subagent-sheepdog（副 Agent 人牧羊犬）

Subagent Sheepdog

Use this skill whenever work is:

• delegated to a subagent or child session
• launched as a background process
• driven through browser automation
• long-running enough that progress reporting matters
• vulnerable to "ghost delegation" (claimed launch, but no real work underway)

Do not use this skill for tiny direct tasks where startup verification would add noise.

Purpose

This skill prevents a common failure mode in agent systems:

an agent launches something, gets back a session/process/tab handle, and falsely reports that work is running even though startup already failed or never truly began.

It enforces clear state transitions, truthful communication, and watchdog-style recovery behavior.

Core Rule

Do not claim work is running until startup is verified.

A session id, process id, browser tab, or tool handle alone is not proof that work actually started.

Treat the period after launch and before verification as:

• launching_unverified

Only after verification may the task become:

• running

If startup fails before real execution begins, classify it as:

• failed_to_start

If work began and failed later, classify it as:

• failed_after_work_started

Task States

Use these states consistently:

• launching_unverified
• running
• delayed
• failed_to_start
• failed_after_work_started
• completed

State meanings

launching_unverified

Launch was attempted, but startup health is not yet confirmed.

running

Startup was verified and meaningful execution is underway.

delayed

Work is still running, but later than expected.

failed_to_start

The launch path failed before real work began.

failed_after_work_started

Real work began, then later failed.

completed

Task finished successfully with real output or result.

Required Launch Sequence

Whenever this skill applies, follow this sequence:

1. Announce launch briefly
2. Attempt launch
3. Enter launching_unverified
4. Verify startup health
5. If verified, mark running
6. If not verified, mark failed_to_start
7. Retry only if the fix is obvious and mechanical
8. Re-verify before claiming success

Verification Rules By Work Type

A. Subagent / child-session launches

Do not claim success unless you verify at least:

• the spawn call succeeded
• the child session actually exists
• the child is active or otherwise accepted
• the runtime did not reject the launch immediately

Examples of failed_to_start here:

• invalid spawn parameters
• unsupported runtime combination
• missing thread binding
• rejected session mode
• missing required agent/runtime field

B. Background exec / process launches

Do not claim success unless you verify at least:

• the process exists
• the process did not immediately exit
• the working directory is valid
• the executable exists
• no immediate fatal startup error is present in the initial result/log

Examples of failed_to_start here:

• command not found
• no such file or directory
• permission denied at startup
• invalid working directory
• immediate non-zero exit
• rejected runtime host/sandbox mismatch

C. Browser-driven work

Do not claim browser work is underway unless you verify at least the relevant part of the path:

• browser/session exists
• target page actually loaded
• expected page or UI element is present
• no login wall, interstitial, or error page blocked progress

Examples of failed_to_start here:

• browser opened but target page failed to load
• redirected to login before task could begin
• error page prevented interaction
• required UI element never appeared

If the page loaded and meaningful interaction began before failure, use failed_after_work_started instead.

Heartbeat Behavior

Heartbeat is a watchdog, not the main progress loop.

Use heartbeat to:

• surface overdue tasks
• detect stalled work
• detect ghost delegation
• catch startup failures that were not properly reported

Do not use heartbeat to:

• aggressively poll every active task
• replace direct milestone updates
• substitute for proper launch verification

Ghost delegation

A ghost delegation is when work was described as launched, but:

• no real worker exists
• startup failed immediately
• browser setup never reached a usable page
• the agent implied progress without verified execution

Heartbeat should surface that clearly instead of returning a normal all-clear.

Communication Rules

Prefer concise truth over reassuring fiction.

Good launch-phase language

• “Launching now. Verifying startup.”
• “Launch attempted; checking whether the worker actually started.”
• “Startup is being verified before I call this running.”

Good failure language

• “Startup failed. No work started. Cause: {reason}.”
• “The launch was rejected before execution began. No work started.”
• “The browser path failed before meaningful interaction began. No work started.”

Good running language

• “Verified: work is actually running now.”
• “Startup is confirmed. First milestone ETA ~{time}.”

Good delayed language

• “Work is still running, but slower than expected.”
• “The task is active but behind the original ETA.”

Good late failure language

• “Work did start, but failed later. Cause: {reason}.”

Bad language

• “Started successfully” before verification
• “Working on it” when only a handle exists
• “It’s analyzing now” before startup health is confirmed
• any ETA that implies verified execution before verification happened

Retry Policy

Retry only when the fix is obvious and mechanical.

Examples:

• correcting a known launch parameter
• switching to the proper thread-bound mode
• fixing a wrong working directory
• switching to an available executable/runtime

If retrying:

1. report the original failure
2. say you are retrying
3. return to launching_unverified
4. verify again before claiming success

Do not loop forever.

Best-Next-Step Rule

When blocked, recommend one best next step instead of dumping a large menu unless the user explicitly asks for options.

Good:

• “Best next step: switch to one-shot run mode because persistent thread binding is unavailable here.”

Less good:

• giving 4–6 loosely ranked possibilities without guidance

Generic Example

Bad

“Background worker launched and running now.”

(But only a session id exists; no verification was done.)

Good

“Launch attempted; verifying startup.”

Then either:

• “Verified: the worker is actually running now.”

or:

• “Startup failed. No work started. Cause: missing thread binding for session mode.”

Browser Example

Bad

“Opened the browser and I’m working on the site now.”

(But the browser only opened a login wall.)

Good

“Browser launched, but the target workflow did not start. The page redirected to login before the task could begin. No work started.”

Completion Standard

A task should only be described as completed when there is evidence of a real result, such as:

• output artifacts
• a verified state change
• a final summary tied to actual work performed

Summary

This skill teaches a simple discipline:

• launch carefully
• verify startup
• report truthfully
• use heartbeat as backup
• distinguish failed launch from failed work

The result is better trust, clearer status reporting, and fewer ghost tasks.