Scheduler & Self-Improvement

This page covers automated scheduling and the self-repair mechanism.

Scheduler skills (isolated namespace)

Scheduler-triggered skills live in a separate tree, not scanned by host.Scanner():

~/.config/agenvoy/skills/scheduler/<short>-<hash8>/SKILL.md

Creation flow

The scheduler-skill-creator skill is the canonical entry for new schedules. It:

1. Pre-flight gate (Step 0): if the user message lacks a time token (+5m / HH:MM / etc.) or task token, it must call ask_user first — no defaulting to +10m, no inferring "probably 9am".
2. Runs python3 scripts/init_scheduler_skill.py <short> to create the skill dir with hash suffix.
3. Patches the SKILL.md body (description + task + output format).
4. Calls add_schedule to bind the schedule.

Direct add_schedule calls are allowed only for rebinding existing schedules (changing the time of an already-created scheduler skill).

Manual execution (/sched-<name>)

The TUI command picker lists every directory under scheduler/ as /sched-<name>. Selecting one reads the body and dispatches it to the current agent with a preamble that blocks weaker models from misreading the SKILL.md-shaped body as a schedule-creation request and re-running the creator.

The preamble enforces:

• Execute the existing scheduler skill immediately and output results
• Do not activate scheduler-skill-creator
• Do not run init_scheduler_skill.py
• Do not call add_schedule

Daemon fire

runtime.SetRunner registers runSkill(ctx, sessionID, skillName). When the scheduler fires (cron tick or one-shot deadline), the runner:

1. Reads body via filesystem.ScheduleSkillBody(skillName).
2. Ensures the session directory exists; writes a default bot.md.
3. Calls exec.ExecWithSubagent(ctx, body, sessionID, "", "", nil) — an in-process subagent with always-allow context.

One-shot tasks are removed and the skill dir is trashed after a successful fire.

Self-improvement (auto-fix on failure)

When a skill execution encounters tool errors, Agenvoy automatically rewrites the skill definition to prevent the same failure next time. This is a closed-loop evolution cycle — no user intervention required.

How it works

1. Trace collection — during Execute, every tool call result is recorded as an execStep{Tool, Error}. If the skill's session used run_skill to activate additional skills, those are tracked too.
2. Trigger — at the end of Execute (in the defer), if the trace contains at least one error, postSkillImprove runs synchronously for each activated skill.
3. Improvement agent — postSkillImprove loads the built-in improve-skill skill body, builds a stateless session with the execution trace, and runs a full Execute loop (2 min timeout, always-allow, no interactive tools).
4. Auto-commit — on success, skill.AutoCommit(ctx, "improve", skillName) commits the rewritten files under ~/.config/agenvoy/skills/ to git, creating a versioned history of improvements.

What the improve-skill fixes

Constraints

• The improvement agent cannot use interactive tools (ask_user), web tools (search_web, fetch_page), media tools (generate_image), or agent-orchestration tools
• Only file read/write and command execution are available — the agent reads the current skill, analyzes the trace, and writes corrected files
• Improvements are scoped to the skill's own files; system prompts and runtime code are never modified
• Each improvement run is a stateless session with no conversation history

Rollback

All skill changes are auto-committed to git. Use git_log (tag=skills) to inspect history or git_rollback (tag=skills) to revert an unwanted auto-fix.

[!NOTE] This document was auto-generated by Claude after reading the full source code.