You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Add: Codex host setup (--client codex)
agent-deck setup --client codex installs or refreshes Agent Deck bootstrap guidance in global or project AGENTS.md, marker-merging it without replacing unrelated instructions. Codex MCP transport still comes from the separately installed Agent Deck plugin (agent-deck mcp-launch) — Codex setup merges AGENTS.md only.
--scope project is now accepted for codex alongside cursor and claude.
Changed: Claude Code registers the trusted stdio bridge
Claude Code setup now registers agent-deck mcp-launch over stdio instead of a bare HTTP URL, carrying custom host/port values in the entry's environment. A bare HTTP entry cannot select a deck.
agent-deck debug-mcp probes with the folder's deck/workspace headers, reports a missing folder assignment explicitly (and skips the authenticated probe) instead of misdiagnosing the expected GRANT_REQUIRED response as a daemon failure, and reports project/user launcher conflicts symmetrically.
debug-mcp now exits 1 when a Claude entry is still a bare HTTP URL instead of the mcp-launch stdio bridge. Recheck anything scripted on its exit code.
Fix: fail-closed Agent Deck session bootstrap
Generated Cursor and Claude harnesses now require get_session_binding → get_bound_deck before task work whenever Agent Deck MCP is configured, including launch-selected unattended sessions that intentionally have no .agent-deck/use.json.
Project-scoped harnesses no longer emit the obsolete bind_workspace flow. agent-deck use <deck> is documented as the optional persistent folder-assignment path, not the source of connection authority.
The session skill permits only Agent Deck configuration detection and read-only connection diagnostics before the bootstrap gate passes.
Connection recovery now distinguishes host transport, folder assignment, instruction discovery, and session bootstrap.
Fix: deck membership outside the deck routes reaches the store files (NOT-121)
Accepting a kind: create playbook proposal now writes the new playbook id into decks/<id>.json. Previously the .md landed but the deck link lived only in SQLite, so agent-deck reindex (and every git-sync pull) dropped the playbook from the deck.
Same flush for every other non-route caller: credential/service/playbook deck links, bundle import, and CLI service delete. Deleting a card now also removes it from the deck files it was on — a deck naming a missing card used to abort the next reindex.
A deck file that cannot be written during a card delete no longer strands the card's own file: the card is still removed from the store before the write error surfaces, so a failed delete can't be undone by the next reindex.
No change to reindex semantics: files still win.
After upgrade
Upgrade the CLI, then agent-deck stop && agent-deck start.
Re-run agent-deck setup --client <host> for each host so the regenerated harness carries the fail-closed bootstrap gate.
Claude Code: migrate an older user-scope HTTP entry — claude mcp remove agent-deck -s user && agent-deck setup --client claude.
Codex: verify the Agent Deck plugin is installed and enabled, run agent-deck setup --client codex, then start a new Codex task so MCP + AGENTS.md guidance reload.
Reload IDE MCP hosts so connections send the deck header.