Changelog

All notable changes to SideButton.

[1.5.9] - 2026-08-31

Agent Server

• publish_artifact works for the whole life of an app editing session — the session-open job completes on its boot turn by design, and the dispatcher clears job-context.json when it does, so every later chat turn ran with no job context and the tool refused with "No active job" for the entire rest of the session. It now falls back to the sticky app-session marker the autosave lane already maintains, and attributes on the session id alone — which the artifacts endpoint has always accepted, resolving the step from it before ever looking at job_id/step_index.
• The legacy SSE transport gets per-client session identity — concurrent MCP clients on the pre-streamable transport (GET /mcp stream + POST /message) shared one implicit session, so under load one client's reply could land in another client's stream, or be silently discarded after the server had already answered 202 — both surfacing as "Request timed out". Each stream now carries its own session id with a real -32002 for unknown ids; replies survive brief disconnects via a bounded replay buffer that retires only what a live stream actually received; and a request whose session goes away mid-flight is answered synchronously instead of being dropped.

[1.5.8] - 2026-08-30

Agent Server

• New workspaces no longer stall on the "Allow external CLAUDE.md file imports?" dialog — Claude Code added a second interactive gate past folder trust, triggered by the portal-written workspace pointer (CLAUDE.md → @AGENTS.md) when a session launches inside a project checkout. The pre-launch trust seeding now approves external includes for every portal-provisioned workspace directory, and also repairs a previously recorded decline (which silently stripped the workspace AGENTS.md from every later session in that directory).

[1.5.7] - 2026-08-30

Agent Server

• Live app-chat streaming finds the transcript again — Claude Code ≥ 2.1.x creates the session transcript under its own uuid, ignoring the pre-set --session-id the dispatcher passes, so the 1.5.6 session watcher could never locate the file and the live stream stayed dark. The watcher now falls back to the single actively-written transcript when the direct lookup misses (only when unambiguous — two candidates track nothing), and keeps reporting under the dispatcher's session id so every portal surface keys as before.

[1.5.6] - 2026-08-30

Workflows & MCP

• browser.screenshot step — captures the page (or a selector / ref / region crop) straight to a PNG file on the agent machine, so an agent can produce a docs screenshot without image bytes entering its context. Brings the step count to 46 (43 implemented).
• screenshot MCP tool takes path — same file output from the tool. With no path it behaves exactly as before and still returns image bytes, so existing callers are unaffected.
• inject_css MCP tool — injects a CSS rule into the page. Combined with the above, browser_batch can run navigate → blur the sensitive selectors → wait → screenshot-to-file in a single call, so no unredacted image ever reaches disk.
• docs_screenshot workflow — bundled single-shot recipe wiring that sequence together.
• Output paths for both writers are contained to the home directory — .. traversal and escaping symlinks are rejected, and nothing is created on disk until the path is cleared — matching the rule publish_artifact already applies, so a captured shot can be published directly. publish_artifact now expands ~/… too, so the path the screenshot tool hands back can be published verbatim.
• A step whose was never passed fails loudly — browser.screenshot and browser.injectCSS refuse an unresolved placeholder instead of treating it as a value. Previously a missing redact_css injected the literal text as a stylesheet, which styles nothing: the run reported success and the screenshot that followed was unredacted. Pass an empty string to mean "no redaction".
• Browser pre-flight gate is recursive — run_workflow now detects browser steps nested inside control.foreach / control.if. Previously only top-level steps were checked, so a nested browser step skipped the friendly "browser not connected" error and failed mid-run instead.
• Running an unrecognised step type on an older server still fails fast at parse time with PARSE_ERROR: Unknown step type, so a workflow using browser.screenshot against a pre-1.5.6 server reports a clear error rather than silently skipping the capture.

Agent Server

• Close the window, end the job — closing a dispatched job's terminal window now kills its Claude session and resolves the job, instead of leaving a headless session running behind a closed window.
• Live chat streaming — an app-session chat turn streams into the portal as the agent writes it, rather than arriving as one block at the end.
• Scripted app-session boot — the dev-session boots via script before Claude launches, cutting session start-up time and making the ready report deterministic.
• Notion for agents — NOTION_TOKEN is delivered into the agent environment with an agent-facing connector doc, and the first (gated) Notion browser-webhook lane shipped.
• GitLab — a glab-based git provider with a host-aware push-PR workflow, a stateless gitlab.com credential helper in both manual lanes, and the workspace attach lane's gaps closed.
• Claude Code plugin install report — /health and the agent detail page now report each plugin's install outcome, and a malformed report can no longer wipe a real one.
• Fleet honesty fixes: "no evidence" now means exactly that everywhere, and the dispatch stability age keeps its own clock.

Default skills

• Agents pack 1.25.0 — the documentation lane ships in the default catalog: six agent_docs_* workflows (scope, decisions, write, verify, post-deploy, release-scan) plus the writer role, powering the Documentation Page playbook end-to-end; also adds agent_se_review (review + fix without merging) and the review-state channel refinements.

[1.5.5] - 2026-08-11

Note on 1.5.4: the 1.5.4 npm packages shipped with a stale server build that did not contain the changes below, despite reporting version 1.5.4 at runtime. 1.5.5 is the corrected release — if you installed 1.5.4, update.

Agent Server

• Preview passthrough — ANY /api/preview/:port/* proxies HTTP and WebSocket to a dev server on the agent machine's loopback, so a live app (including hot module reload) can be viewed through the portal. Guarded by the same bearer token as the rest of /api/*; SIDEBUTTON_PREVIEW_PORTS narrows which ports are reachable.
• Trustworthy reboots — POST /api/system/reboot now runs the privileged reboot wrapper first, awaits the result, and reports what actually happened instead of replying ok before attempting anything.
• Build verification on publish — the npm packages now rebuild themselves at publish time (prepublishOnly), so a release can no longer ship a stale compiled output.

Default skills

• Agents pack 1.22.0 — the dev-session boot report gains an optional ROUTES: block (the app's pages, read off the router) and a [viewing /path] context line on user turns, and the ops playbook's app_edit_session workflow carries the full per-project app contract.

[1.0.12] - 2026-03-28

Highlights

• MCP OAuth 2.1 — full OAuth discovery and registration for Claude Code 2.1.84+ compatibility
• Temporal orchestration panel — real-time workflow execution visibility in job detail view
• Mobile-responsive portal — Fleet Control pages adapt to mobile/tablet viewports
• Plugin system — extend the MCP server with custom tools via handler scripts in any language
• 45 step types — added issues, git, chat, and data step categories (42 implemented, 3 chat pending)
• Knowledge pack CLI — sidebutton init, validate, and publish for creating and sharing knowledge packs
• Context system — persona, roles, and targets injected into every LLM call via both REST and MCP
• Community roles — 15 built-in role templates (SE, QA, PM, SD, HR, Sales, and more)
• Abstract providers — issues and git steps work across Jira, GitHub, and other platforms via auto-detected providers

Portal & Dashboard

• Temporal orchestration panel on job detail page with phase timeline
• Sidebar count badges for Jobs and Queue
• Clickable navigation across all dashboard tables
• Removed collapsibles and Steps section from job detail for cleaner UX
• Mobile-responsive Fleet Control pages
• "All agents" targeting option for automations

MCP & Agent Infrastructure

• MCP OAuth 2.1 discovery and /register endpoint for Claude Code compatibility
• Fixed OAuth discovery returning HTML 404 to Bun-based clients
• Agent 10-minute inactivity timeout for Claude Code busy status
• Default LLM model updated to gpt-5.4-nano

New Step Types

• browser.extractMap — extract structured data with field mapping
• browser.fill — set input values directly (React-compatible)
• browser.select_option — select dropdown options by value or label
• browser.scrollIntoView — scroll elements into viewport
• browser.injectCSS / browser.injectJS — inject styles and scripts
• browser.snapshot — capture accessibility tree for LLM analysis
• llm.decide — LLM picks from a list of actions based on context
• data.get — get list item by index
• variable.set — set variables directly
• issues.* — create, get, search, attach, transition, comment (6 types)
• git.* — listPRs, getPR, createPR, listIssues, getIssue (5 types)
• chat.* — listChannels, readChannel, readThread (3 types, pending provider)

---

[1.0.0] - 2025-12-27

Initial Release

The first public release of SideButton.

Features

• Workflow Engine

  • 20 step types (browser, shell, LLM, control flow)
  • YAML-based workflow definitions
  • Variable interpolation with {{variable}} syntax
  • Nested workflow support via workflow.call

• Browser Automation

  • Chrome extension with WebSocket connection
  • Navigate, click, type, scroll, extract
  • Recording mode for capturing actions
  • Embed buttons for in-page automation

• MCP Server

  • Tools for workflow and browser control
  • Compatible with Claude Code, Cursor, VS Code, Windsurf
  • SSE transport over HTTP

• Dashboard

  • Svelte-based web UI
  • Workflow management
  • Recording management
  • Run log inspection
  • Settings configuration

• Desktop App

  • Electron-based native app
  • macOS, Windows, Linux support

• LLM Integration

  • OpenAI and Anthropic support
  • llm.classify for categorization
  • llm.generate for text generation
  • User contexts for customization

Workflow Library

Includes example workflows for:

• Wikipedia summarization
• News site aggregation
• GitHub release creation
• General browser automation

---

Versioning

This project uses Semantic Versioning:

• MAJOR: Incompatible API changes
• MINOR: New functionality (backward-compatible)
• PATCH: Bug fixes (backward-compatible)

Reporting Issues

Found a bug? Open an issue on GitHub.