Skip to content
neuroflow0.2.31 · docs

← neuroflow

The claude.ai workbench

Your own private page over your neuroflow memory: the task boards with drag and drop, your hives' project lists and the wikis — mirrored from git, with every change going back through Claude Code.

What it is

The workbench is one private claude.ai page per person. It has a tab for each level you mirror — your flowie, each hive you joined, and each project you export it from — and every tab is drawn the same way:

  • Board — the level's Kanban: drag a card to another column, or use the keyboard (Enter or M opens Move to, Shift + ← / → moves a card one column, U undoes your last waiting move, N adds a task).
  • Projects — the level's project registry, with each card's status and phase.
  • Wiki — the level's wiki, with [[links]] between pages working.
  • The level — the flowie's ideas, notes and meetings; a hive's team text, people, ideas and errata; a project's phase map, dates and integrity markers.

Git stays the only store. The page shows a snapshot of your files, and a click on it is a request — an intent — that Claude Code checks against git and applies when you run /neuroflow:tasks --apply-web. Until then the card shows waits for Claude Code. Nothing on the page is research state that git does not hold: if the page were deleted, you would lose only the clicks still waiting.

It is yours alone. The page's database rules let only its owner read or write anything, so a page shared by mistake shows others nothing. You can change how it looks — theme, density, which fields a card shows, the order of the tabs, hidden columns, display-only column limits, grouping by project or owner — and Claude Code copies those settings into your flowie, so a rebuilt workbench keeps them.

How a click lands in git

  1. You move a card on the page. It writes an intent into your private database and draws the card in its new column, marked as waiting.
  2. In Claude Code you run /neuroflow:tasks --apply-web. After a one-line preview and your yes it claims the waiting intents, so no other machine applies them, then checks each one against git — has the task changed since your snapshot, is it still in the column you saw — and moves the task files exactly as /tasks does. Two moves of one card become one move; a card you added and then moved is written once, in its last column.
  3. Each level follows its own rule: your flowie is committed and synced; a hive's commit is pushed only after your yes; a project's change stays in your working tree for you to commit with /git. Every commit names the clicks it holds in a Neuroflow-Intent trailer, so nothing is applied twice.
  4. The outcomes and the refreshed levels go back to the page in one upload — after a one-line preview and your yes.

The page's top line counts the clicks still waiting and says how old the oldest is. With the neuroflow mod, the band above the prompt says how many workbench changes wait and a puts /neuroflow:tasks --apply-web in the prompt; the boards mark those cards with ↻. Without the mod, /tasks and /dashboard say the same in their text.

Setting it up

Run /neuroflow:dashboard --web in one of your projects. The first time it explains what the workbench is and what would leave your machine, asks for your consent, a name for this machine and which levels to mirror, shows a one-line preview — and only after your yes publishes the page and uploads the levels. After that, the same command refreshes it; --web --status reports without uploading, and --web --levels changes which levels are mirrored.

You need a claude.ai subscription with Artifacts, and Claude Code signed into the same claude.ai account (an API key, Bedrock or Vertex cannot publish one). The workbench belongs to the account that published it: if Claude Code is signed into another account — a personal plan at home, an institute's organization at work — neuroflow stops and names the account the workbench was recorded under, rather than making a second one.

Who decides what is mirrored

Level You (~/.neuroflow/user.yaml) The team (shared file)
your flowie workbench: on, and the level in workbench_levels — it is yours
a hive the same workbench_mirror in the frontmatter of the hive's hive.md
a project the same workbench_mirror in the project_config.md frontmatter

workbench_mirror is allow, allow-names (adds the names and roles the roster lists — never an email address) or forbid. Without the key, nothing of that hive or project is mirrored: it is the team's decision, recorded in a commit the team can see, and /dashboard --web never writes it for you. It fetches first and follows the stricter of what your clone holds and what the team pushed, so a team's forbid takes effect before you pull, and a value nobody pushed does not start a mirror. Your other keys: workbench_machine (the name each export shows), workbench_index_only (levels whose wiki goes up as an index without page bodies), workbench_account (which recorded workbench this machine uses) and pii_config (the path of your pii_scan.py config — ID patterns and a participant roster — which every export applies; without it, ID numbers and participant names are not checked). The record itself — the page's address, the account label, the page version and a copy of your settings — lives in your flowie's workbench.json, or in user.yaml when no flowie is linked.

What leaves your machine, per source

Everything uploaded comes from this list; python <neuroflow-core>/scripts/workbench.py sources prints the same table, and a test keeps the two equal.

Source Levels Read from What is uploaded
tasks flowie, hive, project tasks/{column}/{slug}.md cards, slug, file, column, title, owner, due, created, updated, phase, tags, project, blocked_by, sha256, notes, legacy, readable — owner only as a roster handle; the first 1,000 characters of the notes; sha256 is the content hash; legacy and readable say whether the file is in an older format or carries a merge conflict
columns flowie, hive, project tasks/config.json columns, id, label, archive, closed, extra, archive_after_days — extra marks a folder that config.json does not name
registry flowie, hive projects/projects.json registry, id, description, status, current_phase, visited_phases, repos, pointer, hive — remote repository URLs only, without a user name, password or query; the first 1,000 characters of a description; a flowie pointer as its id and hive only
wiki flowie, hive, project wiki/pages/**, wiki/index.md pages, id, title, type, tags, status, updated, summary, sources, body — no body with index only; [[links]] resolved to pages; sources, raw/ files and images named, never uploaded
people flowie, hive, project your handle, members.md, collaborators people, handle, name, role — name and role only with allow-names; your own handle only at flowie level; never an email address
ideas flowie, hive ideas.md, ideas-inbox.md (flowie) ideas, ideas_inbox — the text, at most 40 KB
notes flowie notes/*.md notes, file, title, date — never the notes themselves; the newest 200
meetings flowie, hive, project meetings/*.md meetings, slug, title, date — never attendees, agenda or notes; the newest 200
team hive hive.md team — the text after the frontmatter, at most 40 KB: the team's name, description, norms and research directions
errata hive errata/*.md errata, dataset, columns, rows — the table rows (the newest 200 per dataset): date, scope, problem, workaround, reported_by
phases project project_config.md, phase folders project_name, active_phase, recommended_phases, visited, ethics_not_applicable, phases, card — visited names the phase folders that exist, card the project's id in your registry; no other config field
dates project timeline.md dates, date, what, gates — the dated rows; gates names the phase a date gates
integrity project ethics/status.md, preregistration/status.md ethics, prereg, status, expires, frozen_at, by_person — by_person says whether a person set the marker; never approval documents
git flowie, hive, project the clone source, git, branch, head, upstream, upstream_head, ahead, behind, fetched_at, shallow, uncommitted_tasks, machine, applied, state, commit — machine is your label for this computer; applied names your own clicks that reached git, by their Neuroflow-Intent trailers
the level flowie, hive, project the export itself snap_schema, level, part, kind, label, panel, sha256, intent_schema, exported_at, policy, working_tree, index_only, parts, withheld, counts, id, bytes, source, reason, cards, open, pages — the level's name and kind, format versions, the export time, the team's workbench_mirror value, part sizes and hashes, counts, and which sources held items back and why — never the held items or their names

Never uploaded: sessions/, review/, paper/xray-*, wiki/.pending/ and every other local-tier path; integrations.json, user.yaml, local-projects.json, boards/; fails/, finance/, reasoning/; the flowie's profile.md and wellbeing/; wiki raw/ files and images; email addresses, anywhere; meeting attendees, agendas and notes; machine-local paths.

Free text (titles, notes, descriptions, a project's status and phase, column labels, wiki pages with their type and status, ideas, team text, people's names and roles, errata rows, dated rows) goes through pii_scan.py's patterns, without its allowance for addresses the project lists: an email address becomes [email removed], a phone number [phone removed], an ID number your pii_scan config names [id removed], and a machine-local path (C:\Users\…, /home/…, ~/…, as local_paths.py finds them) [local path removed]. An item with secret-like text or a name from that config's participant roster is held back whole, and so is one whose file name, tags or links carry personal data, as they travel unchanged. Your pii_scan config is the file pii_config names in ~/.neuroflow/user.yaml (its id_patterns and roster); without it, ID numbers and participant names are not checked. The preview names each item held back; the page only counts them.

Every upload — the first publish, each refresh, each claim of your clicks, each write-back of outcomes — needs a one-line preview and your yes in that turn. Nothing is uploaded on a timer or at the end of another command. Teammates' text reaches your page as text: it is never run as HTML, and only http and https links leave it.

Limits

  • Not live. A teammate's move appears on your workbench after they push and you refresh. Each tab says what it was made from: working tree on main @ 3f5a2c1 + 2 uncommitted task changes · upstream fetched … · from laptop.
  • Claude Code has to run for a click to reach git; clicks wait until you run /neuroflow:tasks --apply-web.
  • One person, one account. There is no shared board and no read-only page for a supervisor in this version; a teammate sees your changes in git.
  • Phones: drag and drop needs a mouse or trackpad; on a touch screen use Move to….
  • The store's limits: 256 KiB per document, 25,000 documents per workbench, and 50 writes and 1 MiB per batch — an upload is split into as many batches as it needs. Finished requests are deleted after 7 days. A level you take off with --web --levels, or whose team stops allowing it, is purged — the preview names it; a level is never purged only because the machine you run it on lacks it, since another of your machines may mirror it.
  • Big wikis: an export over 8 MB is flagged as slow to load; workbench_index_only sends a level's wiki as an index without page bodies.

When something goes wrong

What happened What neuroflow does
Claude Code is not running when you move cards The cards wait, marked as such; nothing is half done in git
A teammate changed or moved the task after your snapshot A conflict: nothing is written, the card says why and where the task is now
A run stopped half way (closed laptop, a refused approval) The claim expires after 30 minutes; the next run finds what the stopped run already changed — in git, and in a note on your machine that lasts until the commit exists — and commits it instead of repeating it, so an added card is never added twice. A stuck card offers Release, which cancels only a change that was not made yet
You declined the hive push The commit stays local and the card says committed, not pushed; a later push updates it
Another machine of yours exported the level Claude Code asks before overwriting, naming both machines; a clone that is behind is skipped
A newer neuroflow wrote the workbench from another machine It is never overwritten: update neuroflow here first
The page is deleted Only the waiting clicks are lost: /dashboard --web rebuilds it from git and restores your settings
The workbench database is full The page says so; /dashboard --web deletes old requests and purged levels