Agent Plugins 0.14.4

From the round-3 trials on 0.14.3, the first product fix they led to (worktree 0.7.1), and eight
findings from a team adopting ESS. Two isolated trials on this version passed (an ESS retrofit and
worktree onboarding).

Setup

• SETUP.md step 1 asks the user before it installs b10x, showing the target, the download URL
  and the install path. The product init skills say they install CLIs only after the user
  confirms.

ess

• ess:specifying: never invent an entity so that a command type-checks. In an interactive session
  the open UNMAPPED: markers become questions to the user; headless runs and the author agent
  keep them and list them. Marker examples no longer say "Ask".
• ess:specifying: a second source document goes into the same system, a new system or an
  external boundary, by what it says about itself; when it says neither, that is an open question.
• ess:specifying: ## Next starts with the owner reviewing every marker and contradiction, then
  planning (aep:planning), then generate and conformance. The author agent ends its report the
  same way.
• ess:specifying: a headless run needs --allowedTools "Bash(ess:*)" (also on the plugin page).
  Where a repository has no gate, validate plus a synthesis with 0 refusals is the check to report.
• Syntax reference: a predicate is one comparison or a bare fact path; &&, || and in are
  refused and a list's length cannot be tested, so "state A or B" is two views. Error outcomes count
  toward "all but one needs a when"; wrong_state outcomes do not.
• ess:retrofitting: a service that publishes no events declares one per success outcome for the
  fact it produces and leaves it out of publishes:, since validate refuses an outcome with
  nothing observable. An entity the code gives no status gets one structural state, marked as such.

worktree

• worktree:init: worktree doctor --check fails with no active profile since worktree 0.7.1;
  the profiles=0 workaround is gone.

Docs and tooling

• The docs intro and b10x skill --help name the current skills (init, specifying, …).
• agentplugins-check trial-isolation: a skill or agent from outside the sandbox fails a run only
  when the run used it; one merely offered (claude.ai account skills in sub-agent sessions) is a
  note.