Panneau Permissions: idea_ask_agents est câblé côté backend MCP mais absent
du groupe « Délégation agents » du panneau — ajout de l'outil et de son
label, catalogue mock MCP aligné en conséquence.
Terminaux: ProjectsView remonte LayoutGrid via une key au changement de
projet, donc refitEpoch n'aide pas ; le montage de TerminalView ne
maintenait pas assez longtemps le fit tant que la nouvelle cellule ne
s'était pas stabilisée. Le chemin de montage garde désormais une traîne de
refits plus longue (MOUNT_SETTLE_REFIT_FRAMES) avant de laisser la main au
ResizeObserver.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds the "Tools MCP IdeA" tab to the project Permissions panel, alongside
the existing "Système" (file/command) tab. Lets the user grant/revoke MCP
tool capabilities per agent or project-wide default, grouped by domain
(Lecture projet, Lecture tickets, Délégation agents, Contexte et mémoire,
Tickets, Travail et exécution, Skills) instead of a flat 25-checkbox list,
per the UX conception in carnet #82.
- domain/ports/adapters (Tauri, HTTP, mock): wire get_mcp_tool_permissions,
update_project_mcp_tool_permissions, update_agent_mcp_tool_permissions
(already merged backend API, #82 lots B1-B4) onto PermissionGateway.
- useMcpToolPermissions: view-model owning the durable MCP tool policy
document, distinct from the file/command permissions in usePermissions.
- McpToolPermissionsPanel: target selector (Défaut projet + agents with
Hérité/Override badges) and grouped editor — inherited agents are
read-only until "Créer un override" (prefilled with the effective
allowlist), per-row Ajouté/Retiré diffing against the project default,
inline confirmation before granting a write tool at project-default
level, and unsaved-draft protection on target change.
- mcpToolGroups.ts: presentational-only domain grouping and short French
labels — the read/write classification itself always comes from the
backend-provided catalogue, never hardcoded here.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The draft-resync useEffect fired on every new `draft` object (initial load,
refresh, another save), even when its content matched `local`. If it flushed
after the user started editing, it silently discarded the edit — flaky in
tests where a passive effect can settle after a synchronous fireEvent
sequence, and a real risk in production if a fetch lands mid-edit. Now it
only resyncs when `local` still matches the last-adopted draft.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
