mycoauthor-harness-tasklisted

Use this skill when designing, writing, configuring, or debugging a new phased executor task for the Myco agent harness — even if the user doesn't explicitly ask for a "task authoring" guide. Applies when adding a new intelligence task, modifying phase structure, tuning turn budgets or model routing, adjusting scheduling triggers or session-gating, designing a tool surface, or debugging silent phase failures or budget exhaustion. Covers: YAML task anatomy and registration; phase decomposition and the judgment/recipe gradient; model selection via the advisor pattern; turn budget calibration including local-model multipliers; scheduling triggers and session-gating; tool surface design and readOnly enforcement; Grove scope iteration patterns; per-project lifecycle management; session lifecycle orchestration and agent runtime coordination; and observability via the agent_runs audit table.
goondocks-co/myco · ★ 13 · AI & Automation · score 79

Install: claude install-skill goondocks-co/myco

# Myco Agent Harness Task Authoring The Myco agent harness is a phased executor running inside the daemon. Each task is an ordered sequence of phases — each phase is a single LLM invocation with a bounded tool surface and a turn budget. This skill covers the full authoring lifecycle: designing the phase sequence, writing the task config, selecting models, calibrating budgets, configuring triggers, designing tool surfaces, managing session lifecycle coordination, and debugging when things go wrong. ## Prerequisites - Daemon is running and `agent.enabled: true` in `.myco/myco.yaml`. - You have read at least one existing task YAML (`packages/myco/src/agent/definitions/tasks/vault-evolve.yaml`) to understand the config shape. - You can describe the new task's purpose in one sentence and identify which vault state it reads and writes. - Familiarity with session lifecycle states (CAPTURING, PROCESSING, COMPLETE) and agent runtime coordination. ## Procedure 1: Design the Phase Sequence ### Apply the judgment/recipe gradient Every phase sits on a spectrum from pure-recipe to pure-judgment: | Pole | Characteristics | Typical examples | |------|----------------|------------------| | **Recipe** (deterministic) | Tight tool allowlist, short budget, script-like | Mark processed, cursor update, dedup gate | | **Judgment** (open-ended) | Broader tool access, longer budget, LLM reasons freely | Extract spores, consolidate, generate skill | Position each phase deliberately. Never blur