Version history
Changelog
Every ResearchForge release, from the project's CHANGELOG.
v0.4.3
Latest on PyPINovember 16, 2026$ pip install -U researchforge
Ported notebooks bring every input- featA ported public notebook brings everything it attaches, not only datasets: propose-from-kernel copies dataset_sources, kernel_sources (other notebooks’ outputs), model_sources (Kaggle Models), docker_image and machine_shape (GPU type) into proposal.json; pull-kernel lists them, and execute-round and run-parallel push them with the round’s notebook. Hand-written proposals can use the same keys
v0.4.2
November 16, 2026Code competitions run the full loop- featA code competition now gets the full experiment loop on Kaggle GPU, where each round runs as a Kaggle notebook; choosing other compute stops setup with the reason. Only data larger than the free disk still gets the ideation plan
v0.4.1
November 16, 2026Dashboard: Needs you- fixDashboard project picker groups projects as Needs you (questions waiting, paused or a run stalled), Active and Idle; a competition with a question waiting is no longer filed under Idle
v0.4.0
November 15, 2026$ pip install researchforge
Kaggle OvernightAuto-runBackground sensorsLive dashboardClaude Code pluginLocal models- feat/overnight <competition-url> skill and researchforge kaggle overnight: reads the competition, finds starter kernels, asks a few setup questions, then runs a propose → run → measure → keep loop on the compute you pick — Kaggle GPU, this machine, Colab, an SSH server or a RunPod / Lambda GPU
- featCompetition data never downloads locally when the compute policy forbids it
- featAnswers that change a setting: kaggle ask --action ties an answer to Run configs changes (compute, budgets, round time limit, Kaggle GPU cap, Colab plan and GPU, metric direction); picking it on the dashboard (marked ⚙, confirmed first) or with kaggle respond applies them, logged as “answer to q-…”. Only a person can pick one
- featThe 🔔 groups questions by competition, each section with its count and Kaggle submissions left today (“7 of 10 submissions left today”); a used-up day is refused before Kaggle is called
- featHuman-gated submissions: experiment submission-candidates ranks candidates by evidence; experiment submit asks for confirmation and records the public leaderboard score back into the experiment graph
- featBackground sensors (kaggle monitor): new public kernels → ideas, submission status, hypothesis-pool growth, hung-experiment heartbeat, arXiv scan and a morning briefing; custom sensors from any shell command
- featkaggle scheduler-install runs the monitor as a macOS launchd service, so it keeps working with Claude Code closed
- featLive local dashboard for every competition and repository project: Home, Automation, Runs, DAG, Hypotheses, Submissions, Timeline and Insights
- featautorun --no-metric-gate records every result instead of auto-rejecting; experiment plan --merge a,b combines two independent winners
- featresearchforge notify setup: ntfy push for every result, merge opportunity and briefing
- featSteer a running loop: kaggle status, pause, resume, respond, compute, add-hypothesis and settings, also as /overnight status, /overnight pause, …; questions never block the loop
- featscheduler-install also on Linux (a systemd user timer) and Windows (Task Scheduler)
- featClaude Code plugin: /plugin marketplace add forger-labs-hq/researchforge, then /plugin install researchforge@researchforge
- featLocal models through Ollama: RESEARCHFORGE_LLM=ollama/<model>, no key and no token cost; Kaggle Overnight runs headless on it
- featautorun --kaggle and --judgment for runs where the local metric is not the real score
- featAuto-run (kaggle auto-run <slug> --on, off by default): an opt-in sensor runs rounds on untested ideas with no chat open, driven by headless Claude Code, the Cursor agent or your API key; gated by pauses, open questions, a daily limit and the Kaggle GPU cap, and never submits
- featOne driver per project, switchable any time (kaggle driver <slug> <driver>), which also does the sensors’ AI work; kaggle driver-login signs a driver in from your own terminal
- featQuestions come to you: kaggle ask pushes to your phone and the dashboard’s 🔔; only the ideas it names wait, and kaggle respond --dismiss closes one unanswered
- featRun configs in the dashboard: every setting for a competition, editable on the fly, with a change log
- featMetric direction read from the leaderboard’s order, with its source shown and guesses flagged; kaggle settings --metric-direction fixes it
- featNotebook ideas show the leaderboard score their author claims and the gap to your best, labelled unverified
- featkaggle archive (reversible) and kaggle delete-round (confirmed); kaggle pull-kernel downloads a public notebook to port from
- featColab without clicking Run all, through Google’s Colab CLI: each round rents a fresh VM with the GPU your plan allows, runs in the background (checked every minute, so a dropped connection no longer kills it), brings back its submission file and releases the VM; used on a real Colab account; Colab Enterprise not supported yet
- featOne history per competition: the Kaggle loop is the only engine; a repository project beside it is folded in (kaggle merge-history) as imported nodes that never count as best or as a parent; with no baseline a round must beat the best earlier round to count as improved
- feat+ New competition in the dashboard: paste a Kaggle URL, answer the /overnight questions as a form, and watch setup run step by step; the answers are a draft Claude Code or Cursor can finish (kaggle draft)
- featKeys & accounts: every credential on one local page and kaggle keys status|open|wait|set|test|remove; /overnight opens the page for missing keys instead of asking in the chat
- featSteer: directions for the next round or every round (kaggle steer), answered by the proposer; pause, resume, re-synthesize and add ideas from the same page
- featApprove & submit from the 🔔: every improved round you could submit is offered, Kaggle’s daily limit is checked first, and a round is never submitted twice; unscored notebook rounds are kept and offered too (kaggle offer-submission)
- featPublic notebooks checked every hour, with a notebook board ranked by claimed score; notebooks ahead of your best leaderboard score rank first; kaggle propose-from-kernel runs one as a round
- featresearchforge auto-run for repositories: a background monitor runs the repository loop with nothing open, plans wait for Approve plan unless you type a standing approval, and Automation → Repo projects controls it (new; tested on the bundled example repository)
- featRound time limit in Run configs (30 minutes by default); per-competition ntfy topic with New topic and Send test alert; submissions checked every 15 minutes
- fixAn unattended run never asks you to run a command; ideas it cannot do are skipped with a reason (kaggle skip-idea). Alerts no longer fall back to a shared public topic
- fixSetup always offers every compute option, and an unknown or unconfigured --gpu target stops setup instead of running somewhere else
- fixStalled rounds recover: a dead round’s lock is released and the round recorded as interrupted, and the monitor runs a missed cycle within a minute of the Mac waking
- fixRunning /overnight again on a set-up project resumes it instead of redoing setup (--refresh starts over)
v0.3.0
GitHub onlySeptember 1, 2026Never published to PyPI on its own; everything in it ships in 0.4.0.
Loop without an API keyCost & token accounting- featautorun --dry-run shows which node the loop would expand next and which hypotheses it ranked there, without running anything
- featexperiment plan --parent exp-NNN builds on a measured winner
- featresearchforge-autorun skill and Cursor rule drive the loop from the editor, with no API key
- featTime economics on the dashboard: compute by stage and by how each experiment ended, and benchmarks avoided by screening or the stall rule
- featToken accounting for Anthropic, Gemini and OpenAI, attributed to what asked for it; unpriced models are named, never billed at zero
v0.2.0
August 30, 2026Standalone AIautorunExperiment graph- featBuilt-in Anthropic, Gemini and OpenAI providers (and Ollama), so every step works from the CLI — Claude Code and Cursor become one option of three
- featresearchforge autorun: unattended rounds with stall, global-stall, max-rounds, max-hours and target stopping rules; --resume continues an interrupted loop
- featExperiments became a graph: build on any earlier winner, or merge two winners to test whether their gains add up; UCB1 node selection
- featPatches written from the real source files, so plans no longer fail to apply
- featNO CHANGE badge, a derived audit trail with gate findings, and a readable experiment graph
- featship pr asks where the pull request should go; opening a PR on a repo you cannot push to always needs typed confirmation
- featbaseline run diagnoses and retries setup failures; baseline run --n-runs N freezes the mean of N runs
v0.1.0
First public releaseJuly 24, 2026Open-source beta, Claude-first.
Full local pipelineClaude Code skills- featResearch intelligence: arXiv discovery, the Claude ↔ CLI synthesis handshake (landscape + hypotheses with graded evidence) and a citation-backed report
- featExperiment contract: researchforge.yaml wizard, semantic validation and typed approval into immutable versions
- featIsolated execution: a detached git worktree per attempt, Docker and .venv runners, path guard, timeouts and secrets redaction
- featExperiment funnel: screening → full benchmark → repeated validation, hard constraints, and rejected experiments preserved
- featShipping: a clean branch rebuilt from the frozen baseline, optional draft PR and an engineering report
- feat12 installable Claude Code skills
Follow along
Star the repo to hear about new releases.