Repository navigation
Releases
Releases · not-so-fat/agent-deck
2.5.3
Fix: connecting an OAuth card works on a second host that shares the deck store (NOT-631)
- A card synced from another machine (deck store shared via git) carries an automatically registered OAuth client bound to that machine's redirect URI. On a host with a different redirect URI (for example a cloud host with
AGENT_DECK_PUBLIC_URL), Connect reused that client and the provider (e.g. Linear) rejected the sign-in. Connect now registers a fresh client for the current redirect URI when the stored one was registered for a different address and the provider supports dynamic registration, and clears any secret left by the previous client. Clients you entered yourself, and providers without dynamic registration (e.g. a Slack app), are unchanged.
2.5.2
Change: folder binding is optional — headerless sessions use your default deck (NOT-598, NOT-600 to NOT-606)
- The connection no longer needs a per-folder pin. Set a default deck once with
agent-deck default <deck>; any Cursor, Codex, Claude Code or Muse chat without a pin binds to it. Existing pins keep working and keep precedence. Shipped agent text, skills and docs (NOT-603) now describe this behaviour.
Fix: a deleted or unreadable default deck is reported, not silently treated as no default (NOT-606)
agent-deck defaultwith no argument now says so when the saved default is broken:Default deck: <name> (deck no longer exists, run agent-deck default <deck>)when the saved deck id is no longer in the store, andDefault deck file is unreadable: <path>whendefault-deck.jsonis corrupt or an unknown version. The exit code stays 0, and the file is never auto-cleared or rewritten.- A headerless session that lands in the explain-only unassigned session because no default applied now gets one sentence naming the reason: the missing default deck, or the unreadable
default-deck.jsonpath. Sessions with no default file see the unchanged message. - The shared reader gains
readDefaultDeckState(none/ok/unreadable) so callers can tell "never set" from "set but broken";readDefaultDeckkeeps its null result.
Fix: a folder assignment to a deleted deck falls through to the default deck (NOT-604)
- A local MCP session whose folder assignment (
.agent-deck/use.json) names a deck that no longer exists now resolves as if the folder had no assignment: it binds the user default deck when one is set, instead of landing in the explain-only unassigned session. A launch deck header naming a deleted deck still fails closed, and the assignment file is never rewritten. - With no usable default deck, the session stays explain-only and the
GRANT_REQUIREDreply names the missing deck by its saveddeckName, so the user can see which assignment went stale.
Fix: agent-deck stop never stops another home's daemon (NOT-605)
runStopnow terminates only the supervisor and listeners recorded in the current home'srun.json. The port-listener fallback runs only for listeners whose process environment or/healthidentity reports the sameAGENT_DECK_HOME; otherwise it prints that the daemon on those ports belongs to another home and exits 0 without signalling it. This fixes an isolated-homestopSIGTERMing the production daemon on:1110/:1111.
Change: doctor says which deck source a folder would use (NOT-602)
agent-deck doctorprintsDeck source: <launch header | folder assignment | default deck | none> (<name>)for the current folder, using the same precedence the daemon binds with (launch deck header, then the v3 folder assignment, then the user default deck). A folder with no assignment and no default saysnoneand names theagent-deck default <deck>remedy instead of hiding behindIssues: none.- Run from the home folder, the Cursor inspection no longer prints
~/.cursor/mcp.jsonas both Global and Project: Project printsnone (run doctor inside a project folder), and the headerless global fallback no longer raises a bogusmissing-workspace-headerissue.
Feature: session receipt marks the default-deck fallback (NOT-601)
- When a session is bound to the user default deck because no folder pin resolved,
display_summaryinget_session_contextandget_session_bindingends withdefault deck, no folder pin, so the silent fallback is distinguishable from a pinned folder on the first turn. Folder-assigned, launch-selected and session-override receipts are unchanged, as are the status line and menubar text.
Feature: headerless sessions bind the user default deck (NOT-600)
- A local MCP session that arrives with no deck header and no usable folder assignment now binds the user's default deck (
$AGENT_DECK_HOME/default-deck.json) instead of landing in the explain-only session, so a missing pin or unapproved project server no longer makes Agent Deck unreachable. Precedence is strict: launch deck header, then folder assignment, then default deck; hosted and grant-bound sessions are unchanged. - Set, show, or clear the default with
agent-deck default [<deck>] [--clear], or from the dashboard decks panel (per-deck "Set as default", current-default row with Clear). No default is ever auto-selected. - Such sessions report binding source
default-deckinget_session_contextandget_session_binding. With no default set, or a default naming a deleted deck, behaviour is unchanged (explain-only session,GRANT_REQUIRED).
After upgrade
- Upgrade the CLI, then restart the daemon (
agent-deck start --daemon --no-openafter it stops), runagent-deck default <deck>once, and reload each host. Existing folder pins need no change.
2.5.1
Change: agent-deck upgrade --scan <dir> migrates every assigned folder (NOT-577)
agent-deck upgrade --scan <dir>(repeatable) walks the directory for folders with a.agent-deck/use.jsonand converts their legacymcp-launchentries to pure MCP in one command, including the Cursor project.cursor/mcp.jsonentry.--scan <dir> --checklists the planned changes and writes nothing. The Codex plugin step runs once per invocation. Seedocs/MCP_LAUNCH_MIGRATION.md.
Fix: upgrade no longer lets the Codex plugin step block the MCP migration (NOT-576)
- The MCP config migration now runs and prints before the Codex plugin step.
- Every
codexsubprocess has a timeout (60 seconds by default,AGENT_DECK_CODEX_TIMEOUT_MSto override), so a hungcodexcall cannot stallupgrade. - When
codex plugin addfails after a successfulremove, the output now includes the cache-clearingrm -rfandcodex plugin addrecovery commands. Agent Deck still never deletes the cache itself.
Change: doctor reports the real connection state per host (NOT-574, NOT-586)
agent-deck doctorprints aHost connectionsblock for Codex, Claude Code, Cursor and Muse with one state and a one-line remedy each:no-entry,entry-daemon-down,entry-daemon-up-no-session,connected,folder-not-trusted,folder-not-repository(Muse) orstale-conflicting-entry(Muse). An installed entry alone is no longer reported as OK.- A stopped daemon is reported as
entry-daemon-downfor every host that has an entry, ahead of any trust remedy.
After upgrade
- Upgrade the CLI, then
agent-deck stop && agent-deck start. Runagent-deck upgrade --scan <your projects directory>once to migrate every assigned folder; restart each host afterwards.
2.5.0
Change: Codex, Claude Code, Cursor and Muse connect over plain MCP (NOT-561 to NOT-572)
- The default connection no longer needs an
agent-deckexecutable onPATH. This fixes the Codex desktop plugin failing to start because the app launches its MCP host with a minimalPATH. Each host now talks HTTP to the daemon directly. agent-deck use <deck>writes a per-folder entry that names the folder in anx-agent-deck-workspaceheader:.codex/config.tomlfor Codex,.mcp.jsonfor Claude Code and Muse,.cursor/mcp.jsonfor Cursor. No deck id is written; the server resolves the deck from the folder's.agent-deck/use.json(NOT-562, NOT-563, NOT-567, NOT-571).agent-deck setup --client codex|claude|cursor|musewrites a global headerless HTTP fallback entry. A folder with no assignment lands in the explain-only unassigned session. Agent Deck's own earliermcp-launchentries are replaced; custom entries are never changed (NOT-565, NOT-567, NOT-570).- The Codex, Claude Code and Cursor plugins no longer declare an MCP server; they supply skills and guidance only. The repository-root
.mcp.jsonis removed (NOT-565, NOT-567). - What each host needs once: Codex requires the folder to be trusted in
~/.codex/config.toml(a per-run override is not enough); Claude Code and Cursor ask you to approve the project server; Muse loads project config only inside a trusted git repository. - The Muse launcher path is retired (NOT-572).
agent-deck mcp-launchand the bridge remain for Claude Desktop, which is still launcher-based.
Fix: sessions survive a daemon restart for hosts that never re-initialize (NOT-569)
- Cursor and Muse do not send a new
initializeafter the daemon answers 404 to a staleMcp-Session-Id, so a daemon restart used to end Agent Deck for the rest of that host session. In local mode, a request with an unknown session id whose headers resolve to a deck (workspace header with an assignment, or an authenticated deck header) is now served on a recovered session and keeps working on the same id. Everything else keeps the 404 (no deck-resolving header, unassigned workspace, a session the server itself closed, hosted and grant-bound sessions)./healthcounts these as recovered sessions.
Fix: doctor reports the real Codex connection (NOT-564)
agent-deck doctorreports the Codex host's real connection state and labels the plugin manifest as manifest-only, instead of treating an installed manifest as a working connection.
Docs
- Design notes with the live evidence behind these changes:
docs/design/launcher-free-host-connection.md(Codex and Claude Code, NOT-561) anddocs/design/muse-pure-mcp-connection.md(Muse, NOT-568).
Release tooling
- The packaged-install smoke asserts Muse's headerless HTTP entry and the current setup next steps (it still expected the old launcher entry).
Not changed
- Claude Desktop keeps using
agent-deck mcp-launch.
2.4.0
Change: delegation is MCP-only for agents (NOT-540)
- The
agent-deck-runskill and the generated harness delegation paragraph now delegate exclusively through Agent Deck MCP:get_session_context, optional human-approvedswitch_deck,start_run(brief built in memory, explicit tool profile, resume id, finalization preparation, pinned playbook version),get_runpaged by cursor to a terminal state, andcancel_run. Each failure (missing MCP,WORKSPACE_REQUIRED, provider unavailable, a rejected switch,error/died, cancellation) stops with a short report instead of a CLI remedy; reports carryrunId,sessionId, tool profile, and terminal state with no host log path.agent-deck setuprefreshes the skill and harness for Claude, Codex, Cursor, and Muse. The host CLIagent-deck runstays available for operators (installation, diagnostics, recovery); seedocs/DELEGATION.md.
Feature: MCP run tools (NOT-539)
- Bound MCP sessions get
start_run,get_runandcancel_run: start or resume a host-local run (tool profile, finalization preparation, one pinned playbook version), page its ordered events with a cursor, and cancel it — without the CLI or host loopback access. The deck is the session's live deck (a session-onlyswitch_deckincluded) and the cwd is the workspace the launcher bound the session to; neither is a tool argument, and.agent-deck/use.jsonis not consulted. A session launched without a workspace getsWORKSPACE_REQUIRED. Reference indocs/MCP_RUN_TOOLS.md.
2.3.2
Feature: pinned playbook runs (NOT-537)
- New
agent-deck run --playbook-version <playbookId>=<versionId>flag runs one session against a stored older version of one playbook, so a baseline-vs-candidate evaluation can test each version without touching the live playbook. The run's MCP session reads that version's title, body and triggers with no open patches (marked withpinnedVersionId) and cannot file playbook proposals; the pin lives only as long as the run's session. - The pin travels through each client's explicit per-run MCP server env (a
-c mcp_servers.agent-deckoverride for Codex, an inline SDK MCP server for Claude), since MCP stdio servers never see inherited process env. Cursor and Muse have no such override, so a pinned run with those clients fails closed before any model work instead of silently reading the live playbook. The pin is recorded in the run'smeta.jsonasplaybookPin.
2.3.1
Feature: evaluation scenarios (NOT-535)
- New dashboard-authenticated REST API for evaluation scenarios, manually curated feedback-derived cases for one target playbook:
POST/GET /api/evaluation-scenarios,GET/PUT /api/evaluation-scenarios/:id. A scenario can only becomereadywith task input, a historical failed output, at least one existing feedback signal and one deterministic expectation. Reference indocs/EVALUATION_SCENARIOS_API.md. - The Feedback page can now curate evaluation scenarios from selected signals with a scenario editor. Expectation types match the
agent-deck experiment evaluateevaluator.
2.3.0
Feature: agent-deck experiment evaluate
- New
agent-deck experiment evaluate <manifest.json> [--json]command deterministically compares frozen baseline and candidate outputs against per-case expectations (contains,not_contains,equals,regex) and reports absolute artifact complexity (bytes, words, non-empty lines). - No optimizer or automatic promotion: candidate generation, live runs and human acceptance stay outside the evaluator.
- Example manifests in
docs/examples/experiments/(basic,rsi-feedback-smoke); the smoke example turns two real feedback signals into cases and proves the data path, not that the candidate is better.
2.2.0
Feature: same-host agent run API (NOT-479 to NOT-486)
- One host-local execution service (
createAgentRunServicein@agent-deck/backend) now owns the client adapters, run records and start/resume lifecycle;agent-deck runand the backend share it with unchanged JSONL andmeta.jsonshapes (NOT-479). - Authenticated HTTP API for remote grants that carry the explicit run capability (default off, owner toggle):
POST /api/agent-runs(Idempotency-Key required, bounded body, deck allowlist), resume, status and cursor-ordered normalized events. Failed bearers are rate-limited in hosted mode (NOT-480). GET /api/agent-runs/capabilitieslists only clients whose real host preflight passes; idempotentPOST /api/agent-runs/:runId/cancelsignals the run-owned process group; running records are reconciled after a backend restart. Caller reference indocs/AGENT_RUN_API.md; real Codex smoke recorded (NOT-481).- Optional finalization behind a separate grant capability (default off):
GET /changesreturns the reviewed path/sha256/mode manifest,POST /finalizecreates exactly one local commit onexpectedBaseCommitwith a compare-and-swap ref move (NOT-486).
Codex run adapter (NOT-482, NOT-484)
- Codex runs emit a stable final assistant result.
- Codex resume allows trusted same-profile edit continuation.
2.1.1
Docs / plugin packaging (NOT-432)
- Cursor Marketplace prep:
.cursor-plugin/plugin.jsonand.cursor-plugin/marketplace.json; public skills cleaned before marketplace submission. - Agent plugin now uses a single
skills/source for Claude, Cursor, Codex and Muse, with delegation and setup prerequisites documented in theagent-deck-runandagent-deck-setupskills. - Operator-verified delegation poll interval, round cap and Muse
initializeinstructions indocs/DELEGATION.md. - No runtime behavior change.