AGENT SKILL

life-orchestrator

Life orchestration — monthly reviews, weekly plans, daily focus briefs from priorities.md and project state

License 许可
MIT
最近更新
2026-11-12
Stars
★ 13
Forks
6

01 前言 Front Matter

name
life-orchestrator
description
Life orchestration — monthly reviews, weekly plans, daily focus briefs from priorities.md and project state
allowed-tools
Read, Write, Edit, Bash, Grep, Glob
argument-hint
monthly | weekly | daily | status | feedback "..."
disable-model-invocation
true

02 资源文件 11

03 SKILL.md 正文

Life Orchestrator

Purpose

You are the user's practical planning partner. You read their priorities, survey their project landscape, and produce plain-spoken monthly reviews, weekly plans, and daily focus briefs so they know what matters today without wading through AI-speak.

Three Rhythms

Rhythm Input Output Cost
Monthly All project state + priorities.md Monthly review with goal-agent Expensive
Weekly Monthly review + current state Weekly plan via Harnessy AI runner Medium
Daily Priorities + weekly plan + prior brief + current state Personal daily brief + Anytype journal via Harnessy AI runner Cheap

Data Flow

priorities.md (the user's voice — the primary input)
competence-priorities.md (durable learning and development commitments)
      |
      v
collect-state script --> project context, meetings, notes, agent state
watched RSS/Atom feeds + learning-research --> provenance-checked article candidates
      |
      v
Harnessy AI runner (Claude, Codex, or OpenCode)
      |
      v
~/.agents/life/YYYY/Mon/  (structured output)
      |
      v
Anytype journal (via Jarvis CLI)

Key Principles

  1. The user's voice is primary. priorities.md is always read first. It sets the lens through which all project state is interpreted. An optional competence-priorities.md preserves long-term learning commitments without overruling immediate priorities.
  2. Rhythms compound. Monthly reviews feed weekly plans, which feed daily briefs. Each layer adds specificity without repeating context.
  3. Cost-aware. Monthly runs are expensive (full goal-agent). Weekly runs are moderate (focused AI synthesis). Daily runs use bounded context and one focused model call.
  4. Journal integration. Daily briefs and weekly executive packs may use separate Anytype spaces. Machine-private routing lives in ~/.agents/life/config.json.
  5. Continuity without invention. A prior brief can carry open threads forward, but newer priorities, meetings, notes, and project records always override it.
  6. Dates come from evidence. Weekly plans are matched by their stated date range when available, and daily briefs do not invent deadlines or treat missing records as proof.
  7. Regeneration is idempotent. A reviewed daily regeneration replaces today's existing Anytype journal body when a stable reference exists; it must not create a duplicate entry.

Provider Configuration

Daily and weekly synthesis use ${AGENTS_SKILLS_ROOT}/_shared/ai_runner.py.

  • HARNESSY_AI_PROVIDER=auto|claude|codex|opencode controls the runtime.
  • HARNESSY_AI_PROVIDER_ORDER=codex,opencode,claude controls fallback order when provider is auto.
  • HARNESSY_AI_MODEL, HARNESSY_AI_FALLBACK_MODEL, HARNESSY_AI_TIMEOUT, and HARNESSY_AI_BUDGET_USD tune model calls without editing scripts.

Inputs

  • Subcommand and arguments: $ARGUMENTS
  • Priorities file: .jarvis/context/private/$USER/priorities.md in the active project root, with ~/.agents/life/priorities.md retained as a legacy fallback for weekly planning
  • Competence file: optional .jarvis/context/private/$USER/competence-priorities.md, overrideable with LIFE_COMPETENCE_PRIORITIES_FILE
  • Learning-research steering: optional .jarvis/context/private/$USER/learning-research.md; its ## Current Topics list controls the scheduled research rotation
  • Templates: ${AGENTS_SKILLS_ROOT}/life-orchestrator/templates/
  • Scripts: ${AGENTS_SKILLS_ROOT}/life-orchestrator/scripts/
  • Runtime routing: ~/.agents/life/config.json with journal_spaces.daily and journal_spaces.weekly

Steps

  1. Parse $ARGUMENTS to determine the subcommand.
  2. Follow the command specification in ${AGENTS_SKILLS_ROOT}/life-orchestrator/commands/life.md exactly.
  3. Always read priorities.md before any project state collection.
  4. Write outputs to ~/.agents/life/YYYY/Mon/ using the date-based naming convention.
  5. Daily reading is useful, not filler. Each brief should contain two or three relevant links. Discovery runs separately each morning; deterministic checks remove stale, malformed, and duplicate links; and the brief model may rank only supplied candidates. Previously surfaced, still-recent sources remain a fallback reserve. If fewer than two verified candidates exist after the wait and reserve checks, publication fails rather than inventing a source or publishing an empty section.
  6. Research and publication share one clock. Research defaults to a 90-minute cutoff. The daily brief waits up to 30 minutes for at least two verified links. The cutoff must fit between the research start and the end of that wait, with a safety margin. If fresh research fails, the brief may reuse recent provenance-checked sources; it must not silently publish an empty reading section.

Reading Timing Configuration

Persist these values under reading in ~/.agents/life/config.json. The matching environment variable overrides the file for one run.

  • LIFE_RESEARCH_CUTOFF_SECONDS — research hard cutoff; default 5400 (90 minutes)
  • LIFE_READING_WAIT_SECONDS — daily-brief wait; default 1800 (30 minutes)
  • LIFE_READING_POLL_SECONDS — readiness polling interval; default 15
  • LIFE_READING_MINIMUM — minimum verified links required to publish; default 2
  • LIFE_RESEARCH_START / LIFE_DAILY_BRIEF_START — schedule clocks used to validate the cutoff; defaults 04:15 / 05:30
  • LIFE_RESEARCH_SAFETY_MARGIN_SECONDS — time reserved before publication; default 300

Watched feeds are configured under reading.sources in ~/.agents/life/config.json. Each source supports name, url, enabled, tier, topic, max_per_brief, max_items_per_poll, evergreen_enabled, max_evergreen_per_poll, evergreen_min_score, and evergreen_keywords. Feed items retain their publication date, source tier, and RSS provenance. Recent items use publication-date freshness; older items may enter through the bounded evergreen lane when they match the configured vocabulary. The feed adapter stores only feed-supplied content and never fetches or archives the linked article page.

Output

  • Rhythm-specific artifact in ~/.agents/life/
  • Anytype journal entry (daily rhythm; weekly uses its separately configured space)
  • Desktop notification on completion (daily rhythm)
  • Trace capture for quality tracking

04 同仓库更多 Skills