Перейти к основному содержимому

Перевод: оболочка страницы на русском; тело навыка (SKILL.md) пока на английском — это исходные инструкции агента.

Progress Ledger

Keep a Magentic-style task and progress ledger.

Метаданные навыка​

ИсточникВстроенный (установлен по умолчанию)
Путьskills/research/progress-ledger
Версия1.0.0
АвторVibeOS
ЛицензияMIT
Платформыlinux, macos, windows
Тегиresearch, orchestration, multi-agent, kanban, planning
Связанные навыкиanalysis-orchestration, professional-research, challenge-gate

Справка: полный SKILL.md​

к сведению

Ниже полное определение навыка, которое VibeOS загружает при активации. Это инструкции, которые видит агент.

Progress Ledger Skill

Maintain a Task Ledger (facts, guesses, plan) and a Progress Ledger (current step, assignee, stall count, replan notes) for long jobs. Inspired by Magentic-One — without peer agent chat. Use files or a Kanban root-task comment; do not invent a new core tool.

When to Use​

  • Multi-step research, creative, or operate jobs longer than a few tool rounds.
  • delegate_task batches or Kanban swarms where the parent must replan on stall.
  • User asks to “track progress”, “don’t get stuck”, or “show the plan”.

Prerequisites​

  • Writable workdir (session cwd or Kanban workspace).
  • Optional: Kanban board for durable jobs (vibeos kanban / kanban_* tools).
  • Related: analysis-orchestration for gatherer/critic/synthesizer roles.

How to Run​

  1. Copy templates/ledger.md to the workdir as progress-ledger.md (or pin the same JSON block as a Kanban comment on the root task).
  2. At job start: fill Task Ledger — known facts, open guesses, ordered plan.
  3. Each major step: update Progress Ledger — current step, who owns it (role/profile), stall_count, last_progress_at, blockers.
  4. If stall_count ≥ 2 with no new artifacts: replan (outer loop) — revise Task Ledger plan, reset stall, note what failed.
  5. On finish: set status: done and point to final artifact paths.

Quick Reference​

FieldWherePurpose
factsTask LedgerConfirmed evidence
guessesTask LedgerUnverified hypotheses
planTask LedgerOrdered steps
current_stepProgress LedgerWhat is in flight
assigneeProgress LedgerRole/profile owning the step
stall_countProgress LedgerConsecutive no-progress ticks
blockersProgress LedgerCAPTCHA, missing keys, waits

Template path (relative to this skill): templates/ledger.md.

Procedure​

Outer loop (Task Ledger)​

When starting or after a stall replan:

  1. List facts you would bet money on (with source pointers).
  2. List guesses separately — never mix into facts.
  3. Write a short plan (5–12 steps). Prefer durable Kanban tasks for steps that must survive restart.

Inner loop (Progress Ledger)​

After each child/delegate_task result or Kanban heartbeat window:

  1. Ask: did we gain a new artifact, finding, or decision?
  2. If yes → update current_step, reset stall_count to 0.
  3. If no → increment stall_count; if ≥ 2, jump to outer-loop replan.
  4. Record human gates (assignee: human, ChallengeGate) as blockers — do not thrash tools behind a wall.

Persistence choices​

Job lengthStore ledger in
Single chat turn / short batchprogress-ledger.md in workdir
Multi-hour / restart-safeKanban root task via CLI (below) + file copy
Cron / monitorJob workdir file; optional archival memory later

Kanban persistence (CLI)​

Upsert the ledger onto the root task (survives restart):

vibeos kanban ledger <root_task_id> --set progress-ledger.md
vibeos kanban ledger <root_task_id> # print body
vibeos kanban ledger <root_task_id> --json # machine-readable

Dashboard: GET/PUT /api/plugins/kanban/tasks/{id}/ledger.
Marker: <!-- vibeos:progress_ledger v1 --> (upsert keeps a single comment).

Promote process-local work → Kanban (manual)​

Background delegate_task is process-local — it dies on app restart. When a job outlives a short batch (Fleet shows a promote hint after delegation.promote_hint_after_seconds, default 600s), manually create a durable task. Never auto-create Kanban rows without an explicit command.

# Flags
vibeos kanban from-delegation --title "Continue research" --goal "…" \
--context "constraints…" --summary "partial findings…" \
[--ledger-file progress-ledger.md] [--assignee profile] [--ready]

# Or JSON handoff file
vibeos kanban from-delegation --from-json handoff.json --json

Then keep updating the ledger on the new task id with vibeos kanban ledger.

Pitfalls​

  • Do not dump the full parent transcript into children — put decisions and constraints into the ledger / delegate_task context package.
  • Do not treat peer agent chat as the ledger — comments drift; structured fields do not.
  • Leaves cannot write memory; durable facts go to the ledger / Board KB / parent promotion paths.
  • Avoid change-detector “progress” (rewriting the same paragraph) — require artifacts or verified findings.

Verification​

  • progress-ledger.md (or Kanban comment) exists with both ledger sections.
  • After a deliberate stall, stall_count rises and a replan note appears.
  • Final status is done or blocked with human-readable blockers.
  • Fleet / user can open the ledger file without reading raw agent logs.