← HomeDocsGitHub

What's shipped.

Every release, generated straight from the repo's changelog. Beacon follows semantic versioning and the Keep a Changelog format.

unreleasedin progress
Added
  • Reddit and Medium platforms (opt-in: beacon config set platform reddit on, beacon config set platform medium on). Reddit drafts a title and a self-post body for a technical subreddit; Medium drafts a full story with title, optional subtitle, tags, and body.
Removed
  • Bluesky and Mastodon platforms. Disable them in existing configs before upgrading if you rely on them — platforms.bluesky and platforms.mastodon are no longer recognized keys.
v0.5.0July 11, 2026latest
Added
  • beacon ui — review the queue in the browser: platform drafts side by side, inline editing, copy to clipboard, approve/discard, and live updates over SSE as new commits draft. Localhost-only with per-session token auth, a host-header allowlist against DNS rebinding, and a strict self-only Content-Security-Policy. Attaches to a running beacon serve instead of starting a second instance.
  • beacon serve — the local review API behind the UI, runnable headless: GET /queue, POST /entries/:id/approve|discard, PATCH /entries/:id/drafts, GET /events (SSE), GET /health.
  • Cross-process queue lock — the git hook, beacon review, and the local API now serialize queue writes through an advisory file lock (~/.beacon/queue.lock), so concurrent writers can no longer lose each other's updates. Stale locks from crashed processes are detected and reclaimed automatically.
  • Per-repository config — a .beacon.json at a repo root overrides the global config for that repo (enabled, platforms, significanceThreshold, language, authorNotes). It is ignored until you approve it with beacon trust, which pins the file's SHA-256; editing the file lapses that approval automatically. A repo config may never set apiKey, baseUrl, provider, or model.
  • enabled: false — opt a repository out entirely. The hook becomes a no-op ahead of any API call, so an opted-out repo costs nothing.
  • beacon doctor now reports which config layers are in effect and warns when a .beacon.json is present but untrusted.
  • npm provenance — releases are published with a signed provenance attestation linking the package to the commit and CI run that built it. CI actions are pinned by commit SHA.
  • SECURITY.md — a written threat model: what leaves your machine, what never does, and what the scanner does not protect against.
  • Ollama auto-detection in beacon init — picking Ollama now finds the running daemon and offers your pulled models as a picker, instead of asking you to type a base URL and a model name from memory.
  • beacon config show is readable — an aligned key/value view that says which API key is actually in effect (and when an env var overrides the stored one). The old raw-JSON output is still there behind --json.
Changed
  • beacon init asks less — setup now needs only the essentials (provider and key); bio, voice notes, and language sit behind a single optional "personalize now?" step, and the significance-threshold question is gone (it defaults to 6 — tune it any time with beacon config set significance-threshold). A failed connection test now offers to re-enter the key or retry instead of marching on with a broken setup.
  • A branded terminal — Beacon's yellow now runs through the whole CLI: a ✦ beacon wordmark opens each command, prompts render as one connected flow, and beacon review shows each draft as a card with per-platform character counts (flagged when a tweet runs past 280).
Fixed
  • Secrets in the commit message reached the model. The scanner only ever saw the diff, while the commit message was sent verbatim to both LLM calls — so git commit -m "rotate sk-ant-…" shipped the key to the provider. Both surfaces are now scanned before either call.
  • The repository was named after the working directory. A commit from a subdirectory told the model (and the drafted post) that the repo was called src, or pipeline; it now uses the repository root.
  • A queue from a newer Beacon reported as corrupt. queue.json is now versioned: an older file is migrated (after its bytes are copied aside), and a newer one prompts you to upgrade rather than implying data loss.
  • Re-running beacon init after switching providers suggested the old provider's model. Choosing Anthropic on a machine previously set up for OpenAI offered gpt-4o-mini as the default; the suggested model now follows the provider you just picked.
  • Piping output could crash. beacon config show | head (or any early- closing reader) died with an unhandled EPIPE stack trace; a closed pipe now ends the process quietly.
Security
  • queue.json was world-readable and stored the raw diff. A warning-level finding (an .env assignment, a private IP) is redacted from the model but does not block drafting, so the raw secret was written to a 0644 file. queue.json is now 0600 and stores only the redacted snapshot; ~/.beacon and every file under it are owner-only, repaired on write.
v0.4.0July 10, 2026
Changed
  • Leaner install — the anthropic provider now calls the Messages API directly instead of through @anthropic-ai/sdk. Beacon's LLM layer ships with no runtime dependencies: seven packages fewer on every install (9.8 MB in the SDK alone), and roughly 30 ms off CLI startup.
  • base-url applies to every provider, not just OpenAI-compatible ones — point anthropic at a proxy or gateway with beacon config set base-url <url>. beacon doctor now reports the base URL whichever provider you use.
Fixed
  • beacon doctor exits non-zero when the live provider ping fails, as documented. An invalid API key previously reported a healthy setup.
Removed
  • ANTHROPIC_BASE_URL is no longer read from the environment. Use beacon config set base-url <url> instead.
v0.3.1July 4, 2026
Added
  • Production website — Next.js port of the project site at beacon-bip.vercel.app; docs in MDX, changelog page generated from this file at build time.
  • README demo GIF showing beacon init → commit → beacon review.
Changed
  • Site redesigned to match the new mockup: full-viewport hero with the platform flip word, responsive nav and type scale on mobile.
  • The beacon GitHub repository is now public.
Fixed
  • punycode DeprecationWarning no longer pollutes CLI output on Node ≥ 21 (upgraded @anthropic-ai/sdk to 0.110).
  • README pipeline description now matches the actual stage order.
  • Site favicon adapts to dark browser chrome.
v0.3.0July 4, 2026
Added
  • beacon init — guided first-run setup: provider, key, model, voice, language, hook install, connection test — ending with a draft from your latest commit so the first session produces real output.
  • beacon doctor — setup diagnostics: node/git versions, config, API key, hook presence and executability, beacon on PATH, and a live provider ping. Exits non-zero when a check fails.
  • Ollama preset in beacon init — fully local, free drafting with no API key (http://localhost:11434/v1 via the OpenAI-compatible provider).
  • Digest drafting — beacon draft --since <when>, --week, and --today combine a range of commits into one draft ("here's what I shipped this week"). Accepts anything git log --since understands.
  • Bluesky and Mastodon platforms (opt-in: beacon config set platform bluesky on). Platform keys in a draft set are now optional; only enabled platforms are drafted and validated, and queue entries from older versions still parse.
  • Configurable author identity and language — new config keys authorName, authorBio, and language drive the drafter's voice prompt (previously hardcoded). Drafts can be generated in any language.
  • Project site (design/beacon-site.html) — single-file landing, docs, and changelog pages; fully responsive, with the main nav collapsing into a toggle menu and a dedicated small-screen typography scale on mobile.
Changed
  • beacon review edit flow — editing now opens a single platform's draft as plain text (tweets separated by ---, markdown for dev.to) instead of raw JSON, with structured parts falling back to the original draft when deleted. Edits are still schema-validated on save.
  • The drafter prompt is composed from enabled platforms only, reducing token cost when platforms are toggled off.
  • CLI output overhauled — semantic color helpers (TTY- and NO_COLOR-aware), a dependency-free spinner that animates on a TTY and degrades to plain lines when piped or inside the git hook, plus a shared first-run nudge and consistent friendly error rendering across commands.
Fixed
  • QuietSpinner constructor call passed an argument it did not accept.
  • OpenAI-compatible 401/403 responses now normalize to AUTH_ERROR (was API_ERROR).
v0.2.1June 25, 2026
Fixed
  • LLM JSON parsing. Prompt templates no longer embed // comments, which models echoed back as invalid JSON; parsing is also more tolerant of trailing commas in model output.
v0.2.0June 25, 2026
Added
  • Multi-provider LLM support. Beacon is no longer Anthropic-only. A new provider abstraction (src/lib/llm/) supports: · anthropic (default) — via the official SDK. · openai — any OpenAI-compatible Chat Completions endpoint (OpenAI, OpenRouter, Groq, Together, a local server, …) over fetch, configurable with baseUrl. No new dependency.
  • New config keys: provider (anthropic | openai) and baseUrl.
  • New commands: beacon config set provider <name> and beacon config set base-url <url>.
  • Provider-aware API-key resolution: ANTHROPIC_API_KEY or OPENAI_API_KEY (env vars take precedence over the stored key).
Changed
  • Pipeline reordered to capture → safety → significance → draft → queue. The safety scanner now runs before every LLM call. Previously the significance filter received a raw (unredacted) diff excerpt before scanning; now both LLM calls only ever see the redacted diff. A critical finding blocks the pipeline before any network call is made.
v0.1.1June 25, 2026
Fixed
  • bin path. bin.beacon pointed at ./dist/cli/index.js, but the build emits ./dist/index.js, so the global beacon command failed to resolve.
  • Git stderr (e.g. fatal: Needed a single revision when probing HEAD~1 on a repo's first commit) no longer leaks into commit/hook output; it is captured instead of inherited.
v0.1.0June 25, 2026
Added
  • Initial release: run, install, review, draft, and config commands.
  • Five-stage local pipeline that drafts build-in-public content for Twitter/X, LinkedIn, and dev.to from git commits.
  • Regex-based safety scanner, significance filter, voice drafter, atomic JSON review queue, and a post-commit git hook.
  • Daily log rotation with a 7-day retention window.