Un groupe lance désormais UNE seule session Claude couvrant tous ses
repos/worktrees (--add-dir) au lieu d'une session par repo, pour
travailler en simultané dans une conversation partagée. Le transport WS,
le flow control et l'attach sont inchangés (toujours 1 session, 1 channel).
- shared: SessionSummary.addedDirs/groupId (champs additifs optionnels) ;
CreateGroupSessionRequest / GroupSessionResponse.
- server: --add-dir dans claude-launcher ; spawn({addDirs,groupId}) +
persistance (migration #8 : sessions.added_dirs/group_id) ;
POST /api/v1/groups/:id/session (résolution des dirs cote serveur) ;
resume re-relie les dirs ; deleteGroup nullifie group_id.
- web: GroupSessionModal (modes feature / checkouts principaux) ;
createGroupFeature/createGroupSession ; badge « groupe · N depots » ;
sessionsInGroup dedupliquee.
- site: textes EN/FR « une seule session Claude » + diagramme hub.
- tests: claude-launcher, pty-manager (groupe), group-manager ;
acceptation P5 etendue (session de groupe multi-repo). 272 tests verts.
Fan-out integration + fixes found by the test/acceptance pass:
- FIX ring-buffer: chunks >= capacity skipped bytes now count into the
monotonic offset (invariant: stream byte k lives at k % capacity) —
window order was corrupted on unaligned big chunks
- FIX auth: non-numeric cookie expiry no longer bypasses expiration
- FIX protocol: safe-integer validation on ack.bytes / hello.protocol
- FIX @fastify/websocket v11: websocket route must be registered in an
encapsulated context after plugin load (handler got REST signature)
- FIX flow-control deadlock found by e2e acceptance: client only ACKs
on data receipt, so pausing with an unACKed residue in (LOW,
ACK_EVERY] stalled both sides at 0.9 MB. ACK_EVERY now 64 KiB (<=
LOW invariant, tested) + trailing debounced ACK in the web client
- Web: Vue 3 + Vite + Pinia + Tailwind 4 + vue-i18n (EN/FR) + xterm 6
(fit + webgl fallback), multiplexed ws-client with reconnect/backoff
and resync epochs
- Tests: 100 vitest (protocol fuzz, ring edges, auth, pty-manager flow
control with mocked pty, REST e2e) ; CI Node 22/24 + pack-smoke
- scripts/acceptance-p1.mjs: real daemon + real WS client — boot,
login, attach, stdin, 10 MB flood w/ ACK (13.7 MB/1.9s, RSS bounded),
brutal disconnect + replay resync, kill broadcast, SIGTERM drain
