hajimi-claw 0.1.3 → 0.1.4

package/README.md

@@ -14,6 +14,8 @@ Single-user Telegram/Feishu-first ops agent in Rust.
 - Windows-safe execution mode with allowlist checks and Job Object cleanup
 - Channel-aware onboarding for Telegram or Feishu plus provider/model setup
 - Configurable multi-agent orchestration with coordinator/worker/integrator flow
+- Executable skills registered as first-class tools with shared approval, audit, and persistence
+- Capability inventory and MCP server status surfaced in chat via `/capabilities`, `/skills`, and `/mcp`
 
 ## Running
 
@@ -171,6 +173,69 @@ Current Feishu limitations:
 `hajimi ask` now records task state and tool invocations in SQLite. If a guarded command blocks on
 approval, `hajimi approve <request-id>` resumes the blocked task instead of asking you to rerun it.
 
+## Skills and MCP
+
+Hajimi now treats native tools, executable skills, and MCP-discovered tools as one capability
+surface.
+
+- Native tools keep their existing names such as `read_file` or `exec_once`
+- When `[telegram].bot_token` is configured, a native `telegram_api` tool is also registered
+- Executable skills are configured in `[skills]` and are registered as tool names like
+  `skill.deploy`
+- MCP tools will be exposed as namespaced tool names like `mcp.<server>.<tool>`
+- All three flow through the same runtime path for approval, audit logging, task persistence, and
+  model tool-calling
+
+`skills.md` remains prompt guidance only. Use it for playbooks, habits, and routing hints. It is
+not the executable source of truth for runnable skills.
+
+### Config
+
+`config.example.toml` now includes these sections:
+
+```toml
+[skills]
+enabled = true
+directory = "./skills"
+manifest_paths = []
+entries = []
+
+[mcp]
+enabled = true
+servers = []
+```
+
+Skill manifests and inline entries deserialize into `ExecutableSkillConfig`, including `name`,
+`description`, `command`, `args`, `cwd`, `env_allowlist`, `requires_approval`, `timeout_secs`,
+`max_output_bytes`, and `input_schema`.
+
+Relative paths in skill manifests, `skills.directory`, `skills.manifest_paths`, and MCP server
+`cwd` values resolve relative to the config file.
+
+### Telegram capability commands
+
+- `/capabilities` — list the effective native tools, executable skills, and MCP tools
+- `/skills` — list executable skills only
+- `/skill run <name> <json-or-text>` — explicitly invoke one configured skill through the runtime
+- `/mcp` — show configured MCP server status
+- `/mcp tools [server]` — list discovered MCP tools, optionally filtered by server
+
+Natural-language requests stay primary. Once a skill or MCP tool is registered in the runtime, the
+model can choose it during normal `/ask` or plain-text requests just like built-in tools.
+
+`telegram_api` calls the configured Telegram Bot API token directly from the native tool layer. A
+typical invocation looks like:
+
+```json
+{
+  "method": "sendMessage",
+  "params": {
+    "text": "deploy finished"
+  },
+  "use_default_chat_id": true
+}
+```
+
 ## Multi-Agent
 
 `hajimi` can split one natural-language request into multiple sub-agents. This is configured in
@@ -210,25 +275,43 @@ Behavior:
 
 ## Persona Files
 
-`hajimi onboard` now creates empty persona files in `~/.hajimi/`:
+`hajimi onboard` seeds persona files in `~/.hajimi/`:
 
+- `identity.md`
 - `soul.md`
 - `agents.md`
 - `tools.md`
 - `skills.md`
+- `heartbeat.md`
+
+`hajimi` reloads persona prompt files on each request, so edits take effect on the next `/ask`.
+
+Layered prompt order:
+
+1. base system prompt
+2. `identity.md`
+3. `soul.md`
+4. `agents.md` / `AGENTS.md` / `tools.md` / `skills.md`
+5. runtime overlays such as shell-session metadata and multi-agent role instructions
 
-`hajimi` reloads these files on each request, so Telegram edits take effect on the next `/ask`.
+Auto-discovery behavior:
 
-- Auto-discovered files: `soul.md`, `agents.md`, `AGENTS.md`, `tools.md`, `skills.md`
-- Search roots: the current working directory, the config directory, and `~/.hajimi`
+- Auto-discovered files: `identity.md`, `soul.md`, `agents.md`, `AGENTS.md`, `tools.md`, `skills.md`
+- Search roots and precedence: `persona.directory` -> config directory -> current working directory
+- `identity.md` and `soul.md` support optional front matter, but plain markdown still works
+- Structured `identity.md` / `soul.md` fields override by precedence while freeform notes accumulate
+- Extension files stay additive in precedence order
+- `heartbeat.md` is runtime config only and is never appended into the prompt
 - Optional explicit list: set `[persona].prompt_files` in `config.toml`
 
 Use these files for:
 
-- `soul.md`: tone, temperament, and high-level persona
+- `identity.md`: user profile, owned systems, environments, durable preferences, and hard constraints
+- `soul.md`: Hajimi's stable role, tone, style, and behavioral stance
 - `agents.md` or `AGENTS.md`: repo or operator instructions
 - `tools.md`: tool-use policy and operational preferences
-- `skills.md`: extra habits, playbooks, and skill-selection hints
+- `skills.md`: extra habits, playbooks, and skill-selection hints for the prompt layer only
+- `heartbeat.md`: daemon heartbeat runtime config
 
 ## Telegram commands
 
@@ -245,10 +328,16 @@ Use these files for:
 - `/model current`
 - `/model use [model]`
 - `/persona list`
-- `/persona read <soul|agents|tools|skills>`
+- `/persona guide`
+- `/persona read <identity|heartbeat|soul|agents|tools|skills>`
 - `/persona write <file> <content>`
 - `/persona append <file> <content>`
 - `/ask <text>`
+- `/capabilities`
+- `/skills`
+- `/skill run <name> <json-or-text>`
+- `/mcp`
+- `/mcp tools [server]`
 - `/shell open [name]`
 - `/shell exec <cmd>`
 - `/shell close`

package/config.example.toml

@@ -4,6 +4,8 @@ kind = "telegram"
 [telegram]
 bot_token = "123456:replace-me"
 poll_timeout_secs = 30
+# When bot_token is configured, hajimi also registers a native `telegram_api` tool.
+# Set `use_default_chat_id = true` in tool input to reuse policy.admin_chat_id.
 
 [feishu]
 # `mode = "webhook"`: expose listen_addr + event_path through your reverse proxy or public host.
@@ -47,12 +49,30 @@ worker_timeout_secs = 90
 max_context_chars_per_worker = 24000
 
 [persona]
-# hajimi onboard creates soul.md, agents.md, tools.md, and skills.md here.
+# hajimi onboard seeds identity.md, soul.md, agents.md, tools.md, skills.md, and heartbeat.md here.
 directory = "~/.hajimi"
-# Optional. If empty, hajimi auto-loads soul.md, agents.md, AGENTS.md, tools.md,
-# and skills.md from the current working directory, the config directory, and persona.directory.
+# Optional. If empty, hajimi auto-loads identity.md, soul.md, agents.md, AGENTS.md,
+# tools.md, and skills.md from persona.directory, the config directory, and the current
+# working directory, in that precedence order.
+# identity.md and soul.md are layered first-class persona files; heartbeat.md is runtime
+# config only and is never appended into the prompt.
 prompt_files = []
 
+[skills]
+# Executable skills are registered as tools named `skill.<name>`.
+enabled = true
+# Optional directory of TOML manifests. Relative paths resolve from this config file.
+directory = "./skills"
+# Optional extra manifest files.
+manifest_paths = []
+# Optional inline entries.
+entries = []
+
+[mcp]
+# MCP servers currently surface inventory/status in the runtime. Tool transport wiring is next.
+enabled = true
+servers = []
+
 [policy]
 admin_user_id = 123456789
 admin_chat_id = 123456789

package/npm/platforms/linux-arm64-gnu/package.json

@@ -1,6 +1,6 @@
 {
   "name": "hajimi-claw-linux-arm64-gnu",
-  "version": "0.1.3",
+  "version": "0.1.4",
   "description": "Prebuilt Linux arm64 glibc binary for hajimi-claw",
   "os": [
     "linux"

package/npm/platforms/linux-x64-gnu/package.json

@@ -1,6 +1,6 @@
 {
   "name": "hajimi-claw-linux-x64-gnu",
-  "version": "0.1.3",
+  "version": "0.1.4",
   "description": "Prebuilt Linux x64 glibc binary for hajimi-claw",
   "os": [
     "linux"

package/npm/platforms/win32-arm64-msvc/package.json

@@ -1,6 +1,6 @@
 {
   "name": "hajimi-claw-win32-arm64-msvc",
-  "version": "0.1.3",
+  "version": "0.1.4",
   "description": "Prebuilt Windows arm64 binary for hajimi-claw",
   "os": [
     "win32"

package/npm/platforms/win32-x64-msvc/package.json

@@ -1,6 +1,6 @@
 {
   "name": "hajimi-claw-win32-x64-msvc",
-  "version": "0.1.3",
+  "version": "0.1.4",
   "description": "Prebuilt Windows x64 binary for hajimi-claw",
   "os": [
     "win32"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "hajimi-claw",
-  "version": "0.1.3",
+  "version": "0.1.4",
   "description": "Single-user Telegram-first ops agent in Rust",
   "bin": {
     "hajimi": "bin/hajimi-claw.js",
@@ -12,10 +12,10 @@
     "node": ">=18"
   },
   "optionalDependencies": {
-    "hajimi-claw-linux-arm64-gnu": "0.1.3",
-    "hajimi-claw-linux-x64-gnu": "0.1.3",
-    "hajimi-claw-win32-arm64-msvc": "0.1.3",
-    "hajimi-claw-win32-x64-msvc": "0.1.3"
+    "hajimi-claw-linux-arm64-gnu": "0.1.4",
+    "hajimi-claw-linux-x64-gnu": "0.1.4",
+    "hajimi-claw-win32-arm64-msvc": "0.1.4",
+    "hajimi-claw-win32-x64-msvc": "0.1.4"
   },
   "files": [
     "bin",