-
Notifications
You must be signed in to change notification settings - Fork 0
Releases: Wintersta7e/agentdeck
Releases · Wintersta7e/agentdeck
v6.10.0 — productivity + plan limits, persisted session history
AgentDeck 6.10.0 replaces the unreliable cost tracking with honest productivity + plan-limits metrics, and makes session history survive restarts.
Productivity tracking (replaces cost)
- Disk-backed daily rollups survive app restart — no more "today resets on relaunch".
- Home KPIs now show active time, files changed, and sessions (7-day spark + per-project), in place of the old cost/token readouts.
- Productivity is recorded from the real session lifecycle in the main process — no API price tables, no log-tailing.
Plan limits
- Real Codex 5h + weekly rate-limit gauges, read account-wide from the rollout log (reset-aware).
- A rolling-5h activity tile for every agent, built honestly from productivity data.
Session history persists
- Timeline + HISTORY now survive close/restart — recording moved renderer → main on PTY spawn/exit.
- Fixes the "session not found" race: user-killed sessions are counted, and records left open by an unclean shutdown are finalized on next launch.
Housekeeping
- Test coverage scoped to testable logic (excludes render + bootstrap).
- Dependency bumps: zustand 5.0.14, lucide-react 1.17.0, lint-staged 17.0.7.
Assets 3
v6.8.0 — reliable WSL agent execution + per-node permissions, 7 workflow blueprints, loop-escape & a broad stability pass
Everything since 6.6.0 (6.7.x was never published). Headline: workflow agents now run reliably in WSL from the packaged app. Plus per‐node agent permissions, seven complex workflow blueprints, a graceful loop‐escape engine, a broad error‐surfacing and stability pass, and a dependency refresh.
Added
- Per‐node agent permissions — Read / Edit / Full on each agent node, mapped to per‐agent flags (codex
--sandbox, claude--permission-mode), with a node‐editor dropdown and seed annotations. - Seven complex workflow blueprints seeded out of the box: Autonomous Bug Fix, Feature Pipeline, Parallel Deep Review, Coverage Loop, Refactor Campaign, Release Readiness, and Cross‐Agent Design Verify.
- Adversarial Reviewer built‐in workflow role.
- Graceful loop escape — exhausted condition loops route to an escape edge instead of silently completing.
- Edit project + Pin actions in the project right‐click menu.
- App version is logged at startup.
Changed
- Agent‐node timeouts retuned: dropped over‐tight per‐node caps from the seed workflows and adopted a 20‐minute idle window, so agents that buffer all their output until done aren't killed mid‐run.
~/.local/binis preferred on PATH so a native CLI install wins over a stale npm‐global copy shadowed by a node version manager.- Workflow scheduler hardened: dedup condition resolution, restore shared loop‐exit targets on reset, and ignore stale ready‐queue entries after a loop reset.
- Dependency refresh: Electron 42, Vitest, TypeScript, type defs, lucide‐react, lint‐staged; transitive advisories (tmp, brace‐expansion) patched; dependabot set to ignore blocked majors.
Fixed
- Agent CLIs now run reliably in WSL. Prompts go over the child's stdin instead of the command line (shell‐significant characters can no longer break the WSL command);
HOME/NVM_DIRare reset to the real Linux home (they could inherit the Windows profile from the launching process); and the agent +nodedirectories are resolved from an interactive login shell and injected into PATH — fixing nvm‐installed agents (e.g. codex) not being found (exit 127). - codex agent nodes run without inheriting interactive hooks; claude agents change directory via
cdrather than an invalid--directoryflag. - Windows folder‐picker / project paths are normalised to WSL paths on save, on startup, and from the picker;
$HOMEresolution is retried after transient WSL failures. - Agent updater only flags an update when the latest version is strictly newer, retries the post‐install binary check, parallelises its failure diagnostics, and uses the refreshed goose installer URL.
- No more silent failures — surface template parse errors, workflow save/unmount failures, invalid‐regex conditions, PTY spawn failures, and previously‐swallowed background errors; distinguish ENOENT from real fs errors.
- Cost history writes are atomic and preserve a corrupt file as
.bad. - Stability: LRU caps on per‐project template watchers, auto‐cancel of stale confirm prompts, xterm disposal on close/eviction, capped cost‐tracker discovery polling,
crypto.randomUUIDfor starter node ids, correct session agent labels for migrated projects, and edge deletion in the workflow editor.
Security
- Workflow import downgrades a
fullagent permission level toedit.