@desplega.ai/agent-swarm 1.20.0 → 1.49.0

package/CHANGELOG.md

@@ -0,0 +1,447 @@
+# Changelog
+
+All notable changes to Agent Swarm are documented in this file.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
+
+## [1.49.0] - 2026-03-21
+
+### Added
+- `agent-swarm onboard` CLI wizard — interactive first-time setup that collects credentials, generates `docker-compose.yml` + `.env`, starts the stack, and verifies health (#206)
+  - Presets: `dev`, `content`, `research`, `solo`
+  - Progress indicator, `ANTHROPIC_API_KEY` support, Ctrl+C handling
+  - Inline validation errors for integration steps (GitHub, GitLab, Sentry, Slack)
+- `agent-swarm docs` command — show documentation URL with `--open` flag to launch in browser
+- `agent-swarm claude` command — run Claude CLI with optional message and headless mode
+- Workflow structured output support — agent-task nodes can define `config.outputSchema` for validated JSON responses (#207)
+  - `store-progress` validates agent output against schema inline
+  - Workspace scoping for agent-task executor via `vcsRepo`
+- Workflow I/O schemas with explicit input mappings and data flow validation (#201)
+- Fumadocs LLMs and OpenAPI integrations for docs site (#205)
+
+### Changed
+- CLI command renames: `setup` → `connect`, `mcp` → `api` (#206)
+- `api` command gains `--db` flag for custom database file path
+- CLI help rewritten as plain `console.log` with per-command `--help` support
+- `connect` command auto-reads `API_KEY` from `.env`, uses random port, supports `APP_URL`
+
+### Fixed
+- Workflow validation: clear `nextRetryAt` when retries are exhausted (#207)
+- Workflow validation: re-run validation after retry poller re-executes a step (#207)
+- Workflow validation: normalize pass/fail across all executor types (#207)
+
+## [1.48.0] - 2026-03-20
+
+### Added
+- Workflow I/O schemas with explicit input mappings and data flow validation (#201)
+  - Node-level `inputs` mapping for cross-node data flow
+  - Static data flow validation for input references
+  - `triggerSchema` for validating trigger payloads
+- Fumadocs LLMs and OpenAPI integrations for docs site (#205)
+  - API Reference pages auto-generated from OpenAPI spec
+  - Project selector for Documentation vs API Reference
+  - `.md` extension support for LLM-friendly content
+- CI merge gate for generated API docs drift detection
+- SEO: automated inbound links to new documentation pages
+
+### Changed
+- API reference consolidated to single page with tag-based subsections
+- Docs site sidebar navigation improved with API Reference visibility
+
+### Fixed
+- Docs site project selector visibility on all pages
+
+## [1.47.0] - 2026-03-20
+
+### Added
+- Linear integration — bidirectional ticket tracker sync via OAuth + webhooks (#161)
+  - OAuth 2.0 authorization flow with PKCE
+  - Webhook handler for issue/comment events
+  - `AgentSession` lifecycle tracking for Linear issues
+  - Generic tracker abstraction layer (`tracker_sync` table) for future integrations
+  - `.env.example` updated with Linear setup instructions
+- Workflow engine redesign — DAG-based workflow automation with improved reliability (#196)
+  - Executor registry architecture for extensible step types
+  - Node I/O schemas with explicit input mappings and validation
+  - Workflow-level `triggerSchema` validation
+  - Static data flow validation for input mappings
+  - Convergence deadlock fix with active edge tracking
+  - Interpolation rewrite with unresolved variable tracking and deep config support
+  - Slack notification executor for workflow steps
+- Portless integration for local development — friendly URLs like `api.swarm.localhost:1355` (#200)
+  - `dev:http` script uses portless by default
+  - New `start:portless` script for production-like local runs
+  - `.env.example` updated with portless configuration instructions
+- `agent-fs` Claude plugin pre-installed in worker containers
+
+### Changed
+- Claude Code version pinned in Dockerfile.worker via `CLAUDE_CODE_VERSION` build arg (default: `2.1.80`) — replaces dynamic installer for reproducible builds (#202)
+- Runner prompt generation is now provider-aware for pi skill prefix
+
+### Fixed
+- Corepack permissions — `COREPACK_HOME` redirected to user-writable directory to avoid "operation rejected by your operating system" errors (#202)
+- `task.cancelled` outbound handler added for proper cancellation event propagation
+- Follow-up tasks properly repoint `tracker_sync` for session lifecycle
+- Read user message from `agentActivity` with proper stop signal handling
+- Avoid duplicate responses — prefer `AgentSession` over issue comments
+- [UI] Use node ID as graph label, remove schema sections from workflow inspector
+
+## [1.45.1] - 2026-03-19
+
+### Added
+- Debug tab with database explorer — SQL query interface in the dashboard with Monaco editor, table browser sidebar, and AG Grid results display
+- `db-query` MCP tool — lead-only read-only SQL queries against the swarm database (capped at 100 rows)
+- `POST /api/db-query` REST endpoint for database inspection
+- Agent-fs native integration — persistent, searchable filesystem shared across the swarm
+  - Auto-registration on first container boot (idempotent)
+  - Lead creates shared org, workers receive invitations automatically
+  - System prompt conditionally includes agent-fs CLI usage instructions
+  - `agent-fs` CLI and Claude plugin pre-installed in worker containers
+
+### Changed
+- Per-session MCP config — each Claude session gets its own `/tmp/mcp-{taskId}.json` config file instead of sharing `.mcp.json`, eliminating race conditions with concurrent sessions (#192)
+- `--strict-mcp-config` flag ensures only per-session MCP servers are loaded (#192)
+- Removed time-based `getAgentCurrentTask()` fallback — uses deterministic `sourceTaskId` only
+- Slack metadata is now auto-inherited from the creator's current task via `X-Source-Task-Id` header — explicit `slackChannelId`/`slackThreadTs`/`slackUserId` params on `send-task` remain available as optional overrides (#191)
+
+### Fixed
+- Concurrency safety for Slack metadata auto-inheritance — pass `sourceTaskId` through MCP session context via `X-Source-Task-Id` header instead of guessing current task (#191)
+- `send-task` now propagates `sourceTaskId` for accurate Slack metadata lookup
+
+## [Unreleased]
+
+### Added
+- Multi-API-config UI for dashboard — connect to multiple swarm instances from a single browser (#189)
+  - Slug-based connection data layer with localStorage persistence (Phase 1)
+  - React context for multi-connection state management (Phase 2)
+  - Sidebar swarm switcher and header connection name display (Phase 3)
+  - Config page multi-connection management with URL param modal (Phase 4)
+  - Health indicator dots in swarm switcher (Phase 5)
+
+## [1.44.5] - 2026-03-17
+
+### Added
+- OpenAPI 3.1 spec at `/openapi.json` (~83KB, ~60 REST endpoints) generated from route registry (#184)
+- Scalar interactive API docs at `/docs` — pre-authentication API explorer (#184)
+- `MODEL_OVERRIDE` and `CAPABILITIES` env vars for content agents in `docker-compose.example.yml` (#165)
+  - `content-writer`: `MODEL_OVERRIDE=opus`, capability: `content-writing`
+  - `content-reviewer`: `MODEL_OVERRIDE=sonnet`, capability: `content-review` (uses Gemini via OpenRouter)
+  - `content-strategist`: `MODEL_OVERRIDE=sonnet`, capability: `content-strategy`
+
+### Changed
+- `route()` factory replaces all raw `matchRoute()` calls — typed route definitions with Zod schemas for params, query, and body validation (#184)
+- Lead agent now posts task results back to originating Slack threads (#183)
+- Worker agents now post start/completion/failure updates to originating Slack threads (#183)
+
+### Fixed
+- Slack thread follow-ups route to lead when assigned agent is offline (#183)
+- `parentTaskId` continuity preserved for follow-up tasks (#183)
+- ARM compatibility for Docker Compose — added `platform: linux/amd64` to all services to fix `no matching manifest for linux/arm64/v8` on Apple Silicon Macs (#180)
+
+### Added
+- Rich Block Kit messages for all Slack responses — structured headers, context, sections, and action buttons (#177)
+- Single evolving message per task — assignment, progress, and completion all update one message via `chat.update` (#177)
+- Slack Assistant sidebar support with thread routing, suggested prompts, and typing status (#177)
+- Interactive actions: follow-up modal for sending follow-up tasks, cancel with confirmation dialog (#177)
+- Markdown-to-Slack format converter (`markdownToSlack`) for consistent formatting (#177)
+- Per-agent write isolation on shared disk (#172)
+  - Each agent can only write to its own subdirectory under `/workspace/shared/{category}/{agentId}/`
+  - PreToolUse hook warns agents before writing to another agent's directory
+  - PostToolUse hook detects "Read-only file system" errors and guides agents to use their own directory
+  - Base prompt updated with per-agent directory convention and discovery commands
+  - Slack download tool saves to per-agent download directory by default
+- Claude credential validation — fail fast if no auth is set
+- Pre-push hooks to match CI merge gate checks
+- Working directory (`dir`) support for agent tasks (#159)
+  - `send-task` and `task-action` accept `dir` parameter (absolute path) to set agent starting directory
+  - Runner resolves `dir` for both new and resumed tasks with fallback chain: `task.dir` > `vcsRepo` clone path > default cwd
+  - System prompt annotated with working directory context when non-default
+- Content agent templates: writer, reviewer, strategist (#160, #162)
+  - 3 new official templates: `official/content-writer`, `official/content-reviewer`, `official/content-strategist`
+  - Docker-compose examples for all 3 content agents
+  - Content reviewer configured with Gemini via OpenRouter (`HARNESS_PROVIDER=pi`)
+- Template defaults applied during worker registration (#159)
+  - Templates can now set `name`, `role`, `capabilities`, `maxTasks`, and `isLead` as fallback defaults
+  - Template fetched before registration so defaults apply to the registration call itself
+- Archil FUSE mount support for persistent workspace storage (#166, #168, #169)
+  - `archil` CLI installed in both API and worker Docker images
+  - FUSE3 and libfuse2 packages added to Docker images
+  - Entrypoint-based mount logic for R2-backed persistent disks
+  - Removed `VOLUME` directives for `/workspace/shared` and `/workspace/personal` to allow FUSE mounts
+- Contribution guidelines (CONTRIBUTING.md) with templates linked in docs and landing page (#158)
+- Templates registry for agent workers (#155, #156)
+  - 6 official templates: lead, coder, researcher, reviewer, tester, forward-deployed-engineer
+  - Templates UI (Next.js) with gallery, detail pages, and interactive docker-compose builder
+  - `TEMPLATE_ID` env var for initial profile fetching on first boot (e.g., `official/coder`)
+  - `TEMPLATE_REGISTRY_URL` env var for custom registry endpoints
+  - Template idempotency: existing profile fields are never overwritten
+  - GitHub issue/PR templates for community template submissions
+- GitLab integration with Provider Adapter Pattern (#153)
+  - `POST /api/gitlab/webhook` route with timing-safe secret verification
+  - Handlers for merge_request, issue, note (comments), and pipeline events
+  - Bot mention detection via `GITLAB_BOT_NAME` env var
+  - GitLab trigger events for workflow engine (`gitlab.merge_request.*`, `gitlab.issue.*`, etc.)
+  - `glab` CLI installed in worker Docker image
+  - VCS provider detection for automatic `gh`/`glab`/`git` clone selection
+  - New env vars: `GITLAB_TOKEN`, `GITLAB_URL`, `GITLAB_WEBHOOK_SECRET`, `GITLAB_BOT_NAME`, `GITLAB_EMAIL`, `GITLAB_NAME`
+- ProviderAdapter abstraction with pi-mono support (#151)
+  - `ProviderAdapter` interface decouples the runner from Claude CLI
+  - `ClaudeAdapter` extracted from monolithic runner (~600 lines)
+  - `PiMonoAdapter` with MCP tool discovery, event normalization, and cost tracking
+  - All 6 swarm hook events mapped to pi-mono extension handlers
+  - Selected via `HARNESS_PROVIDER=claude|pi` env var
+  - Docker multi-provider support in Dockerfile.worker and entrypoint
+
+### Changed
+- API data disk switched from Archil FUSE to Fly volume for reliability
+- Shared disk uses exclusive Archil mounts with `--force` for stale delegation recovery
+- Template fetching refactored to run before agent registration (cached and reused for identity files)
+- Docker workspace volumes replaced with FUSE mount points for Archil compatibility
+
+### Fixed
+- Thread follow-ups now route correctly after task completion — `getAgentWorkingOnThread` checks all statuses (#177)
+- Docker entrypoint runs as root for FUSE mounts, then drops to worker user via `gosu` before exec
+- Archil FUSE mount fixes: read-write mounts, per-agent subdirectory checkout, POSIX signal names in entrypoint, shared flag for mount calls
+- `dir` validation added to MCP tool schemas with inner type cast fix
+- Workspace `mkdir` made non-fatal for read-only Archil mounts
+- VOLUME directives removed from Dockerfile.worker to unblock FUSE mounts on Fly.io
+
+### Changed
+- Memory system enhancements (#148)
+  - Epic-linked task completions auto-promote to swarm scope (visible to all workers)
+  - `inject-learning` creates swarm-scoped memories
+  - Mandatory `memory-search` directive in base prompt
+  - Follow-up tasks include epic context (goal, plan, progress, nextSteps)
+  - Server-side memory injection enriched with epic name/goal and recent task summaries
+  - New `nextSteps` column on epics (migration 005)
+- Base prompt updated with VCS CLI comparison table (gh vs glab)
+- DB migration 006: renames `github*` columns to `vcs*`, adds `vcsProvider` column
+
+### Fixed
+- Prevent duplicate review tasks and fix PR Lifecycle workflow (#150)
+  - Dedup guard for review task creation
+  - Action filtering fixes in webhook handlers
+  - Webhook enrichment improvements
+
+- Workflow automation engine with DAG-based node execution (#142)
+  - Trigger nodes: task created/completed, GitHub events, Slack messages, email, webhooks
+  - Condition nodes: property-match, code-match (sandboxed JS), LLM-classify
+  - Action nodes: create-task, send-message, delegate-to-agent
+  - Template interpolation with `{{variable}}` syntax in node configs
+  - Async node support with pause/resume for long-running actions
+  - Stuck run recovery and retry-from-failure support
+  - 9 MCP tools for workflow CRUD, triggering, and run management
+  - REST API endpoints for workflows and runs
+- Workflows UI with React Flow graph visualization (#144)
+  - Interactive DAG visualization with dagre auto-layout
+  - Custom node components (TriggerNode, ConditionNode, ActionNode) with status overlays
+  - Workflow runs table with execution status tracking
+  - Step detail drill-down panel
+  - Workflows section in dashboard sidebar under Operations
+- E2E workflow test with Docker worker integration
+- Database migration system with numbered `.sql` files and incremental runner (#133)
+- Lightweight code-level heartbeat module for swarm triage without spinning up Claude sessions (#124)
+  - 3-tier approach: preflight gate, code-level triage, Claude escalation
+  - Auto-assignment of pool tasks to idle workers
+  - Stall detection for in-progress tasks
+  - Worker health status correction
+  - Configurable via `HEARTBEAT_*` environment variables
+
+### Changed
+- Migrated inline `try { ALTER TABLE } catch {}` schema blocks to `src/be/migrations/` folder
+
+### Fixed
+- `property-match` workflow node crash when config uses flat format (`property`/`operator`/`value`) instead of `conditions` array (#146)
+- API migration Dockerfile fix for workflow schema
+
+## [1.43.0] - 2026-03-12
+
+### Added
+- Slack thread follow-up routing — @mentions in threads route directly to the worker already active in that thread, bypassing lead delegation
+- Additive Slack buffer (`ADDITIVE_SLACK=true`) — non-mention thread replies are debounced and batched into a single follow-up task with dependency chaining
+- `!now` command for instant buffer flush without dependency chaining
+- `HEURISTICS.md` documenting all Slack routing rules and buffering behavior
+- `reactions:write` Slack scope for visual buffer feedback (:eyes:, :heavy_plus_sign:, :zap:)
+
+### Changed
+- Eliminated inbox message system — all Slack and AgentMail messages now route directly as tasks
+- Leads poll for tasks like workers (removed poll-task lead block)
+- Child tasks auto-inherit Slack/AgentMail metadata from parent tasks
+- Removed `inbox-delegate` and `get-inbox-message` MCP tools
+- Removed fuzzy name matching from Slack router (replaced by task-based routing)
+
+### Fixed
+- AgentMail sender domain filter now correctly handles "Name \<email\>" format
+
+## [1.36.0] - 2026-03-06
+
+### Added
+- One-time (delayed) scheduled tasks alongside recurring schedules
+  - New `scheduleType` field: `recurring` (default) or `one_time`
+  - `create-schedule` accepts `delayMs` (relative delay) or `runAt` (absolute ISO datetime) for one-time schedules
+  - One-time schedules auto-disable after execution
+  - `list-schedules` hides completed one-time schedules by default (`hideCompleted`)
+  - UI shows type badges (amber=one-time, emerald=recurring)
+- AgentMail webhook domain filters: `AGENTMAIL_INBOX_DOMAIN_FILTER` and `AGENTMAIL_SENDER_DOMAIN_FILTER` env vars to filter incoming webhooks by inbox and sender domain
+
+### Changed
+- Docker worker improvements: streamlined `Dockerfile.worker` and `docker-entrypoint.sh`
+
+## [1.35.2] - 2026-03-05
+
+### Fixed
+- Avoid duplicate heartbeat triage task creation for the same stalled task set
+- Run stale heartbeat resource cleanup even when preflight triage gate bails
+
+## [1.35.1] - 2026-03-05
+
+### Fixed
+- Use unique port variables per service in `docker-compose.example.yml` to avoid conflicts (#137)
+- Clarified that port variables are examples and that isolated network namespaces can share ports
+
+### Changed
+- Added internal cross-links across docs pages and blog/examples navigation (#135)
+- Added canonical URLs and JSON-LD structured data to docs pages
+
+## [1.34.0] - 2026-03-04
+
+### Added
+- Task cost tracking and display in task details page (#131)
+- Schedule and epic HTTP API endpoints for CRUD operations
+- Exhaustive HTTP API integration test suite (#132)
+- `claude-context-mode` as default context management plugin for workers (#125)
+- Base prompt test coverage
+
+### Changed
+- Refactored monolithic `src/http.ts` into modular route handlers under `src/http/` (#132)
+- Abstracted route matching into `matchRoute` utility with dedicated tests
+- Converted handler dispatch to registry-based for-loop pattern
+- Improved system prompt assembly in `base-prompt.ts`
+
+### Fixed
+- Context-mode marketplace plugin ID in install command (#130)
+- Lint warnings and type errors across HTTP route handlers
+
+## [1.32.0] - 2026-03-03
+
+### Added
+- Model control per task, schedule, and global override — `model` parameter (`haiku`/`sonnet`/`opus`) on `send-task`, `task-action`, `create-schedule`, and `update-schedule` (#127)
+- Schedule-to-task linking via `scheduleId` — tasks created by schedules have a direct back-reference and `get-tasks` supports filtering by `scheduleId` (#127)
+- Multi-credential support — `CLAUDE_CODE_OAUTH_TOKEN` and `ANTHROPIC_API_KEY` accept comma-separated values for load balancing across subscriptions (#119)
+- `ANTHROPIC_API_KEY` as alternative credential to `CLAUDE_CODE_OAUTH_TOKEN`
+- x402 payments guide page and environment variables reference in documentation site
+
+## [1.31.0] - 2026-02-28
+
+### Added
+- x402 payment capability for agents — automatic USDC micropayments for x402-gated APIs (#108)
+- Dual signer support: Openfort (managed wallet in TEE) and viem (raw private key)
+- Openfort backend wallet signer with v-value normalization for USDC settlement
+- x402 CLI for testing payments (`check`, `fetch`, `status` commands)
+- Spending tracker with per-request and daily limits
+- Real testnet E2E tests with x402.org facilitator on Base Sepolia
+- Landing site: x402 example page, blog section with Openfort hackathon post and swarm metrics post
+
+### Fixed
+- Openfort signature v-value normalization (v=0/1 to v=27/28) for on-chain USDC settlement
+- Network chain passthrough to Openfort signer (was hardcoded to baseSepolia)
+
+## [1.30.1] - 2026-02-28
+
+### Added
+- Agent `lastActivityAt` timestamp for stall detection (#105)
+- Slack attachment handling — voice memos, images, and file uploads are now processed as messages (#103)
+- `includeHeartbeat` filter for `get-tasks` — heartbeat/system tasks are excluded by default (#102)
+- Tool annotations (`readOnlyHint`, `destructiveHint`, `idempotentHint`) on all 36 MCP tools for improved Tool Search discoverability (#95)
+
+### Changed
+- Pinned Dockerfile builder to `bun:1.3.9` for reproducible builds
+- Dockerfile improvements: `pipefail`, consolidated `RUN` layers, `--no-install-recommends` for Node.js and GitHub CLI
+- Removed `cc-ai-tracker` from worker image agent tools
+- README optimized for GitHub star conversion: badges, hero, issue/PR templates (#104)
+
+## [1.28.1] - 2026-02-27
+
+### Added
+- Fumadocs documentation site at docs.agent-swarm.dev (18 pages across architecture, concepts, guides, and reference sections)
+- Agent-swarm.dev landing page
+- Agent artifacts feature via localtunnel — SDK, CLI command, `/artifacts` skill, and Docker support
+- Depot build system for Docker images
+- Slack offline message queuing — @mentions when no agents are online are now queued as tasks
+- `AGENTMAIL_DISABLE` env var to skip AgentMail integration
+
+### Changed
+- Server-side aggregation for usage pages (performance improvement)
+- Removed old `ui/` directory in favor of `new-ui/`
+
+### Fixed
+- Usage pages performance issues (5 review fixes: full table scan, SQL parameterization, useMemo deps, groupBy validation, test coverage)
+- CI path filtering to skip workflows for docs-site and landing directory changes
+
+## [1.28.0] - 2026-02-17
+
+### Added
+- New dashboard UI ("Mission Control" theme) with AG Grid, command palette, and dark mode
+  - Phase 1-6: project scaffolding, app shell, config page, agents/tasks/epics pages, chat/schedules/usage pages, polish
+- Comprehensive env vars reference and agent configuration docs
+- Active sessions table for lead concurrency tracking
+- Concurrent context endpoint for lead session awareness
+- Task deduplication guard to prevent concurrent lead duplicates
+- Workers wake on in-app chat @mentions
+- Delete-channel MCP tool (lead-only)
+
+### Changed
+- README and docs cleaned up for public launch
+- Polished env examples and DEPLOYMENT.md
+
+### Fixed
+- New UI: CSS vars instead of hardcoded oklch in charts
+- New UI: swapped theme and sidebar active state
+- New UI: stale config dialog, chat URL params; removed dead code
+- Zombie task revival — prevent completed tasks from being revived
+- Task pool claiming made atomic to prevent race conditions
+
+## [1.25.0] - 2026-02-07
+
+### Added
+- Agent self-improvement mechanisms (7 proposals implemented)
+- Follow-up task creation for lead on worker task completion
+- `/internal/reload-config` endpoint and config loader extraction
+- Session error tracking with meaningful error reporting for failed worker sessions
+
+### Fixed
+- Graceful fallback when session resume fails with stale session ID
+- Lead task completion polling prioritization and increased concurrency
+- Slack initialized flag reset on stop
+- AgentMail `from_` type fix
+
+## [1.21.0] - 2026-01-28
+
+### Added
+- MCP tools for swarm config management and server config injection
+- AgentMail webhook support
+- Persistent memory system with vector search
+- Centralized repo management
+- Persistent setup scripts and TOOLS.md for agents
+- Soul/identity editors in UI profile modal
+- Session attachment with `--resume` logic in runner for session continuity
+
+### Fixed
+- Permanent notification loss from mark-before-process race
+- 404 handling in task finalization
+- Config upsert with NULL scopeId for global config
+
+## [1.16.3] - 2026-01-14
+
+### Added
+- Epics feature for project-level task organization
+- Lead-only authorization for epic tools
+- Slack user filtering by email domain and user ID whitelist
+- Scheduled tasks feature (cron-based recurring task automation)
+
+### Fixed
+- Task totals to show absolute counts
+
+## [1.15.8] - 2026-01-07
+
+_Initial tracked version. Earlier changes are not included in this changelog._