mohammadhprp/teammate/recover-run

Claude Code and agent skill by mohammadhprp/teammate

Trust tier 3
  • community
Source check pending

Checking that the SKILL.md file is still in the repository.

Reconcile the durable ledger with the evidence a run left behind after a primary restart or a lost worker: `tm` cannot list live subagents, so classify each task from its brief, report, and diff, cancel what is unsafe, re-delegate what was lost, and never resend a brief that may already have landed. Use when the primary session restarted or compacted mid-run, when a worker is missing or stalled, or when the developer asks what happened to work that was running.

Install

npx vskill@latest install mohammadhprp/teammate/recover-run

SKILL.md

The same SKILL.md was published earlier as mohammadhprp/teammate/recover-run.

Recover a run

A run breaks in two ways: the primary's context is lost (restart or compaction) while workers keep running, or a worker stops making progress. The ledger survives the primary; the subagent calls do not. Because tm is ledger-only it cannot list live subagents, so recovery reconciles the ledger against the evidence on disk — the brief, .teammate-report.md, and the diff.

The dangerous mistake is re-sending a brief whose first delivery may already have taken effect. A worker that received it may already have edited files, and a resend duplicates that work. When delivery cannot be ruled out, escalate: a question is cheaper than a duplicated change.

When to use

  • The primary restarted or compacted while work was in flight.
  • A worker is orphaned, stalled, or never returned.
  • The developer asks what happened to work that was running.

When not to use

  • A healthy run that is merely slow: that is monitor-agents.

Inputs

  • The ledger (state_dir, default ~/.teammate/).
  • The briefs under <state_dir>/<project>/briefs/ and the reports under

<state_dir>/<project>/reports/, or the project's .teammate-report.md.

  • python3 scripts/tm.py diff --cwd "<root>" for what actually changed.
  • Each task's project root and linked worker.

Classify before acting

tm cannot list live subagents, so classify from the ledger and the evidence:

ClassHow you recognize itDefault action
healthya report or a progressing diff exists, and no conflictkeep monitoring (monitor-agents)
orphaneda recorded worker with no report and no matching changedecide from evidence whether to re-delegate or escalate
stalledno report, no new change, past the expected run timeinspect, then re-delegate or escalate
returned a questionthe report or last result asks a questionpause and escalate; never answer it
unresolvedthe evidence is ambiguoustreat as unresolved; inspect or escalate, never as success

Never re-send blindly

Before any re-delegation, check whether the first brief took effect: read .teammate-report.md with python3 scripts/tm.py report, check the change with python3 scripts/tm.py diff, and look for the task's files. If it did any of the work, do not send the same brief again — resume from where it is, or escalate the partial state. Re-delegate only when you can show the original delivery never landed.

Procedure

  1. Rebuild from the ledger. Memory is gone after a restart; the ledger is

not:

``bash python3 scripts/tm.py task list python3 scripts/tm.py task find --worker "<name>" ``

Active statuses are planned, working, awaiting_review, rework, and ready_for_approval; anything else is already closed. task list shows the open session by default, so reconcile only that session — an earlier run's tasks are history, not this run's orphans. If the live ledger is cluttered, python3 scripts/tm.py task prune archives closed tasks (it moves them to <state_dir>/<project>/archive/, never deletes).

  1. Read the evidence for each task. There is no live-worker list; the report

and the diff are what tell resume from redo:

``bash python3 scripts/tm.py report "<name>" python3 scripts/tm.py diff --cwd "<root>" --stat ``

  1. Classify each task with the table above. Classify only after inspecting:

a returned worker is not proof of success, and a report that asks a question still holds useful output.

  1. Cancel what is unsafe. Cancel workers blocked on a decision you cannot

make, writing outside their scope, or racing another stream (see parallel-coordination) through the harness's own control, then record the task cancelled:

``bash python3 scripts/tm.py task update "<id>" --status cancelled ``

  1. Re-delegate what was lost.
  • Still running and its task is intact: keep it, re-establish monitoring

with monitor-agents, and continue the loop.

  • Gone, or the task lost: mark the old task failed, then re-delegate it

to a new worker (delegate-task). If any of its writes may already be in the tree, brief the new worker to reconcile the existing state rather than redo it, or escalate first.

  1. Re-verify recovered work. A tree that survived an interruption has not

been checked since it changed; run verify-evidence before review-work treats it as recovered.

  1. Escalate ambiguity. When you cannot tell whether a brief was delivered, or

what a dead worker changed, stop and ask the developer with the ledger row and the evidence. Duplicating work is worse than one question.

Output

A ledger that matches the evidence on disk, with every task classified and the action taken recorded via python3 scripts/tm.py task update.

Failure and escalation

  • Ambiguous recovery — escalate rather than duplicate work.
  • A worker that returned a question always pauses for the developer.
  • Unresolved evidence is never success; do not mark a task done to clear the

ledger.

Security

  • Security scanPassed, 90 of 100Pattern checks on the skill's files.
  • Model reviewPassed

Quality evaluation

  • Task evaluationNot runNo evaluation yet. Run one in Skill Studio.

Source

Popularity

7-day trend score
0

Not refreshed yet.

Works with

Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.

  • Claude Code
  • Cursor
  • GitHub Copilot
  • Windsurf
  • Codex
  • Gemini CLI
  • Cline
  • Amp
  • Roo Code
  • Goose
  • Continue
  • Kilo Code
  • Junie
  • Trae
  • OpenHands
  • Qwen Code
  • AI Agent
  • Tabnine Chat
  • Sourcegraph Cody
  • Amazon Q
  • Firebase Genkit
  • Sweep
  • Devin
  • Phind
  • Mentat
  • Claude Desktop
  • Aider
  • GPT Pilot
  • Plandex
  • bolt.new
  • v0
  • Replit Agent
  • Anse
  • TypingMind
  • BetterChatGPT
  • LibreChat
  • Warp AI
  • Pieces
  • Fig AI
Embed badge
vskill VERIFIED badge
[![vskill VERIFIED](https://verified-skill.com/api/v1/skills/mohammadhprp/teammate/recover-run/badge)](https://verified-skill.com/skills/mohammadhprp/teammate/recover-run)

Something wrong with this skill or its page? Tell us.