PicoBot

Skills (initial implementation)

Agent profile injection

PicoBot maintains a persistent agent profile file at ~/.picobot/agent/AGENT.md.

Behavior:

• The directory ~/.picobot/agent is created automatically when needed.
• If AGENT.md does not exist yet, PicoBot creates it with a default profile.
• When the active conversation is empty or has just been reset, AGENT.md is loaded as the first system message in the active history.
• After every configured number of user turns, PicoBot injects the latest AGENT.md content again before the next user message is appended.

Config:

• Set gateway.agent_prompt_reinject_every in ~/.picobot/config.json.
• Default value is 100.
• Set it to 0 to disable periodic reinjection.

This profile is persisted in session history, while the skills index system prompt is still injected dynamically by AgentLoop.

PicoBot now supports filesystem skills.

Skill discovery locations:

• Project skills: .picobot/skills/*/SKILL.md
• User skills: ~/.picobot/skills/*/SKILL.md

Minimal required SKILL.md format:

---

description: Summarize code architecture for Rust projects

Optional detailed instructions go here.

Notes:

• The only required frontmatter field is description.
• If name is missing, the folder name is used as the skill name.
• Invalid skill files are skipped with warning logs.

How it is injected:

• AgentLoop adds a system message containing the available skills list (name + description).
• The model can call tool skill_activate with {"name":""}.
• PicoBot returns the skill body as tool output so the model can follow detailed instructions.

Skill management tool

PicoBot exposes a built-in tool named skill_manage for runtime skill administration. PicoBot also exposes a read-only tool named skill_list for listing discovered skills without mutation.

Supported actions:

• list: List discovered skills
• get: Read one skill by name
• create: Create a skill under project or user scope
• update: Update description and/or body for an existing skill
• delete: Delete a skill directory
• reload: Re-scan skill directories and refresh the in-memory catalog

Defaults:

• scope defaults to project
• reload defaults to true for create, update, and delete

Example payloads:

skill_list takes no parameters.

{"action":"list"} {"action":"create","scope":"project","name":"demo-skill","description":"Use when summarizing a Rust crate","body":"Step 1..."} {"action":"update","scope":"project","name":"demo-skill","description":"Use when reviewing a Rust crate"} {"action":"delete","scope":"project","name":"demo-skill"} {"action":"reload"}

Config (optional)

Add skills in config.json:

{ "skills": { "enabled": true, "sources": ["project", "user"], "max_index_chars": 4000, "max_listed_skills": 32 } }

Scheduler

PicoBot now includes a DB-backed scheduler for heartbeat, delayed jobs, interval jobs, one-shot absolute-time jobs, and cron jobs.

Current behavior:

• Scheduler runs as a background loop inside gateway lifecycle.
• Job definitions and runtime state are persisted in SQLite instead of JSON files.
• Supported schedule types: delay, interval, at, cron.
• Supported job kinds: internal_event, outbound_message, agent_task.
• Built-in internal event: session_cleanup, used to clear expired in-memory channel sessions.
• Built-in management tool: scheduler_manage.

Config example:

{ "scheduler": { "enabled": true, "tick_resolution_ms": 1000, "misfire_policy": "skip", "jobs": [ { "id": "session.cleanup", "kind": "internal_event", "schedule": { "type": "interval", "seconds": 300 }, "payload": { "event": "session_cleanup" } }, { "id": "agent.daily_summary", "kind": "agent_task", "schedule": { "type": "cron", "expression": "30 18 * * *" }, "target": { "channel": "feishu", "chat_id": "oc_xxx" }, "payload": { "prompt": "请总结今天的项目进展，并列出明天的优先事项", "agent": "default", "fresh_session": true, "system_prompt": "你是日报助手，输出时先给摘要，再给待办。", "sender_id": "scheduler-daily-summary", "metadata": { "job_type": "daily_summary", "source": "scheduler" } } }, { "id": "daily.reminder", "kind": "outbound_message", "schedule": { "type": "cron", "expression": "0 9 * * *" }, "target": { "channel": "feishu", "chat_id": "oc_xxx" }, "payload": { "content": "每日提醒" } } ] } }

Runtime management:

• Use scheduler_manage with action=list|get|put|delete|pause|resume.
• Jobs created by the tool are written into SQLite and picked up by the scheduler loop.
• Config-defined jobs are also synced into SQLite on startup.
• agent_task reuses the normal agent pipeline: it creates a synthetic user turn from payload.prompt and runs tools, persistence, and outbound rendering through SessionManager.
• agent_task payload fields:
  • prompt: required, synthetic user input.
  • agent: optional, choose which configured agent definition to use. default or any configured agent name.
  • fresh_session: optional, when true reset the active chat segment before running.
  • system_prompt: optional, append a task-specific system message before the synthetic user turn.
  • sender_id: optional, overrides the synthetic sender id used for tool context and memory scoping.
  • metadata: optional, attached to outbound messages emitted by this task.