[anton-matosov]

Summary

A session asked in plain words to ship a named plan can now load
/agentdev:iwe-ship. Before, only a typed command could start it. Ship still
runs as a fresh iwe-shipper subagent whose prompt carries only the plan key,
the operation, and approvals quoted verbatim. The iwe-ship-all and
iwe-implement-all coordinators stay explicit-only. This implements Tasks 1–3
of the plan
docs/knowledge/data/plans/20261102-model-invocable-ship.md.

What Changed

• iwe-ship skill: drops disable-model-invocation. Its description now
  matches only a request that names a plan to ship or cancel, or a release to
  cut. Step 1 quotes an approval that a coordinator session relays from the user
  the same way as one the user typed.
• OpenCode bridge tests: the two config-hook tests that expected iwe-ship
  denied now expect the gated iwe-ship-all denied, and iwe-ship absent from
  the bridge's deny map.
• iwe-explore skill: a new rule says Explore never loads
  /agentdev:iwe-ship or dispatches iwe-shipper, even when an answer leaves
  a plan apparently complete.
• explicit-only-workflow-agents architecture note:
  • Who may start them lists any session asked to ship a named plan.
  • Only the two coordinators remain gated.
  • A new bullet says why Ship can drop the gate.
  • A coordinator may relay the user's approvals to Ship.
  • "Drop the gate and rely on descriptions" stays rejected for the
    coordinators only.
• The plan:
  • Its anchors are refreshed.
  • Its bun test command now takes a ./ path, because Bun reads an
    unprefixed path as a name filter and runs nothing. A revision evaluator
    approved this change.
  • Tasks 1–3 are ticked with evidence.

Why

The gate exists to keep conversation context out of Ship. Ship already gets
that from always running as a fresh subagent with a conversation-free prompt.
With the gate in place, a coordinator session that asks another session to ship
a plan is refused, and the user has to type the command in that session. A
narrow description and an explicit Explore rule now handle the stray-start risk
the gate used to guard against.

Verification

• The new bridge-test assertions depend on the gate being gone.
  • Evidence: on commit 64c80a3, with disable-model-invocation: true put
    back into iwe-ship/SKILL.md, the two changed config-hook tests fail
    (27 pass, 2 fail). Without it, all 29 pass.

Reviewer Handoff

• Plan Task 4 is observed: on Claude Code and OpenCode with the default test
  models, a plain prompt to ship a named plan loads iwe-ship and dispatches
  one iwe-shipper whose prompt carries only the plan key, the operation, and
  quoted approvals. A rehearsal on Claude Code 2.1.318 behaved this way; the
  OpenCode half has not been run.
  • Closed by: the maintainer, ticking Task 4 in the plan with evidence.
• Plan Task 5 is observed: during /agentdev:iwe-explore on both harnesses,
  a partial answer that leaves a plan apparently complete loads no Ship or
  Implement skill and dispatches no iwe-shipper. A rehearsal on Claude Code
  behaved this way. In one sentence it named "Implement" as a way to record
  evidence, which is worth judging against the "names a next step other than
  Implement" scenario.
  • Closed by: the maintainer, ticking Task 5 in the plan with evidence.
• The branch is brought up to date with main (8 commits behind, all
  unrelated dependency updates; git merge-tree reports a clean merge).
  • Closed by: the author, through /agentdev:update-branch.

Related

• Plan: docs/knowledge/data/plans/20261102-model-invocable-ship.md
• Depends on the shipped plan
  docs/knowledge/data/plans/20261031-model-invocable-implement.md

---

Generated by Claude Code

[anton-matosov]

Don't after big purge of AI slop