5.2-chat",
+      "name": "gpt-5.4-mini",
       "variant": "thinking-medium-azure",
       "platform": {
         "name": "azure",
@@ -925,8 +925,149 @@
       "model": {
         "format": "openai-responses",
         "config": {
-          "model": "gpt-5.2-latest",
-          "reasoning": { "effort": "medium", "summary": "auto" }
+          "model": "gpt-5.4-mini",
+          "reasoning": { "effort": "medium", "summary": "auto" },
+          "store": false,
+          "include": ["reasoning.encrypted_content"]
+        }
+      },
+      "cost": {
+        "currency": "USD",
+        "unit": "1M",
+        "costs": {
+          "input_tokens": 0.75,
+          "input_tokens_details.cached_tokens": -0.675,
+          "output_tokens": 4.5
+        }
+      }
+    },
+    {
+      "name": "gpt-5.4-mini",
+      "variant": "thinking-high-azure",
+      "platform": {
+        "name": "azure",
+        "variant": "openai"
+      },
+      "model": {
+        "format": "openai-responses",
+        "config": {
+          "model": "gpt-5.4-mini",
+          "reasoning": { "effort": "high", "summary": "auto" },
+          "store": false,
+          "include": ["reasoning.encrypted_content"]
+        }
+      },
+      "cost": {
+        "currency": "USD",
+        "unit": "1M",
+        "costs": {
+          "input_tokens": 0.75,
+          "input_tokens_details.cached_tokens": -0.675,
+          "output_tokens": 4.5
+        }
+      }
+    },
+    {
+      "name": "gpt-5.4-mini",
+      "variant": "thinking-xhigh-azure",
+      "platform": {
+        "name": "azure",
+        "variant": "openai"
+      },
+      "model": {
+        "format": "openai-responses",
+        "config": {
+          "model": "gpt-5.4-mini",
+          "reasoning": { "effort": "xhigh", "summary": "auto" },
+          "store": false,
+          "include": ["reasoning.encrypted_content"]
+        }
+      },
+      "cost": {
+        "currency": "USD",
+        "unit": "1M",
+        "costs": {
+          "input_tokens": 0.75,
+          "input_tokens_details.cached_tokens": -0.675,
+          "output_tokens": 4.5
+        }
+      }
+    },
+    {
+      "name": "gpt-5.5",
+      "variant": "thinking-medium-azure",
+      "platform": {
+        "name": "azure",
+        "variant": "openai"
+      },
+      "model": {
+        "format": "openai-responses",
+        "config": {
+          "model": "gpt-5.5",
+          "reasoning": { "effort": "medium", "summary": "auto" },
+          "store": false,
+          "include": ["reasoning.encrypted_content"]
+        }
+      },
+      "cost": {
+        "currency": "USD",
+        "unit": "1M",
+        "costs": {
+          "input_tokens": 5,
+          "input_tokens_details.cached_tokens": -4.5,
+          "output_tokens": 30
+        }
+      }
+    },
+    {
+      "name": "gpt-5.5",
+      "variant": "thinking-high-azure",
+      "platform": {
+        "name": "azure",
+        "variant": "openai"
+      },
+      "model": {
+        "format": "openai-responses",
+        "config": {
+          "model": "gpt-5.5",
+          "reasoning": { "effort": "high", "summary": "auto" },
+          "store": false,
+          "include": ["reasoning.encrypted_content"]
+        }
+      },
+      "cost": {
+        "currency": "USD",
+        "unit": "1M",
+        "costs": {
+          "input_tokens": 5,
+          "input_tokens_details.cached_tokens": -4.5,
+          "output_tokens": 30
+        }
+      }
+    },
+    {
+      "name": "gpt-5.5",
+      "variant": "thinking-xhigh-azure",
+      "platform": {
+        "name": "azure",
+        "variant": "openai"
+      },
+      "model": {
+        "format": "openai-responses",
+        "config": {
+          "model": "gpt-5.5",
+          "reasoning": { "effort": "xhigh", "summary": "auto" },
+          "store": false,
+          "include": ["reasoning.encrypted_content"]
+        }
+      },
+      "cost": {
+        "currency": "USD",
+        "unit": "1M",
+        "costs": {
+          "input_tokens": 5,
+          "input_tokens_details.cached_tokens": -4.5,
+          "output_tokens": 30
         }
       }
     },
@@ -978,6 +1119,30 @@
         }
       }
     },
+    {
+      "name": "glm-5",
+      "variant": "bedrock",
+      "platform": {
+        "name": "bedrock",
+        "variant": "default"
+      },
+      "model": {
+        "format": "openai-messages",
+        "config": {
+          "model": "zai.glm-5",
+          "reasoning_effort": "high"
+        }
+      },
+      "cost": {
+        "currency": "USD",
+        "unit": "1M",
+        "costs": {
+          "prompt_tokens": 1,
+          "completion_tokens": 3.2
+        }
+      }
+    },
+
     {
       "name": "glm-5.1",
       "variant": "fireworks",
@@ -1071,6 +1236,21 @@
         }
       }
     },
+    {
+      "name": "kimi-k2.6",
+      "variant": "azure",
+      "platform": {
+        "name": "azure",
+        "variant": "default"
+      },
+      "model": {
+        "format": "openai-messages",
+        "config": {
+          "model": "Kimi-K2.6",
+          "reasoning_effort": "high"
+        }
+      }
+    },
 
     {
       "name": "deepseek-v4-pro",
@@ -1118,6 +1298,21 @@
         }
       }
     },
+    {
+      "name": "deepseek-v4-pro",
+      "variant": "azure",
+      "platform": {
+        "name": "azure",
+        "variant": "default"
+      },
+      "model": {
+        "format": "openai-messages",
+        "config": {
+          "model": "DeepSeek-V4-Pro",
+          "reasoning_effort": "high"
+        }
+      }
+    },
 
     {
       "name": "minimax-m2.7",

package/package.json

@@ -1,13 +1,16 @@
 {
   "name": "@iinm/plain-agent",
-  "version": "1.10.15",
-  "description": "A lightweight CLI-based coding agent",
+  "version": "1.10.17",
+  "description": "A lightweight coding agent for the terminal.",
   "license": "MIT",
   "type": "module",
   "repository": {
     "type": "git",
     "url": "git+https://github.com/iinm/plain-agent.git"
   },
+  "keywords": [
+    "coding-agent"
+  ],
   "bin": {
     "plain": "bin/plain",
     "plain-notify-desktop": "bin/plain-notify-desktop",

package/src/cli/formatter.mjs

@@ -844,7 +844,7 @@ function renderPatchBlock(block, originalLines, nonce) {
     out.push(
       styleText(
         "cyan",
-        `@@@ ${nonce} ${block.start}:${block.startHash}-${block.end}:${block.endHash}`,
+        `REPLACE ${nonce} ${block.start}:${block.startHash}-${block.end}:${block.endHash}`,
       ),
     );
     if (originalLines) {
@@ -872,7 +872,9 @@ function renderPatchBlock(block, originalLines, nonce) {
     }
   } else {
     const afterSuffix = block.afterHash ? `:${block.afterHash}` : "";
-    out.push(styleText("cyan", `@@@ ${nonce} ${block.after}${afterSuffix}+`));
+    out.push(
+      styleText("cyan", `INSERT_AFTER ${nonce} ${block.after}${afterSuffix}`),
+    );
     for (const line of block.body) {
       out.push(styleText("green", `+ ${line}`));
     }
@@ -890,8 +892,8 @@ function highlightPatchPlain(patch) {
   if (!patch) {
     return "";
   }
-  // Patch headers look like "@@@ <nonce> ...".
-  const headerRegex = /^@@@\s+\S+(\s.*)?$/;
+  // Patch headers look like "REPLACE <nonce> ..." or "INSERT_AFTER <nonce> ...".
+  const headerRegex = /^(REPLACE|INSERT_AFTER)\s+\S+(\s.*)?$/;
   return patch
     .split("\n")
     .map((line) => {
@@ -912,8 +914,8 @@ function highlightPatchPlain(patch) {
  * @returns {string | null}
  */
 function extractPatchNonce(patch) {
-  const match = patch.match(/^@@@\s+(\S+)/m);
-  return match ? match[1] : null;
+  const match = patch.match(/^(REPLACE|INSERT_AFTER)\s+(\S+)/m);
+  return match ? match[2] : null;
 }
 
 /**

package/src/providers/openaiCompatible.mjs

@@ -8,6 +8,7 @@ import { styleText } from "node:util";
 import { noThrow } from "../utils/noThrow.mjs";
 import { retryOnError } from "../utils/retryOnError.mjs";
 import { loadAwsCredentials, signAwsRequest } from "./platform/awsSigV4.mjs";
+import { getAzureAccessToken } from "./platform/azure.mjs";
 import { readBedrockStreamEvents } from "./platform/bedrock.mjs";
 import { getGoogleCloudAccessToken } from "./platform/googleCloud.mjs";
 
@@ -43,6 +44,8 @@ export async function callOpenAICompatibleModel(
           return `${baseURL}/model/${modelConfig.model}/invoke-with-response-stream`;
         case "vertex-ai":
           return `${baseURL}/endpoints/openapi/chat/completions`;
+        case "azure":
+          return `${baseURL}/models/chat/completions?api-version=2024-05-01-preview`;
         default:
           throw new Error(`Unsupported platform: ${platformConfig.name}`);
       }
@@ -64,6 +67,15 @@ export async function callOpenAICompatibleModel(
             ...platformConfig.customHeaders,
             Authorization: `Bearer ${await getGoogleCloudAccessToken()}`,
           };
+        case "azure":
+          return {
+            ...platformConfig.customHeaders,
+            Authorization: `Bearer ${await getAzureAccessToken(
+              platformConfig.azureConfigDir
+                ? { azureConfigDir: platformConfig.azureConfigDir }
+                : undefined,
+            )}`,
+          };
       }
     })();
 
@@ -83,7 +95,11 @@ export async function callOpenAICompatibleModel(
         case "vertex-ai":
           return {
             ...modelConfig,
-            model: modelConfig.model,
+            stream: true,
+          };
+        case "azure":
+          return {
+            ...modelConfig,
             stream: true,
           };
       }

package/src/tools/patchFile.mjs

@@ -29,18 +29,18 @@ When editing multiple locations in the same file, include all blocks in a single
           patch: {
             description: `
 Format — a single patch string may contain multiple blocks:
-@@@ ${nonce} {start}:{startHash}-{end}:{endHash}
+REPLACE ${nonce} {start}:{startHash}-{end}:{endHash}
 replacement for lines {start}-{end}
-@@@ ${nonce} {N}:{hash}
+REPLACE ${nonce} {N}:{hash}
 replace just that one line
-@@@ ${nonce} {start}:{startHash}-{end}:{endHash}
+REPLACE ${nonce} {start}:{startHash}-{end}:{endHash}
 (empty body deletes the range)
-@@@ ${nonce} {N}:{afterHash}+
-appended content after line N
-@@@ ${nonce} 0+
-prepended content at beginning of file
+INSERT_AFTER ${nonce} {N}:{afterHash}
+new content after line N
+INSERT_AFTER ${nonce} 0
+content at beginning of file
 
-- Each block's content starts right after its @@@ header line and ends at the next @@@ or the end of the string. Any blank lines between the header and the content become part of the replacement.
+- Each block's content starts right after its header line and ends at the next header or the end of the string. Any blank lines between the header and the content become part of the replacement.
 - The nonce "${nonce}" is constant; always use the exact value shown above.
 - Hashes are 2-character hex hashes of each line's full content as shown by read_file.
             `.trim(),
@@ -61,7 +61,7 @@ prepended content at beginning of file
         const blocks = parseBlocks(patch, nonce);
         if (blocks.length === 0) {
           throw new Error(
-            `No patch blocks found. Each block must start with "@@@ ${nonce} ...".`,
+            `No patch blocks found. Each block must start with "REPLACE ${nonce} ..." or "INSERT_AFTER ${nonce} ...".`,
           );
         }
 
@@ -91,7 +91,8 @@ prepended content at beginning of file
  * @returns {PatchBlock[]}
  */
 export function parseBlocks(patch, nonce) {
-  const openPrefix = `@@@ ${nonce} `;
+  const replacePrefix = `REPLACE ${nonce} `;
+  const insertPrefix = `INSERT_AFTER ${nonce} `;
   const lines = patch.split("\n");
   // Drop trailing empty element produced by split() when patch ends with \n.
   if (lines.length > 0 && lines[lines.length - 1] === "") {
@@ -101,14 +102,17 @@ export function parseBlocks(patch, nonce) {
   /** @type {number[]} */
   const headerIndices = [];
   for (let i = 0; i < lines.length; i++) {
-    if (lines[i].startsWith(openPrefix)) {
+    if (
+      lines[i].startsWith(replacePrefix) ||
+      lines[i].startsWith(insertPrefix)
+    ) {
       headerIndices.push(i);
     }
   }
 
   if (headerIndices.length === 0) {
     throw new Error(
-      `No patch blocks found. Each block must start with "@@@ ${nonce} ...".`,
+      `No patch blocks found. Each block must start with "REPLACE ${nonce} ..." or "INSERT_AFTER ${nonce} ...".`,
     );
   }
 
@@ -117,8 +121,19 @@ export function parseBlocks(patch, nonce) {
   for (let i = 0; i < headerIndices.length; i++) {
     const headerLineIdx = headerIndices[i];
     const headerLine = lines[headerLineIdx];
-    const headerArgs = headerLine.slice(openPrefix.length);
-    const header = parseHeaderArgs(headerArgs);
+
+    /** @type {"replace" | "insert"} */
+    let op;
+    let headerArgs;
+    if (headerLine.startsWith(replacePrefix)) {
+      op = "replace";
+      headerArgs = headerLine.slice(replacePrefix.length);
+    } else {
+      op = "insert";
+      headerArgs = headerLine.slice(insertPrefix.length);
+    }
+
+    const header = parseHeaderArgs(headerArgs, op);
 
     // Body: from the line after the header to the line before the next header (or EOF)
     const bodyStart = headerLineIdx + 1;
@@ -126,9 +141,9 @@ export function parseBlocks(patch, nonce) {
       i + 1 < headerIndices.length ? headerIndices[i + 1] : lines.length;
     const body = lines.slice(bodyStart, bodyEnd);
 
-    if (header.op === "insert" && body.length === 0) {
+    if (op === "insert" && body.length === 0) {
       throw new Error(
-        `Insert block "@@@ ${nonce} ${headerArgs}" has empty body. Use a replace block to delete content.`,
+        "Insert block has empty body. Use a replace block to delete content.",
       );
     }
     blocks.push({ ...header, body });
@@ -221,61 +236,69 @@ export function applyBlocks(original, blocks) {
 
 /**
  * @param {string} headerArgs
+ * @param {"replace" | "insert"} op
  * @returns {{ op: "replace"; start: number; end: number; startHash: string; endHash: string } | { op: "insert"; after: number; afterHash: string }}
  */
-function parseHeaderArgs(headerArgs) {
-  // Replace form: "{start}:{startHash}-{end}:{endHash}"
-  const replaceMatch = headerArgs.match(
-    /^(\d+):([a-f0-9]{2})-(\d+):([a-f0-9]{2})\s*$/,
-  );
+function parseHeaderArgs(headerArgs, op) {
+  if (op === "replace") {
+    // Replace form: "{start}:{startHash}-{end}:{endHash}"
+    const rangeMatch = headerArgs.match(
+      /^(\d+):([a-f0-9]{2})-(\d+):([a-f0-9]{2})\s*$/,
+    );
 
-  if (replaceMatch) {
-    const start = Number(replaceMatch[1]);
-    const end = Number(replaceMatch[3]);
-    if (start < 1) {
-      throw new Error(
-        `Invalid replace range "${headerArgs}": start must be >= 1.`,
-      );
-    }
-    if (end < start) {
-      throw new Error(
-        `Invalid replace range "${headerArgs}": end (${end}) must be >= start (${start}).`,
-      );
+    if (rangeMatch) {
+      const start = Number(rangeMatch[1]);
+      const end = Number(rangeMatch[3]);
+      if (start < 1) {
+        throw new Error(
+          `Invalid replace range "${headerArgs}": start must be >= 1.`,
+        );
+      }
+      if (end < start) {
+        throw new Error(
+          `Invalid replace range "${headerArgs}": end (${end}) must be >= start (${start}).`,
+        );
+      }
+      return {
+        op: "replace",
+        start,
+        end,
+        startHash: rangeMatch[2],
+        endHash: rangeMatch[4],
+      };
     }
-    return {
-      op: "replace",
-      start,
-      end,
-      startHash: replaceMatch[2],
-      endHash: replaceMatch[4],
-    };
-  }
 
-  // Replace form: "{N}:{hash}" (single line replace — shorthand for N:hash-N:hash)
-  const singleReplaceMatch = headerArgs.match(/^(\d+):([a-f0-9]{2})\s*$/);
-  if (singleReplaceMatch) {
-    const start = Number(singleReplaceMatch[1]);
-    if (start < 1) {
-      throw new Error(
-        `Invalid replace range "${headerArgs}": start must be >= 1.`,
-      );
+    // Replace form: "{N}:{hash}" (single line replace — shorthand for N:hash-N:hash)
+    const singleMatch = headerArgs.match(/^(\d+):([a-f0-9]{2})\s*$/);
+    if (singleMatch) {
+      const start = Number(singleMatch[1]);
+      if (start < 1) {
+        throw new Error(
+          `Invalid replace range "${headerArgs}": start must be >= 1.`,
+        );
+      }
+      return {
+        op: "replace",
+        start,
+        end: start,
+        startHash: singleMatch[2],
+        endHash: singleMatch[2],
+      };
     }
-    return {
-      op: "replace",
-      start,
-      end: start,
-      startHash: singleReplaceMatch[2],
-      endHash: singleReplaceMatch[2],
-    };
+
+    throw new Error(
+      `Invalid replace header arguments: ${JSON.stringify(headerArgs)}. Expected "{start}:{startHash}-{end}:{endHash}" or "{N}:{hash}".`,
+    );
   }
 
-  // Insert form: "0+" (no hash — there is no line 0 to verify)
-  if (/^0\+\s*$/.test(headerArgs)) {
+  // op === "insert"
+  // Insert form: "0" (no hash — there is no line 0 to verify)
+  if (/^0\s*$/.test(headerArgs)) {
     return { op: "insert", after: 0, afterHash: "" };
   }
 
-  // Insert form: "{N}:{afterHash}+"
-  const insertMatch = headerArgs.match(/^(\d+):([a-f0-9]{2})\+\s*$/);
+  // Insert form: "{N}:{afterHash}"
+  const insertMatch = headerArgs.match(/^(\d+):([a-f0-9]{2})\s*$/);
 
   if (insertMatch) {
     return {
@@ -286,7 +309,7 @@ function parseHeaderArgs(headerArgs) {
   }
 
   throw new Error(
-    `Invalid block header arguments: ${JSON.stringify(headerArgs)}. Expected "{start}:{startHash}-{end}:{endHash}" or "{N}:{hash}" or "{N}:{afterHash}+" or "0+".`,
+    `Invalid insert header arguments: ${JSON.stringify(headerArgs)}. Expected "{N}:{afterHash}" or "0".`,
   );
 }