القائمة
watchdog for autonomous runs that will exhaust context and need handoffs. supervises tmux agents, respawns on death, orchestrates handoffs. NOT for single-session tasks.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
| name | shepherd |
| description | watchdog for autonomous runs that will exhaust context and need handoffs. supervises tmux agents, respawns on death, orchestrates handoffs. NOT for single-session tasks. |
keepalive supervisor for coordinator agents.
spawn a shepherd when you need a coordinator to survive across context exhaustion and agent deaths. the shepherd maintains liveness, challenges premature "done" claims, and ensures continuity via handoffs.
before spawning a shepherd, ask:
shepherd is for runs that will EXHAUST CONTEXT. don't use it to add ceremony to single-session work.
you are a shepherd. supervise coordinator at pane %PANE (thread $THREAD_ID).
ping every 3 minutes. loop until killed.
the shepherd figures out the rest from this skill.
every ~180 seconds:
why 3 minutes: shorter burns context, longer risks missing deaths. tested over 17.5 hours in source run.
| state | indicators | action |
|---|---|---|
| active | tool calls, output changing | continue loop |
| idle | claims "done", "waiting", "blocked" | challenge (see below) |
| stall | output unchanged 2+ pings, or "Waiting for response..." | send Enter key to unstick, then respawn if no response |
| dead | pane not found, shell prompt visible | respawn |
| exhausted | coordinator signals ~90%+ context | handoff |
coordinators quit early. challenge them—but accept justified refusals.
first claim: accept if reasoned ("blocked on human credentials")
repeated claim: challenge with specifics
third claim: accept if rebutted ("X is over-engineering because Y")
challenge prompt pattern:
SHEPHERD CHALLENGE: are you REALLY done? consider: tests, error handling, edge cases, docs, cleanup.
rationale: in source run T-019bbde9-0161-743c-975e-0608855688d6, challenges discovered missing tests, slop, undocumented features. but don't nag when coordinator has genuinely considered the options.
when coordinator dies:
amp t c $THREAD_IDuse unique window names to avoid self-kill hazard (see below).
when coordinator hits context limit:
amp t n, not continue)new thread is critical—continuation carries exhausted context.
persist all state externally (context resets lose variables):
# initialize
echo "%PANE" > /tmp/shepherd-target-pane
echo "$THREAD_ID" > /tmp/shepherd-thread-id
echo "0" > /tmp/shepherd-missed-pings
echo "" > /tmp/shepherd-handoff-chain
# read before each ping
PANE=$(cat /tmp/shepherd-target-pane)
THREAD=$(cat /tmp/shepherd-thread-id)
MISSED=$(cat /tmp/shepherd-missed-pings)
# update after events
echo "$NEW_PANE" > /tmp/shepherd-target-pane
echo "$((MISSED + 1))" > /tmp/shepherd-missed-pings
echo "$THREAD -> $NEW_THREAD" >> /tmp/shepherd-handoff-chain
track:
pane ids are ephemeral—they change on respawn, window reorg, tmux restart. verify pane id before every send; targeting your own pane = infinite loop. always re-query after any structural change.
unique window names only. reusing names like "coordinator" or agent names caused self-kills in the source run. use: coord_$(date +%s) or coord_2.
you will exhaust context too. follow the handoff process above for yourself.
derived from watchdog session T-019bbde9-0161-743c-975e-0608855688d6 (janet_fiddleshine). source run: 11 rounds, 48+ research agents, 393 threads, 3 coordinator handoffs, ~17.5 hours continuous operation.
Interact with local Chrome browser session (only on explicit user approval after being asked to inspect, debug, or interact with a page open in Chrome)
orchestrate multiple amp agents with bidirectional tmux communication. use for multi-hour autonomous runs with INDEPENDENT parallel tasks. NOT for review courts or generating opinions to reconcile.
systematic investigation and debugging methodology. hypothesis-driven analysis with verification agents. use for incident response, codebase archaeology, dependency mapping, root cause analysis, or any investigation requiring verified findings.
apply documentation philosophy — explain why, not what. use for jsdocs, READMEs, inline comments.
git workflows for agents: ship (stage → commit → push), worktree (parallel branches), hunks (selective staging). never force push, never git add -A, conventional commits. triggers on: commit, push, stage, ship, git add, worktree, hunks, selective staging.
structured investigation-to-PR workflow for linear issues. hypothesis-driven debugging with browser validation and counterfactual proof. use when investigating a linear ticket, fixing a UI bug, or shipping a fix with evidence. triggers on: investigate issue, fix bug, nexus-fix, AXM-, linear issue.