Ship Fleet

Whether an edit to plugin/team-lead-fleet/ reaches anyone depends on how each session resolved the plugin, and that is not visible from the repo, the version number, or the drift check.

Two resolution modes exist on this machine at once:

• Source-resolved — the session runs the plugin straight out of the repo checkout. An edit is live at that session's next SessionStart. No version bump, no claude plugin update.
• Cache-resolved — the session runs a version-keyed copy under ~/.claude/plugins/cache/. claude plugin update skips that copy when the version is unchanged, so edits reach it only after a version bump.

Settled 2026-09-15: the fleet is cache-resolved, uniformly

The marketplace moved from the local directory ~/dev/ai-team-lead/plugin to the GitHub repo fryanpan/ai-team-lead. There is no source-resolved path any more — every session reads a version-keyed copy under ~/.claude/plugins/cache/, so a version bump is always required and an edit in the working tree reaches nobody until it is merged to main and pulled.

The probe below is kept because it is the way to confirm what a peer actually loaded, and because a peer can still be running an older cached version. But its two-modes framing is history: the answer to "which mode" is now always cache-resolved.

The rule

Determine resolution mode before you decide what shipping requires — and confirm delivery from a peer's own injected context, never from the repo.

The probe

Grep a peer's latest injected rules block for a string that exists in exactly one of the two copies. A file deleted from the repo but still in the cache is ideal: present means cache-resolved, absent means source-resolved.

Read the latest injection only. A transcript accumulates every SessionStart it has ever had, plus tool output and messages that merely mention the string. Counting matches across the whole .jsonl measures history, not current state — that error produced a confident, wrong fleet-wide claim on 2026-08-20.

python3 - "$TRANSCRIPT" <<'PY'
import json, sys
last = None
for line in open(sys.argv[1]):
    if 'team-lead-fleet rule:' not in line: continue
    o = json.loads(line)
    if o.get('type') != 'attachment': continue   # injections only
    last = json.dumps(o)
print('cache-resolved' if 'DELETED_FILE_NAME' in last else 'source-resolved')
PY

Filtering on type == "attachment" is what separates a real injection from the session talking about one.

Confirming which hook body ran

Resolution mode can also be established positively: put a side effect in the repo copy of hooks/session-start.sh that the cached copy does not have — a line appending to a log — and check whether it fires. On 2026-08-20 that log had 88 entries while the cached hook contained no such line, which proves the source hook is what executes. A side effect present in only one copy beats any amount of inspecting both.

Steps

1. Probe resolution mode for the peers you care about. It is per-session, not fleet-wide.

2. If any peer is cache-resolved: publish (scrub + squash onto main — this repo is public), bump the version in both plugin/team-lead-fleet/.claude-plugin/plugin.json and .claude-plugin/marketplace.json at the repo root, then claude plugin update team-lead-fleet@team-lead-fleet. The two manifests have fallen out of step before; an unbumped version makes every later step a silent no-op.

   The marketplace manifest is at the repo root, not under plugin/. It moved there when the plugin started being served from GitHub, and this step named the old path for weeks afterwards — a path that does not exist, in the one step whose whole purpose is keeping the two versions in step. Following it literally bumps plugin.json alone, which is precisely the silent no-op the sentence warns about. Corrected 2026-09-17, both manifests confirmed at 0.9.0. Check the path resolves before you edit it; the file has moved once and can move again.

3. Deliver into the session. Rules arrive only through the SessionStart hook, so the peer needs a real SessionStart event: restart, --continue resume, /clear, or /compact. Verified from the hook payload log — it fires on startup, resume, compact, and clear, and nothing else.

4. Re-probe. Report shipped only after a peer's own injected context contains the new text.

What to avoid

• Don't infer current state from a whole-file grep. It counts historical injections and incidental mentions. Read the last injection.
• Don't assume the fleet is uniform. Resolution mode is a per-session property; a peer respawned at a different time can differ from its neighbour.
• Don't treat a respawn as delivery for a cache-resolved peer. It re-reads the same stale cache and comes back looking healthy.
• Don't rely on the drift check alone. It compares repo to cache — which tells you nothing when nobody loads the cache. It also skips github-source plugins entirely.
• Don't put this skill in the plugin. A skill that ships the plugin must not be delivered by the mechanism it exists to fix.