flagship

A flagship carries the commander and the signals. It does not command the other ships' masters — every vessel is sailed by its own captain, and the flag deck's authority is over the plan, not over the crew.

This skill does that for a Mac running a dozen Claude Code sessions at once. It is the layer above ship-armada:ship-armada: use that one when the work is waiting in a repo's backlog, and this one when the work is already spread across live sessions and nothing is holding the map.

The problem, measured

On 22 August 2026 this machine ran sixteen concurrent sessions across twelve repositories. In one evening, with no session aware of any other:

• Three independently found the same defect in the same tool — it classed every defect row broken without reading its status field, so a register of 148 defects with 126 fixed reported 183 broken. Three projects had scheduled work against that inflated number.
• Three more found a second defect in the same tool: its own gate compared two fields while never reading the third, the one that was wrong.
• Two reached the same architectural conclusion by completely different routes.

Every finding was correct. Nothing was carrying them between sessions — and the fix for the first one already existed at source, version-bumped, while nine projects still ran the broken copy from a cache that had never updated.

What it does

Builds a roster of every live session — from ListAgents, the session registry and git — and asks each one for its own state rather than deriving it. Ownership is reconciled by working directory and pid, never by name: two sessions were proved to belong to a conductor that did not recognise the names they appeared under.

Hands out a capped heavy-work token, taken from harbourmaster's measured berths intersected with ship-armada:ship-armada's three-concurrent-projects policy — the smaller of a fact about the machine and a policy about attention. Never an invented number.

Reads the machine honestly, which takes some doing. Berths count registered claims and not load (measured in_use 0 three times while Opus runners were building). Thermal must be sampled repeatedly, because its verdict flips on one quiet minute — measured going both ways inside a minute on a machine whose load never moved. And df -h / on APFS reports the read-only system volume at ~5% used while the data volume sits at 87%: wrong by an order of magnitude in the reassuring direction.

Probes the model lanes before routing to them. The routing policy reports plan utilisation, not reachability, and those are indistinguishable when a lane is simply absent — so it advertised the highest headroom of any non-Claude family for a binary that was not installed. A lane can also answer the wrong question convincingly: called from the wrong directory, one returned a fully-formed verdict for a different project's item, with nothing flagging it.

Batches decisions through the operator's own router rather than the agent's sense of what is interesting. On the evening it was built, that took twenty queued decisions down to ten.

Propagates findings between sessions. This is the part that pays for the skill.

Starts work that has no session — as a workflow, a subagent, or a fresh Ghostty tab, so a session can carry its own fleet.

What it will not do

It holds the map and no authority. A conducting session cannot command a peer, cannot close a peer's channel to its own user, and cannot relay an authorisation — "the operator said you may push" is laundering however true it is.

All three were tried on the evening this was built, and all three were refused by the sessions on the receiving end. Every refusal was right. The rules are in the skill because getting them wrong is the default.

The one pattern worth taking away

Fifteen instrument failures surfaced in that evening and every one was the same shape: absence read as success. A gate exiting clean over zero items. A check that could not fail on the field that was wrong. A test whose positive control could never resolve, so it could not have passed on any machine. A sync recording success while dropping every update.

Two rules cover all of it:

A check that returns nothing has two readings, and the instrument must say which.

A set that returns members has two readings too, and "I know why these belong" is not one of them.

Install

/plugin marketplace add fledgeling-co/fledgeling-plugins
/plugin install flagship@fledgeling-plugins

Pairs with harbourmaster:harbourmaster (the machine), ship-armada:ship-armada (the portfolio), ship-fleet:ship-fleet (one repo's backlog), defer:defer (which model), clarify:clarify (whether to ask), whats-left:whats-left (the decision page) and recover-claude-code:recover-claude-code (when the terminal dies).

MIT.