Skip to content

CHANGELOG.md

Latest commit

 

History

History
218 lines (171 loc) · 17.1 KB

CHANGELOG.md

File metadata and controls

218 lines (171 loc) · 17.1 KB
Raw

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[0.4.1] - 2026-06-11

Added

  • AWS CodeCommit as a third supported git platform alongside GitHub and GitLab. Auths via AWS access keys or IRSA/instance profile on EKS; PR ops go through @aws-sdk/client-codecommit. CI checks return [] (auto-merge fires on checksStatus="none"); reviewTrigger="on_pr" is recommended over the default on_ci_pass (#529).
  • Ruby and Dart agent image presets — new language-specific base images available in the agent image picker (#470).
  • Live model discovery — agent model lists are now fetched directly from provider APIs at runtime rather than being hard-coded, so newly released models appear automatically (#543).
  • Claude Fable 5 and GitHub Copilot added as supported agent vendors (#542).

Fixed

  • Git token resolution in repo pods is now more resilient against transient credential lookup failures (#525).
  • scope='global' secrets now correctly enforce workspace_id IS NULL in the database, preventing a constraint mismatch that could cause secret lookups to fail (#509).
  • Missing auth.oidc block validation added to the Helm chart, preventing silent misconfiguration when using generic OIDC login (#528).

[0.4.0] - 2026-04-27

Added

  • Persistent Agents — a third Task tier alongside Repo Tasks and Standalone Tasks. Long-lived, named, message-driven agents that wake on user messages, agent messages, webhooks, cron ticks, or ticket events. Each agent has a stable slug, addressable by other agents in the same workspace via an inter-agent HTTP API. Three configurable pod lifecycle modes: always-on, sticky (default, with idle warm window), and on-demand. Cyclic state machine reconciled by the existing K8s-style control plane (now a fourth RunKind: persistent-agent). New /agents UI with chat, turn history, live activity stream, and pause/resume/restart/archive controls (#510). See docs/persistent-agents.md and the four-agent demo in demos/the-forge.
  • Issues as a top-level nav item — the GitHub Issues queue is now its own page at /issues, fanning out across multiple configured ticket providers.
  • Reviews as a top-level nav item — code-review subtasks plus external PR reviews now live at /reviews (and /reviews/:id), with their own reconciler RunKind (pr-review).
  • Workspace member management UI — invite, list, and remove workspace members by email with a server-side duplicate-member 409 guard (#496, #499).
  • Agent-aware review configuration — pick the review agent type per repo so reviews can run on a different vendor than the authoring agent (#504).
  • Skills marketplace — install skills from any git URL, with agent-typed scoping and a multi-file skill-directory layout so a single skill can target one or many agent types.
  • Persistent Agents and Sessions stats bars on the overview page (#521).
  • Persisted session chat history — re-opening a session now shows the prior conversation (#517).
  • Rich rendering in the agent log viewer — markdown tables and rich blocks render in agent text output (#520).
  • Examples directory — runnable, self-contained agent configurations under examples/, starting with two Persistent Agent setups (Forge and Mars Mission Control). Each example is idempotent — re-running setup.sh is safe.

Changed

  • Sidebar nav reorganized. The hub-and-tabs /tasks page is gone. Each tier has its own dedicated route, grouped into Run (Tasks · Jobs · Reviews · Issues · Scheduled) and Live (Agents · Sessions). The Library group renamed "Templates" to Prompts to free that label. Legacy /tasks?tab=… URLs redirect to the dedicated pages.
  • User-facing names finalised. Repo Tasks → Tasks; Standalone Tasks → Jobs (matching the existing /api/jobs URL); PR Reviews → Reviews; Persistent Agents → Agents; Templates (in the Library) → Prompts. Backend table names (tasks, task_configs, workflows, prompt_templates) are unchanged.
  • Setup wizard defaults agent secrets to global scope with an explicit scope toggle, reducing per-user secret sprawl for the common case (#503).
  • README leads with differentiation — self-hosted, BYO-K8s, multi-vendor — to better orient first-time readers (#519).

Fixed

  • Sessions: stop echoing the agent's final line in the chat view (#516).
  • Auth: /me now returns the enriched user object with the caller's workspace role attached (#508).

Migration notes

Upgrading from 0.3.x — there are no required user actions. URL/nav changes:

  • /tasks?tab=standalone → /jobs (auto-redirected)
  • /tasks?tab=issues → /issues (auto-redirected)
  • /tasks?tab=prs → /reviews (auto-redirected)
  • The "Templates" sidebar item is now labeled Prompts but still points to /templates.
  • The new /agents route requires the v0.4 schema migration (1777200001_persistent_agents.sql) — applied automatically on API startup.

No data migration is needed. Existing tasks, workflows, triggers, and templates continue to work unchanged.

[0.3.2] - 2026-04-24

Added

  • External PR auto-review — review agent for PRs on external repos with chat + one-click merge, lifted into its own primitive alongside task-generated reviews.
  • Google Vertex AI authentication mode for Claude Code — route Claude through GCP Vertex AI using CLAUDE_VERTEX_PROJECT_ID / CLAUDE_VERTEX_REGION and an optional (encrypted, global-scope) service account key, with workload-identity fallback (#478).
  • Workload identity support for agent pods, plus fixes to repo pod lifecycle (#486).
  • User-scoped secrets — keep identity tokens out of the pod env and scope them per user (#474).
  • Secrets injected into pod env for setup commands (#471), and an OAuth refresh widget on /secrets that hides the banner when visible.
  • Resume stopped agents on chat message — sending a chat message to a stopped task resumes the agent (#488).
  • Multi-repo + multi-tracker ticket integration — redesigned setup flow (#489).
  • Dynamic per-provider model & options picker with a refresh button for agent settings (#493).
  • Gemini model options updated with new preview models (#490).
  • GKE & Gateway deployment enhancements in the Helm chart (#461).
  • Diagnostic logging for raw error detection in agent adapters (#467).

Changed

  • PR reviews folded into the Tasks page — removed the sidebar duplicate; task and PR-review detail views now share primitives (#494, #485, f1a6da4).
  • Repo settings page — split external PR review out and tabified agent settings (#487).
  • Standalone Tasks pipeline stats bar restored on the overview page.
  • Opus model option bumped from 4.6 to 4.7 (#491).

Fixed

  • Reconciler: guard PR-reactive actions (auto-merge, complete-on-merge, review launch) to coding tasks only so external PR reviews don't trip them (#480).
  • Reviews: stop writing external PR URLs to pr_review task rows (#481).
  • Secrets: downgrade scope='user' to 'global' when auth is disabled.
  • API: derive Claude/Codex/Gemini mode from secret names on public /setup/status (#477).
  • Auth: add OIDC routes to public auth routes so login works before a session exists (#479).
  • Helm: restore chown capabilities in postgres init containers (#482); fix postgres volume permissions and decouple isSetUp from runtime health (#472).
  • Images: change agent user UID from 1000 to 1001 to avoid conflicts on managed node images (#466).
  • Gemini agent: settings validation, parser crash, and exit-code inference (#463).
  • Correct sub-hour timezone drift in getETDate (#462).

[0.3.1] - 2026-04-20

Fixed

  • Ticket sync: fall back to the configured GitHub App (or GITHUB_TOKEN PAT) when a GitHub ticket provider has no inline token or provider-specific secret. Previously sync hard-failed with "GitHub provider requires token, owner, and repo in config" even when a GitHub App was fully configured (#458).

[0.3.0] - 2026-04-20

Added

  • Pooled standalone-task pods — runs within a workflow now share pods, scaling out to workflows.maxPodInstances replicas each hosting up to workflows.maxAgentsPerPod concurrent runs (mirrors repo pod scaling). Runs track assigned pods via workflow_runs.pod_id with last_pod_id for retry affinity, and pool selection follows preferred → least-loaded → scale-up → overflow. Fixes a leak where a burst of triggers would spawn one pod per run even though only a few ran at once.

Changed

  • Reconciliation control plane is now authoritative — the K8s-style reconciler (shadow mode in 0.2.0) now owns PR-driven transitions, auto-merge, complete-on-merge, fail-on-close, auto-resume, review launch, stall detection, pod-death detection, and control intent (cancel/retry/resume/restart) for both Repo Tasks and Standalone Tasks.
  • Shared auth banner, state badge, and metadata card across task pages for a consistent UX.

Fixed

  • Reconciler: clear stale finishedAt when retrying a standalone run.
  • Reconciler: use unique jobIds for executor enqueues to prevent BullMQ dedup collisions.
  • Agent adapters: include cache_read and cache_creation tokens in input totals (#457).
  • API: trigger auth banner when the usage endpoint detects an expired OAuth token (#455).
  • API: detect Claude auth failures mid-run in standalone task runs and override nominally-successful exit codes.

Docs

  • Document the unified reconciler and the Repo vs Standalone Task model.

[0.2.0] - 2026-04-17

Added

  • Unified Task model — single polymorphic /api/tasks HTTP resource covering Repo Tasks, Repo Task blueprints, and Standalone Tasks; unified resolver across tasks, task_configs, and workflows
  • Standalone Tasks (Agent Workflows) — agent runs with no repo checkout, {{PARAM}} prompt templates, four trigger types (manual / schedule / webhook / ticket), isolated pod execution, WebSocket log streaming, auto-retry with exponential backoff, clone, visual editors, search and filters
  • Connections — external service integrations via MCP with built-in providers (Notion, GitHub, Slack, Linear, PostgreSQL, Sentry, Filesystem) plus custom MCP servers and HTTP APIs; three-layer model of providers → connections → per-repo/agent-type assignments
  • Reconciliation control plane (shadow mode) — K8s-style reconciler for task and pod state, running in observe-only mode
  • StatefulSets for repo pods, Jobs for workflow pods — native K8s controllers replace ad-hoc pod management
  • Generic OIDC OAuth provider — self-hosted SSO via OIDC_ISSUER_URL + OIDC_CLIENT_ID + OIDC_CLIENT_SECRET
  • OpenTelemetry instrumentation — Fastify HTTP metrics plugin and wired-up callsites
  • OpenAPI + Swagger UI at /docs — Zod type-provider migration across all routes (10-phase rollout covering tasks, workflows, repos, sessions, PR reviews, issues, workspaces, notifications, analytics, setup, secrets, optio, cluster, auth, GitHub)
  • Workspace-level audit log and activity feed
  • Outbound webhooks — fire on workflow run events with UI management
  • Expanded dashboard analytics — performance, agents, and failure insights
  • Planning mode and message bar improvements for agent interaction
  • OpenClaw agent runtime adapter
  • OpenCode custom OpenAI-compatible endpoints
  • Multi-arch image publishing — amd64 + arm64 for all service and agent images
  • Ticket trigger UI in TriggerSelector and task forms
  • Ticket-provider auth failure handling — surfaced in UI with auto-disable
  • Stale Claude OAuth token detection — surface before 401s
  • nodeSelector and tolerations for api, web, optio, postgres, redis, and agent pods
  • OPTIO_ALLOW_PRIVATE_URLS — SSRF-check bypass for private network integrations

Changed

  • Overview panel redesign — reordered sections, side-by-side recent tasks and pods, responsive multi-column / masonry grid with auto-fit minmax
  • Replaced connections modal with inline form
  • Renamed "Workflows" to "Agent Workflows" in UI; docs consolidate Schedules + Workflows into a unified Tasks section
  • Removed redundant templates and schedules — superseded by agent workflows
  • Workflow tables replaced with new Workflows data model

Removed

  • Top Failures and Performance dashboard panels
  • "N tasks failed today" dashboard banner

Fixed

  • Classify agent auth failures as run failures rather than global failures
  • Escalate repo tasks to needs_attention when the agent completes without opening a PR
  • Prevent false task failures when agent creates a PR but exits non-zero
  • Detect and clean up zombie workflow_runs with terminated pods
  • Six K8s infra bugs blocking standalone/scheduled runs and repo pods
  • Pod securityContext and explicit UID for PVC permissions on GKE
  • Re-read task state before orphan reconciliation transitions
  • Use KubernetesObjectApi for merge-patch annotations; fix scale API
  • Persist workflow run logs and publish to per-run channel
  • Allow access to workflows with null workspaceId
  • Treat empty-string env vars as missing in parseInt parsing
  • JSON.parse error handling for agent scheduling env vars
  • Health check passes when ClusterRole is not deployed
  • Record GitHub 401s to auth_events for banner detection
  • Dismiss GitHub/Claude token banners immediately after save
  • Clear stale auth-failure banner when token is updated
  • Scope auth failure detection to distinguish provider vs global token failures
  • Replace Drizzle migrate() with hash-based runner; add missing 0046 migration entry to Drizzle journal
  • Merge new chart defaults on update-local upgrade
  • Rename /docs/guides/workflows route to /docs/guides/standalone-tasks

[0.1.0] - 2026-03-24

Added

  • Pod-per-repo architecture — long-lived Kubernetes pods with git worktrees for concurrent task execution per repository
  • Task orchestration — full task lifecycle with state machine (pending, queued, provisioning, running, pr_opened, completed, failed, cancelled, needs_attention)
  • Priority queue with concurrency limits — global and per-repo concurrency controls, priority-based scheduling, and task reordering
  • Subtask system — child, step, and review subtask types with parent blocking and completion tracking
  • Code review agent — automatic PR review as a blocking subtask, configurable triggers (on CI pass or PR open), and dedicated review prompts
  • PR watcher — polls GitHub PRs for CI status, review status, merge/close events; auto-completes on merge, auto-fails on close
  • Auto-resume on review — re-queues tasks with reviewer comments when changes are requested
  • Auto-resume on CI failure and merge conflicts — detects failures and re-queues the agent to fix them
  • Auto-merge — merges PRs automatically when CI passes and reviews are approved
  • Auto-close linked GitHub issues — closes the originating GitHub issue when a task completes
  • GitHub Issues integration — browse issues across repos, one-click assign to create tasks, bulk assign all
  • Linear ticket provider — sync tasks from Linear projects
  • Structured log streaming — real-time NDJSON parsing of Claude Code output with typed log entries (text, tool_use, tool_result, thinking, system, error, info)
  • WebSocket event streaming — live task state and log updates pushed to the web UI via Redis pub/sub
  • Web UI — Next.js 15 app with task list, task detail with log viewer, repo management, cluster health, secrets management, and setup wizard
  • Cluster health dashboard — expandable resource usage graphs, pod health monitoring, and stale task detection
  • Pod health monitoring — automatic detection of crashed/OOM-killed pods, auto-restart, orphan worktree cleanup, and idle pod cleanup
  • Secrets management — AES-256-GCM encrypted secrets with global and repo-scoped support
  • Prompt templates — configurable system prompts with template variables and conditional blocks; per-repo overrides
  • Per-repo agent settings — configurable Claude model, context window, thinking mode, effort level, and max turns
  • Auto-detect image preset — detects project language (Node, Python, Go, Rust) from repo files and selects the appropriate container image
  • Agent adapters — pluggable adapter interface with Claude Code and OpenAI Codex implementations
  • Container runtimes — Docker and Kubernetes runtime backends
  • Authentication — API key and Max Subscription (OAuth) modes for Claude Code
  • Error classification — pattern-matching error classifier with human-readable titles, descriptions, and remediation suggestions
  • Helm chart — full Kubernetes deployment with configurable Postgres, Redis, ingress, RBAC, and secrets
  • Pre-commit hooks — Husky with lint-staged, Prettier formatting, ESLint, typecheck, and conventional commit enforcement
  • CI pipeline — GitHub Actions for format checking, typechecking, testing, web build, and Docker image build