Release notes

Unreleased

• Job search, first cut (docs/design-job-search.md): a career/ folder the person owns (profile, master resume, applications tracker, one folder per application). Tools in the loop: career_profile, career_resume (ingest an attached resume into markdown), jobs_search on Greenhouse, Lever and Ashby boards with no login, jobs_record for postings found in the browser, jobs_posting, career_score (fit against a floor), career_tailor with a truth check that traces every line to the master and cuts what it cannot, PDFs through a local Chromium, and career_track for the board. "Submit application", "Apply now" and "Easy apply" clicks stop at the gate. Daily application cap in the profile.

0.2.0

• Attach files to a message: paperclip, drop or paste in the composer, up to 8 files of 10 MB. jigbee reads plain text and code as is, Word/RTF/ODT documents, PDFs, and the text in images. The first 16k characters ride along with the message; the rest is one read away for the running job. Files show as chips on the message.
• Mac app: a core left running from an older build is replaced on launch instead of reused, so a new build shows its own page. jigbee --version and /api/health carry a build stamp.
• Chat shows reasoning and progress in short form; the browser can press key combinations (Cmd+Enter, Shift+Tab).
• jigbee mcp: a stdio MCP server so Claude Desktop, Cursor or any MCP client can start and watch jobs.
• Coding jobs: a native coder (no external CLI needed) alongside Claude Code and Codex; the coder is picked by its record on this machine after three runs.
• Per-job spend cap counts what the job already spent; declining a login continues the job without that site.
• Each launch gets its own API token; cross-site requests to the local API are refused.
• Ctrl-C closes the browsers a job parked.

Most of what follows was in the first build on the site (0.1.0) too, and is listed here for the first time.

• One agent loop for browser and desktop, replacing the per-mode loops and the keyword router. The model decides chat vs job; jobs run plan → act → observe with a running progress note.
• Guard rails in the loop: one-time desktop consent, a stop-and-ask gate for anything irreversible (pay, send, delete, book, submit), loop detection, retries, step budget → checkpoint.
• Steering reaches the running job on its next step instead of queueing behind it.
• Engram v2 and ReasoningBank v2: plain JSON in <home>/memory/. The model writes memories and a procedure at the end of each job; the next job starts with them. Visible and deletable in Settings.
• The Jev client and the heuristic reasoner are gone. Internal reasoning no longer appears as chat text.
• Mac app: picks a free port, restarts the core, opens OAuth and external links in the default browser, native menus, Dock badge, notifications.
• jigbee for Chrome: an extension that lets jobs run in the person's own Chrome, signed in as them, in a window of its own. Paired with a six-digit code from Settings; the built-in headless browser remains the fallback.
• First run: three steps in the empty state (model, browser, a starter job) with a one-click Chromium download for the built-in browser.
• Replies can carry a table, links or a list, and the last screen of the job, with Reveal in Finder.
• Every finished job is recorded against the procedure it used. Repeat jobs show "N% fewer steps than the first time" on the reply; Settings has a "so far" card with jobs, steps, spend and repeat savings.
• Spend cap per job (JIGBEE_JOB_BUDGET_USD, default $1): the loop pauses and asks before passing it.
• Evals: a fixed job set run cold then warm, a nightly launchd job, and a badge on the site that shows the last run's pass rate and warm-run savings.
• Requests to write or change software get a short reply pointing to Claude Code or Codex. Out of scope by design.

0.1.0

• Bring your own model: Sign in with ChatGPT (uses your plan), OpenAI, Anthropic, xAI, or any OpenAI-compatible endpoint; picked in Settings or .env.

First public release.

• Agent loop with Engram-style memory and ReasoningBank strategies on local Qdrant; ask and path replay the reasoning behind any answer.
• Computer use: browser via Playwright MCP, desktop via a Cua computer server, whole web tasks via Browser Use. Grok picks each action.
• Long sessions: queued tasks with a checkpoint after every action, pause after the current step, resume with a note of steps already taken, steer by sending a message mid-run, reload-as-paused after a crash.
• Local app: sessions sidebar, step thread, live screen and action log, cost meter. Demo mode with no keys.
• Mac app: bundles the core and Qdrant, keeps config and sessions in ~/Library/Application Support/jigbee, in-app Settings for keys. Ad-hoc signed; notarization when a Developer ID is configured.