Changelog

All notable changes to mdnest are documented here.

v4.5.3 — Install with one compose file

Installing mdnest meant cloning the repo, running a setup script and building both images on your own machine. That's a lot of ceremony for anyone who already runs their own networks, reverse proxy and certificates, and wants a compose file they can read and edit. GitHub issue #112 said so plainly, after 30 minutes in the docs without a running install. There is now one file, nothing to build and nothing to clone.

Added

• A plain docker-compose.yml. deploy/compose/docker-compose.yml pulls the published images and runs single-user mdnest with your notes as plain files in ./notes:

  mkdir mdnest && cd mdnest
  curl -fsSLo docker-compose.yml https://raw.githubusercontent.com/mahsanamin/mdnest/main/deploy/compose/docker-compose.yml
  echo "MDNEST_PASSWORD=$(openssl rand -base64 18)"  > .env
  echo "MDNEST_JWT_SECRET=$(openssl rand -hex 32)"  >> .env
  docker compose up -d
  

  Every optional setting is a comment in the file: extra namespaces, the task board, drawings and slides, a reverse-proxy network, and a multi-user block with Postgres. If a secret is missing, compose refuses to start and names the variable, rather than booting with a default password.

• A docker run version, and a guide to both. docs/setup.md now opens with this install. It covers the usual changes in one table, the upgrade command (docker compose pull && docker compose up -d), and which mdnest.conf keys only mean something to the guided setup. The README Quick Start offers it as option A. The guided setup.sh path is unchanged as option B.

Fixed

• The published images now run on ARM. They were amd64 only, so pulling them on Apple silicon, a Raspberry Pi or an ARM cloud host failed with "no matching manifest for linux/arm64". Releases now publish linux/amd64 and linux/arm64. The builds cross-compile natively rather than under emulation, and local setup.sh builds are unaffected.
• The clone URL in docs/setup.md pointed at the wrong GitHub owner.

Tests

• tests/compose-example.sh runs in the pre-push hook. It checks that every setting the compose file offers is one the backend actually reads, that nginx's proxy target backend is a service in it, that both images share a tag, that the file isn't gitignored, that releases publish arm64, and that the documented one-liner downloads the right path. With Docker present, it also checks the missing-secret refusal. Each check was confirmed to fail when its condition is broken.

v4.5.2 — A task board that does what you meant

Dragging a card between columns worked for exactly one gesture: grab the thin title strip, release squarely over another column's cards. Everything else failed quietly or, worse, landed somewhere you didn't aim. The second move from the same note was thrown away with a flash of "Loading tasks…". A refresh on the board dropped you back into the editor, and switching to the List view on a big workspace froze the app for ten seconds. All of that is fixed, along with one byte the CLI added to every note it read.

Fixed

• Grab a card anywhere. Only the title strip used to start a drag, although the whole card wore a grab cursor. The whole card is the handle now; its buttons and step checkboxes still click.
• The column under the pointer is the target. It used to be the column the card's rectangle overlapped most, and a column is only as tall as its cards — so a release below a short column's last card, or on the collapsed Done strip, hit nothing and the card snapped back. Anywhere in a column's lane counts now.
• The board no longer slides sideways mid-drag. dnd-kit measured its auto-scroll edge zone against the source column's box while scrolling the board, so leaving the starting column counted as "at the edge" and the card landed one column over. Its auto-scroll is off; the board scrolls only when you hold a card at its own left or right edge.
• The second drop from a note lands too, without a reload flash. A move writes a status: line under the task, shifting every task below it; the board kept the old line numbers, the server rightly refused the next move, and the board answered with a full reload that discarded it. A refused move now re-finds the task and retries quietly, moves save one at a time in drop order, and cards keep a stable identity across refreshes — a refresh landing mid-drag used to bind the drag to a different card.
• The task board has a URL — #!board/<workspace>/<note> — so a refresh keeps you on it, and a board can be bookmarked or shared. The link keeps the note underneath, so the back button and This note scope survive a reload.
• The List view is instant on a big workspace. It put every task into the page (6,327 rows, ~39k elements on the sample project) and froze the app for ~10 seconds. It now shows 200 at a time with a Show more button, like the Kanban columns' 100. Filters still cover every task.
• mdnest read prints a note byte for byte. It always appended a newline, so a note that already ended in one — anything saved by write or the web UI — read back with an extra blank line, and mdnest read | diff - note.md failed after every write though nothing had changed.

Security

• gRPC bumped to v1.83.2 (GO-2026-6443, GO-2026-6441, GO-2026-6348, all in google.golang.org/[email protected], published after v4.5.1; v1.83.1 fixes two of them, 6443 needs v1.83.2). gRPC arrives indirectly through the Firebase / Google Cloud client libraries; the bump pulls matching minor versions of OpenTelemetry, genproto and golang.org/x/{crypto,net,sys, sync,text} with it. govulncheck -mode=binary (what CI and the pre-push hook run) reports no vulnerable code reachable from mdnest after it.

Tests

• board-drag.spec.js drags by the title (both scopes), by the card body, into the empty lane, onto collapsed Done, to an off-screen column via the edge, and three cards from one note back to back. Every case also fails if any other task anywhere changed — an early flaky version of the back-to-back case moved real cards in the dev instance's sample data, and nothing noticed.
• board-deep-link.spec.js, a List paging case in board-scale.spec.js, and a byte-exact read check in cli-smoke-test.sh.
• tests/e2e-browser.sh and tests/e2e-docker.sh mount their throwaway notes directory from inside the repo instead of the system temp dir, which Colima does not share — the run died at "could not seed note" before a single test ran.

v4.5.1 — The CLI installer survives GitHub's CDN

curl -fsSL https://raw.githubusercontent.com/.../install-cli.sh | bash was returning 503 and nobody could install the CLI. The repo was fine, GitHub was fine, and the network was fine — the 503 came from Fastly's own edge: Backend.max_conn reached, served by one POP, which takes out every install in that region for as long as it lasts. Both the installer and mdnest update were hardcoded to that single host with no fallback and no retry, and reported the failure as "check your network" — sending people to look at the one thing that was working.

The install command is now:

curl -fsSL https://mdnest.dev/install.sh | bash

The GitHub URL is the same script and keeps working.

Fixed

• The installer and mdnest update try three independent hosts, twice each — GitHub raw first, then jsDelivr, then mdnest.dev. GitHub stays first deliberately: it publishes the instant a fix lands on main, whereas jsDelivr caches a branch ref for hours, and the CLI is pull-only so an update that arrives half a day late is its own problem.
• A failure now names itself. Every source's actual HTTP status is printed, plus a line saying that a 503 from raw.githubusercontent.com is GitHub's CDN rather than your network or the repo. Same rule as curl_reason and the pre-push audit check: keep the reason when the reason is what the reader has to act on.
• A downloaded CLI is verified before it replaces a working one. The old check was a shebang, which a captive-portal page and a truncated download both pass; it now also requires the MDNEST_CLI_VERSION marker. mdnest update also downloads once instead of twice — it used to fetch the file to read the version and fetch it again to install it, doubling the exposure to exactly this outage and leaving room for the two to disagree.
• mdnest.dev mirrors the CLI, pulled from main by the site's own deploy step rather than copied by hand — a mirror that can drift would hand people a stale CLI precisely when the canonical source is unreachable and nobody could tell.
• The installer's closing hint is pasteable. It printed mdnest login <server-url> <api-token>, and <server-url> is a shell redirection — the instruction meant to get you started was the next thing to fail. The project already had this rule for CLI output and the web UI; the installer sat outside both.

Security

• Two transitive high advisories, neither from this change. They are here because --audit-level=high is a required check on main, so they blocked the hotfix outright.

  • js-yaml 4.3.1 -> 4.3.2 (GHSA-2883-xcg3-v3hh), reached via @marp-team/marpit. An in-range lock-file bump, no package.json change.

  • @xmldom/xmldom forced to ^0.9.12 with an overrides entry (GHSA-6gmq-8vp8-gcm6 and twelve siblings), reached via speech-rule-engine. v4.4.0 deliberately left this one alone while it was moderate — below the gate — because the override then made the tree invalid to npm's legacy quick-audit endpoint. The re-rating to high met the stated condition for revisiting it, and under CI's actual environment (node:20, npm 10.8.2) the override no longer breaks the audit. [email protected] still pins 0.9.10 exactly and both marp-core and mathjax-full are already at their latest, so there is no other route.

    Two things made it safe to take rather than merely necessary: xmldom and speech-rule-engine are tree-shaken out of every shipped chunk, so the browser bundle does not change at all; and Marp still renders — verified by rendering a deck with front-matter, pagination and math through marp-core in Node, since the unit tests cover our own Marp detection and not marpit's YAML parsing, which is what the js-yaml bump touches.

Notes

• tests/cli-unit.sh gains a source-chain suite: there must be more than one source, they must be on genuinely different hosts, GitHub must stay first, mdnest.dev must not be offered for a non-main branch, and the two copies of the list — the CLI's and the installer's, which cannot share code because the installer is fetched on its own — must match. The errexit lint now covers install-cli.sh too, since it runs under set -e and is no longer a straight-line script.

v4.5.0 — Stickies

A personal sticky board, kept beside your notes and never mixed in with them. Open it as a drawer on the right to jot something down without leaving the note you are reading, or full screen as a corkboard where cards are dragged and resized freely. Each card is a title, some free text, and a checklist — all optional, so the same card covers a scribbled reminder and a small to-do list.

Stickies are deliberately not notes. They never appear in a namespace and never reach a git remote: they live in mdnest's secrets volume (Postgres in multi mode), which the git-sync sidecar cannot see. That storage location is the privacy guarantee, which is why there is nothing to encrypt and no key to manage. The tradeoff is stated plainly in the app and the docs — a sticky has no git history and no off-server copy, so anything you would be upset to lose belongs in a real note.

Nothing about the default install changes: no new dependency, no new environment variable, no new failure mode for an operator who never opens it.

Added

• Sticky notes — GET/PUT /api/stickies, available in both auth modes. Each user gets exactly one board, keyed server-side from the authenticated identity: there is no user id, path or namespace in the request, so reading someone else's board is not a check that can be forgotten, it is not expressible. Boards are stored in Postgres (user_stickies, migration 016) in multi mode and in stickies.json in the secrets volume in single mode — the same volume as auth.json and tokens.json, so a board survives ./mdnest-server rebuild and stays invisible to git-sync.
• The drawer — a right-edge panel sharing geometry and remembered width with the comments sidebar; only one of the two is open at a time. Its open state is remembered per browser, so a refresh keeps it, but it stays out of the URL: a shared note link should never force someone else's stickies open.
• The full board — a corkboard at #!stickies with its own URL, so a refresh returns to the board rather than to the last note. Drag a card anywhere, resize it by its bottom-right corner, or use Tidy up (which asks first) to put everything back on the grid. On a phone it falls back to a flowing grid; free positioning on a 380px screen is a board you have to pan around to read.
• Checklists — "done" lives on the item, not the card. A single card-level flag forces "buy milk, call bank, post form" to be either three separate notes or one note you can only tick when all of it is finished. Enter opens the next line, Backspace on an empty one removes it, and long to-dos wrap.
• Limits, enforced server-side — 200 stickies per board, 200 bytes of title, 4 KB of text and 50 checklist items per card, a five-colour enum, and a 150–600px card width. The endpoint is writable by any authenticated user, so without them a board is an unbounded per-user blob store; this is the same reasoning that gave preferences a key allowlist.

Notes for operators

• No configuration. There is no flag to turn stickies on, no env var to set, and setup.sh, docker-compose.yml, the Dockerfiles, nginx.conf and mdnest-server are untouched. The cost to an install that never uses the feature is +1.7 KB of JavaScript and +0.7 KB of CSS, gzipped, and one Postgres table in multi mode.
• Boards are not backed up. This is deliberate and is the flip side of the privacy guarantee: a sticky has no git history and no off-server copy. It survives a rebuild because the secrets volume is declared, but not the loss of that volume.
• A board that cannot be read is an error, never an empty board. Because a board is written back whole, a "gentle" empty result on a failed read would be replaced by the next keystroke. Both the client and the server refuse to save until a read has actually succeeded, so unreadable data is left in place to be recovered by hand.

v4.4.0 — Edits that don't overwrite someone else

Changing one line in a note used to mean rewriting the whole file. Both surfaces agents reach mdnest through — the MCP server and the CLI — could only replace a note wholesale, so the only way to edit part of one was to read it, rebuild it, and write it back. Anything saved in that window by the web UI, git-sync, or another agent was silently overwritten, and the write reported {"status":"ok"} while doing it.

Both now have an exact-string edit that carries the version it read, so a concurrent save comes back as a 409 with nothing lost. Plus the Preview pane renders images again.

Added

• edit_note MCP tool — replace an exact string in a note instead of rewriting the whole file. Zero matches, or more than one without replace_all, is an error naming the count rather than a guess; the replacement is spliced literally, so $& and $1 in a pasted shell snippet stay as written; and the write carries If-Match, so a save that landed since the read surfaces as a 409 instead of being clobbered. Contributed by Luigi Lotito (@lglot) (#107).
• mdnest edit in the CLI — the same capability, for the same reason, on the surface that had the same gap. Changing one line used to mean read, rebuild the note, write it back, which silently overwrote anything the web UI, git-sync or another agent had saved in the meantime — and reported {"status":"ok"} while doing it. edit matches literally, refuses an ambiguous match unless you pass --replace-all, and guards the write with the version it read. See docs/cli.md.

Fixed

• Preview showed every image broken. A relative ![](shot.png) resolved to /api/files/… but carried no ?token=, and a browser <img> GET cannot send an Authorization header — so each one 401'd. The Live editor had always appended the token; the Preview renderer never did. The token is attached only for our own /api/files path and never rides along to a foreign host. Contributed by @bilal-wego (#108).

• An image whose filename began with a scheme name never loaded in Preview. The absolute-URL test was a startsWith('http') prefix check, so an ordinary uploaded http-flow.png was treated as an external URL, never got its /api/files/ prefix, and rendered broken in Preview while working fine in the Live editor. Both renderers now share the Live editor's test, which also covers blob: and uppercased schemes.

• A note whose content started with @ could not be written at all. The CLI passed note bodies to curl -d, where a leading @ means read this file — so mdnest create/write/append/prepend on a note beginning @mention … failed with curl's exit 26, reported as "couldn't reach the server". Present since v1.0. Now --data-raw, which is -d without that special case.

• The pre-push hook called an npm outage a vulnerability. npm audit exits non-zero both when it finds advisories and when it cannot reach the advisories endpoint, and the hook discarded stderr and treated every non-zero exit as VULNERABILITIES FOUND — so a run of 503s from /-/npm/v1/security/advisories/bulk blocked the push while naming the wrong cause, and npm audit fix silently applied nothing for the same reason. An unreachable endpoint now SKIPs with the reason stated, the way the hook already handles govulncheck without a host Go toolchain; CI's required checks remain the authoritative gate. A real finding still blocks — all seven outcomes are probed in tests/pre-push-audit.sh.

• Four transitive security advisories. qs 6.15.2 -> 6.16.0 and fast-uri 3.1.5 -> 3.1.7 (mcp-server), browserslist 4.28.2 -> 4.28.8 (frontend) — all three in-range lock-file updates, no package.json change. None of these came from this release's own changes — they are newly published advisories against dependencies that were already there, and both npm audit jobs are required checks, so browserslist (the only high) would have blocked the release PR to main outright.

  @xmldom/xmldom (moderate, GHSA-6gmq-8vp8-gcm6, reached via speech-rule-engine) is knowingly left in place. There is no semver-compatible fix — speech-rule-engine declares it as exactly 0.9.10 — and an overrides entry forcing 0.9.12 was tried and reverted: it makes the tree invalid to npm's legacy quick-audit endpoint, which then refuses the whole audit (400 Bad Request … Invalid package tree). That trades one moderate advisory for losing the audit signal entirely, which is strictly worse. Both audit jobs run --audit-level=high, so a moderate is below the gate by deliberate policy (see the severity note in security-audit.yml). Revisit if it is ever rated high or speech-rule-engine relaxes the pin.

• The errexit lint flagged arithmetic as a command substitution. c=$((c + 1)) is arithmetic expansion and carries its own exit status, but the lint's pattern saw the leading $( and reported it. Its self-proof now includes an arithmetic line, so a relapse shows up as a miscount.

v4.3.3 — Instructions that survive a paste

A short one, entirely about the second half of the v4.3.2 report: someone hit a CLI bug, and nothing in the product told them their CLI was stale or how to fix it. v4.3.2 taught the CLI to say so. This does the same for the app.

Fixed

• The commands in Settings → CLI can be pasted. Nine of them carried <your-token> or <namespace>, and < is a shell redirection — pasting mdnest login https://notes.example.com <your-token> into zsh or bash gives no such file or directory: your-token. Every one of those blocks has a Copy button that reproduces the text verbatim, so the instruction meant to get someone started was the next thing that broke for them. They use literal stand-ins now (mdnest_yourtoken, notes).

  The CLI learned this in v4.1.3 and tests/cli-unit.sh has asserted it for CLI output ever since. The web UI was simply never covered by that rule — which is the more interesting failure: the convention existed, was written down, and had a test, and the surface that most needed it sat outside the test's reach.

Added

• "Keeping it up to date" in Settings → CLI. It states the thing nobody had been told: the CLI does not update itself and nothing pushes to it — it is a script on your machine. It shows mdnest update and mdnest version, and names the version this server is running, so there is something concrete to compare against rather than a vague suggestion to check.

Guarding it

frontend/src/__tests__/pasteable-commands.test.js fails the build on a new bracketed command. Two things it has to get right, both found by testing the guard itself against the pre-fix file rather than trusting it:

• A shell block is checked in full, not line-by-line from a mdnest-prefixed start. The Copy button copies the whole block, and one of the nine offenders began echo "text" | mdnest append <namespace>/log.md — which a "line starts with mdnest" rule walks straight past.
• JSON config blocks are excluded. The MCP tab's claude_desktop_config.json mentions node and mdnest, but it is pasted into a file, not a shell, where <your token> is a perfectly good placeholder. Flagging it would have been a false positive that teaches people to ignore the check.

Verified both directions: all nine flagged against the old file, and the false positive stays quiet.

v4.3.2 — The CLI stops dying quietly

One bug report, one root cause, five places it was hiding. mdnest servers printed the table header and then nothing at all — no rows, no error — and exited with curl's 28, any time a registered server was unreachable. Since the server list is globbed alphabetically, one dead server also hid every healthy server sorting after it, so a mdnest login that had worked perfectly looked like it had failed.

Nothing was wrong with the networking, and nothing was wrong with the error handling either: the labels, the hint, and the messages had all been written. They were simply unreachable code. The CLI runs under set -e, and in bash a plain assignment from a command substitution takes the substitution's exit status — so cfg=$(curl ...) killed the whole script the moment curl could not connect, mid-loop, before the first row. The leaked exit code was the tell: a command that should always exit 0 was handing back curl's 28.

The same shape turned up in four more places, each one silently killing the CLI in place of an error message that already existed.

Fixed

• mdnest servers lists every server when one is unreachable. The probe assignment is guarded (... || curl_rc=$?), so the loop survives a failed connection and prints the row it was always meant to: the URL, the unreachable (DNS | refused | timeout | TLS) label naming the actual reason, and the "works in your browser?" recovery hint. A dead server no longer hides the ones after it, and the command exits 0. The trailing curl_rc=$? is deliberately gone — left in place it overwrites the real curl code with the status of the now-successful assignment and reports unreachable (curl 0).
• Reads and writes against an unreachable server say so. api() had the identical unguarded assignment, so mdnest read @unreachable/ns/x.md printed nothing whatsoever and exited 28. Its three error messages — can't resolve the host, connection refused, connection timed out — now actually run.
• mdnest servers -v survives a failed namespace probe rather than aborting partway through the listing.
• A broken jq yields an empty field, not a dead CLI. json_top_string's jq tier returned jq's exit status from a bare return, which aborted the caller's assignment instead of falling through to an empty value.
• mdnest list <missing-folder> reports the missing folder on a fresh machine. The pure-awk tier — the one that runs when neither python3 nor jq is installed — signals "not found" with exit 3, and unguarded that killed the CLI two lines before the path not found in namespace message. It printed nothing and exited 3. Same for the jq tier.

Testing

• tests/cli-unit.sh gains an unreachable-servers suite that runs the real CLI against a throwaway HOME pointed at two closed loopback ports — instant, no network, no Docker — and pins the exit status, one row per registered server, the label, the hint, and api()'s error text. Ten of its eleven checks fail against the unpatched CLI. The eleventh exists to catch the partial fix that guards the assignment but leaves the stray curl_rc=$? behind.
• tests/cli-smoke-test.sh now requires the missing-subfolder case to say why it failed, not merely to exit non-zero. The bare assert_fails passed the entire time the awk tier was dying silently — a reminder that asserting a non-zero exit proves nothing about whether the error path ran. tests/e2e-docker.sh runs this suite in a bare alpine with neither parser installed, which is where that tier gets exercised.

Added

• The CLI tells you when it is out of date. One line, wherever you are already looking at versions — mdnest servers, mdnest whoami, mdnest login — naming both versions and the exact command:

    Your mdnest CLI is v4.3.1; @work is running v4.3.2.
    Update it with:  mdnest update
  

  This is here because of what the bug above revealed rather than the bug itself. mdnest update is pull-only — nothing pushes it, and the in-app banner tracks the server, not the CLI. The only version check the CLI ever had was a major-version mismatch at login, so a client could sit on a stale point release indefinitely without a word. That is precisely how a bug introduced in v1.0 stayed invisible through 47 releases: everyone running it was told nothing.

  Three deliberate limits, stated rather than discovered later. It is not printed on every command — the CLI keeps no update cache, and a check on every read would be its own bug. It does not contact GitHub; the version comes from the /api/config call the CLI already makes, so there is no extra round-trip and no new failure mode — the trade-off being that if your server is also out of date, nothing tells you. And it never nags a pre-release about its own release: 4.3.2-dev is older than 4.3.2 and newer than 4.3.1, matching the in-app banner's isVersionNewer.

  version_gt() is pure bash — no python3, no jq, and no sort -V, which busybox sort does not have — so it works on the same fresh-machine tier as everything else in the CLI.

Guarding against the next one

The bug above was one root cause in five places, and it had been in the codebase since the first release. Two things now make a sixth harder.

• Every plain command-substitution assignment in the CLI is guarded, and tests/cli-unit.sh fails the build on a new unguarded one. The lint proves itself against a probe file — one unguarded form flagged, three guarded forms not — because a green lint that cannot fail is worse than no lint. It exempts local x=$(...), which was verified against bash rather than assumed: local is a builtin, so the assignment's status is the builtin's own and errexit does not fire.
• The lint is scoped to mdnest, not everything. That is the script downloaded onto other people's machines, where a silent death is invisible. mdnest-server runs on the operator's own box and still has unguarded sites — a separate audit, named here so it is a known gap rather than a quiet one.

A third bug fell out of building the notice, and it is worth naming because it had been sitting there quietly: mdnest login read the server's version with grep -o '"version":"[^"]*"', and /api/config carries two matches — the top-level version and latestRelease.version, the nested one emitted first. The value was two lines, not one. Nothing noticed, because the only consumer was ${SERVER_VER%%.*}, which still yields 4 from a doubled string. The moment this release started printing and comparing that value, login began reporting is running v4.3.1\n4.3.1. and the comparison read the third field as 14, telling a 4.3.2 CLI to update to a 4.3.1 server. json_top_string is the depth-aware parser written for exactly this — its own header comment names latestRelease.version as the trap — and the login path simply was not using it. Fixed at the source, and version_gt now takes only the first line of each argument so a sloppy caller cannot fabricate an upgrade either way.

Worth recording, because it is the honest result: while writing the update notice, the same class of bug was reintroduced inside the fix for it — a trailing [ -n "$x" ] && … as a function's last statement, where a false test becomes the return value and set -e exits the CLI with 1. The lint did not catch it; a lint cannot see that shape. The unreachable-servers suite added earlier in this release caught it within a minute. Which is the actual lesson: the behavioural test is the guard, and the lint is the cheap second net.

Reported against v4.3.1 on Linux. Not platform-specific — it is shell semantics, and it reproduces anywhere with a server pointed at a blackhole address.

v4.3.1 — Text you can actually read

v4.3.0 added a light theme. Shipping a second theme turns out to reveal every place the first one was getting away with something, and this release is the cleanup: four separate cases where mdnest drew text you could not read, three of them found by simply looking at the app in the other theme.

The Mermaid ones are the worst of them. A diagram with an author's own classDef fill: — the ordinary way anyone colours a flowchart — rendered its labels in the light ink while in dark mode, so the text was near-invisible on its own node. That is not a subtle miscalculation; the brightness maths was right all along, and the walk that fed it was measuring the wrong shape.

Fixed

• Mermaid labels are readable on coloured nodes, in both themes. Mermaid nests an empty <rect> spacer inside every flowchart node's label group. It paints nothing, but it inherits the themed mainBkg — #313244 in dark — and that was the first shape the contrast pass found. So a pale #cfe4ff node was told its background was dark and got light ink. Light mode was correct only by luck, because its mainBkg happens to be bright too. The pass now measures each candidate before trusting its colour and skips anything with zero area. Measured on a pale node, the ink/fill luminance gap goes from 20 to 193 in dark and 144 in light.
• Mermaid edge labels too. An edge label has no shape at all — Mermaid paints it with a CSS background-color on the HTML inside the foreignObject, which the pass never looked at, so it climbed past the chip to an unrelated node. Brightness is now composited over the diagram's own ground as well, because that chip is half-transparent and judging the declaration instead of the pixel got the answer right only by accident. Gap: 20 to 173.
• Commented text is no longer white on yellow in light mode. A commented passage is painted with --highlight (a bright yellow in both themes) and was inked with --text-inverse, which means "text that sits on the accent" — #1e1e2e in dark, #ffffff in light. So the light theme put white on yellow at 1.32:1. The ink has its own token now.
• The toolbar no longer draws its controls on top of each other. On a narrow editor — a 13" laptop, or any window once the comment panel takes its 330px — the bar ran past its container and painted the filename over the comment button, and Rename/Delete over the theme and settings icons. It is a single non-wrapping flex row whose groups are all flex-shrink: 0, so the only thing that could give was the path in the middle, and it could not. The filename now ellipsizes, the path clips instead of painting outside itself, and the bar wraps rather than overflowing.

Documentation

• The README leads with what mdnest is rather than a feature list, and every screenshot is new — taken at the size it renders, and shipped light and dark through <picture> so it follows the reader's theme. It also gains a diagram of the browser, the CLI and an AI agent all reaching the same files, because most people meet mdnest as a web app and never learn there is a CLI.
• The CLI section teaches the current syntax. It documented mdnest note list, the legacy form, which hides the whole point: one machine holds as many servers as you like and addresses them as @work/… and @home/… from the same shell.

Testing

• mermaid-contrast.test.js and theme-contrast.test.js pin the decisions behind the fixes above; tests/browser/mermaid-contrast.spec.js and toolbar-fit.spec.js pin the rendered result in a real browser, in both themes. Each was confirmed to fail against the pre-fix code rather than assumed to.
• Two notes for whoever writes the next one of these. The overlap check skips nested pairs — Rename and Delete live inside .toolbar-path, and a parent enclosing its children is not a collision — and it confirms each overlap by hit-testing the intersection, because a control clipped by an overflow: hidden ancestor still reports its full rectangle while painting nothing.
• The browser E2E stack now sets ENABLE_LIVE_COLLAB, and its Mermaid fixture carries a pale-filled node and a labelled edge — the two shapes that were broken.

v4.3.0 — Light mode

mdnest has been dark-only since the first commit. Some people don't want that, and until now the only answer was "use a different app after sunrise".

This release adds a light theme that follows your operating system by default, a toggle that takes one click, and — the part that matters more than it sounds — a choice that is stored against your account rather than your browser. Pick light on your laptop and your phone agrees.

It also takes a pass at the top toolbar, which had quietly become a wall of equally-spaced buttons.

Added

• A light theme. Catppuccin Latte alongside the existing Mocha, covering the whole app: sidebar, editors, task board, drawings, diagrams, modals, and the native scrollbars and date pickers the browser draws for us.
• It follows your system by default. New installs start on auto, which tracks your OS light/dark setting and changes with it live — no reload.
• A toggle in the toolbar. The sun/moon button top-right flips light and dark; the icon shows the theme you would switch to.
• Settings → Appearance carries the full three-way choice — Match system, Light, Dark — because "follow my system" is a third state one button cannot express without becoming a menu.
• Your theme follows you, not your browser. It is stored server-side: in Postgres in multi-user mode, in the secrets volume in single mode. A new browser, a phone, or a fresh private window gives you the theme you picked.
• DEFAULT_THEME in mdnest.conf (auto | dark | light) sets the starting point for people who have never chosen. It is a default, not a lock — any user's own choice overrides it. Also available as ui.defaultTheme in the Helm chart, and rejected at install time if it isn't one of the three.
• GET/PATCH /api/preferences — per-user UI settings, available in both auth modes.

Changed

• Drawings follow the app theme instead of remembering their own, and repaint as soon as you change it. The canvas had its own light/dark button, which existed because mdnest was dark-only and a drawing had no other way to be light. Now that the app has a theme, that button was a second control doing almost the same job a few centimetres from the first, so it is gone.
• The toolbar groups its controls. Every button used to sit the same 8px from its neighbour, so "Rename" was no more visibly related to "Delete" than to the filename beside it. Related controls now sit close together with wider space between groups, and two dividers separate the actions that change a file from the ones that just navigate to it. No buttons were added or removed.
• Slide decks keep their own theme. A Marp deck is something you authored to look a particular way, and it renders that way regardless of your app theme.

Fixed

• @milkdown/crepe is now a declared dependency. The Live editor imported it directly while it resolved only as a transitive dependency of @milkdown/react. A lockfile regeneration or a Milkdown bump could have removed it and broken the editor build with an error pointing at node_modules rather than at the missing declaration.
• The Settings tab row wraps instead of pushing "Credentials" off the edge of the dialog.

Under the hood

• Colour is now a two-layer token system (frontend/src/theme.css): a raw palette, and semantic tokens that every stylesheet uses. 930 hex literals were replaced. The indirection is what makes a second theme possible at all — mdnest used #313244 as a background 78 times and as a border 102 times, and in a light theme those must diverge, since a border has to be darker than what it encloses while a raised surface stays lighter than the page.
• The light palette is measured, not eyeballed. Stock Latte tunes its accents for use as accents; mdnest uses them as body text. Latte yellow measures 2.31:1 on the page background against a 4.5:1 AA floor. Five hues are darkened to the first step that clears AA both as text and under white. 63 contrast assertions run in the test suite, including one that light is never less readable than dark for the same pair.
• New tests: theme resolution and its precedence, the token layer's completeness, contrast in both themes, that setup.sh actually delivers DEFAULT_THEME to the container, and six browser specs — one of which clears the browser entirely and proves the theme still comes back.
• tests/setup-marp-themes.sh, added in v4.2.0 but never invoked by anything, now runs in the pre-push hook alongside the new setup test.

Upgrading: nothing to do. Existing installs keep dark unless a user chooses otherwise or you set DEFAULT_THEME. Multi-user installs pick up one new table (user_preferences) automatically on first start.

v4.2.2 — A board you can actually use

Almost all of this release came out of running the task board on a real project with roughly 12,000 checkboxes, where it went from useful to unusable. None of it was where it looked: the server answered in ~100 ms, and the cost was the browser being handed every card at once.

The rest is the board finally saying what it is. It has been a third button inside the Basic/Live control, then a row in the sidebar, and is now one button that names where it takes you — with no way out of it until this release, which is the part that should not have shipped in the first place.

Fixed

• One toolbar button now swaps between your note and the board. On a note it reads Board; on the board it reads Editor and brings you back — the label always names where it will take you. Previously the board was a third button inside the Basic/Live control, where all three read as one choice, even though Basic and Live are ways of editing the file you have open while the board leaves the file entirely. Basic/Live are hidden while the board is open, since there is no note on screen for them to act on, and so is the "No file selected" placeholder. The board's own header still starts with a back button naming the note you came from. The sidebar is now purely your files.
• The sidebar's create buttons read Folder, Note, Drawing — containers before the things that go inside them.
• Card order is now yours to choose, which matters once a column pages. Tasks arrive in note order, so with a column painting 100 cards at a time an overdue task in a late-alphabet file sat on page 64 with no way to know it was there. A Sort control in the filter bar offers due date, then priority, and remembers the choice. The default stays note order deliberately: it is what the board has always shown, it mirrors the files the tasks live in, and most boards never page at all — quietly reshuffling them would be its own kind of broken. Sorting adds no measurable cost (typing measured at 86 ms in note order and 74 ms by urgency on a 12,000-task board).
• The task scan is cached, so the board stops re-reading every note. Every board request read and parsed every note in the namespace. Measured on 420 notes holding ~12,000 checkboxes: walking the tree costs ~6 ms while reading and parsing costs ~100 ms, so the parse is now remembered and only the walk repeats — a warm request drops from ~130 ms to ~20 ms, and "All workspaces" to ~18 ms. It cannot go stale on its own: every request still walks the namespace and a cached answer is used only when the file set, sizes, newest timestamp and column layout are all unchanged, so an edit from the API, the CLI, git-sync or an editor on the host invalidates it without being told. Refresh sends refresh=1 and re-scans regardless. The cache lives in memory, not in a file under your notes — a cache file there would be synced by git-sync and destroyed by a rebuild.
• "All workspaces" asks before it scans. It reads every note in every workspace you can see, which is the one board action that can visibly pause on a large install. It now explains that before starting, with a Don't show this again that is remembered.
• Presence no longer flickers, and no longer moves your work. With more than one person on a note, collaborators appeared and disappeared repeatedly. Two separate faults. Departures were applied the instant they arrived, so any momentary gap — a reconnect, a presence snapshot racing a join — read as someone leaving and immediately returning; a departure is now held for a few seconds and cancelled outright if they come back, so a blip never reaches the screen while a real exit still registers. And the presence bar sat in the document flow as a full-width strip, so each of those blips reflowed everything below it — the editor, a Mermaid diagram, a drawing canvas all jumped. It is now an overlay pinned inside the content area, so presence can come and go without the page moving. This affected every note, not just drawings.
• The task board stays usable on a big project. Enabling the board on a namespace with ~12,000 checkboxes made it crawl. The server was not the problem (it answers in ~100 ms); the browser was being handed every card at once. A column rendered its entire contents, so ~6,300 cards and 50,649 DOM nodes went onto the page, and every keystroke in the filter box re-filtered and re-rendered the lot — 456 ms per key. Columns now paint 100 cards at a time with a Show more button, cards are memoised so one change no longer re-renders its neighbours, and filtering follows a deferred value so typing stays responsive. The column header still reports the true total, so nothing is hidden silently. Measured on the same 11,994-task namespace: board opens 1,802 ms → 844 ms, DOM 50,649 → 950 nodes, filter keystroke 456 ms → 87 ms.
• The task board has a visible way out. It replaces the editor pane, and its onClose was accepted but never rendered — so once you were on the board, the only way back to your note was the toolbar's Basic/Live pair, which says nothing about the board. The header now starts with a back button naming the note you came from, and the sidebar entry toggles the board closed again.
• A missing build asset now 404s instead of being served the app shell. The SPA fallback also caught /assets/, so a content-hashed chunk that no longer existed was answered with index.html and a 200. A tab still running a previous deploy asked for filenames that had been replaced, received HTML where it expected a JavaScript module, and broke in a confusing half-alive way instead of failing cleanly — and the chunk retry could not rescue it, because the cache-busted URL returned the same HTML. Unknown routes still serve the app shell; only /assets/ is strict.

v4.2.1 — Nothing silently disappears

A bug-fix release with one theme: mdnest should never lose your work or hide it somewhere you can't reach. Every fix here is something that failed quietly — no error, no indication anything had gone wrong.

Fixed

• A drawing no longer loses its last strokes when you switch files. Opening another note cancelled the previous file's queued autosave outright, so any edit made inside the debounce window was discarded. Drawings hit this on almost every switch: the canvas debounces its own scene for 500ms before handing it to the app, and you stop drawing at exactly the moment you reach for the next file — a newly created drawing could stay 0 bytes on disk. The pending save is now flushed rather than dropped, before the next note loads (the etag is shared, so the order matters). The drawing canvas hands over its own debounced scene at the same point, and a stray post-unmount timer that could push one file's content at another is cleared.
• A failed code-split chunk no longer blanks the whole app. A lazily-loaded editor whose chunk didn't download threw during render, reached React's root and unmounted everything — no sidebar, no toolbar, no way to open another note. Only the Live editor was guarded; drawings, the task board and the slide renderer had a bare <Suspense>, which handles a pending import but not a rejected one. They now show an error with a way out, and a failed import is retried — the last attempt with a cache-busting query, which recovers from a proxy that cached an error response for an immutable asset URL.
• Dropping a file into an open folder puts it in that folder. Only the folder's own row accepted a drop; its expanded contents area had no handlers, so a drop there bubbled up and was treated as "move to the namespace root" — aiming carefully inside a folder moved the file out of it. Dropping onto a file did nothing at all, because a file row swallowed the event before bailing.
• The task board scrolls sideways instead of hiding columns. With enough columns the right-hand ones became unreachable, with no scrollbar. The board panel had no min-width, so it refused to shrink below the intrinsic width of its columns, grew past the viewport, and was clipped by its container — the scroll container was never smaller than its own contents.

Changed

• "+ Note", "+ Drawing" and "+ Folder" create where you are. They always created at the namespace root, however deep in the tree you were. Clicking a folder now aims them at it — shown in the tree and named in each button's tooltip, since the destination was otherwise invisible — and opening a file hands the target back to that file's folder.
• "Basic" now does something on a drawing. A .excalidraw.md is a real markdown file, but the canvas short-circuited the editor entirely and the Basic/Live buttons sat there inert. The toggle now offers the two views a drawing actually has: the canvas, or the markdown behind it. Live is deliberately not offered — the rich editor would reformat the scene JSON, the same hazard that already forces Marp decks to raw editing.
• Drawings open in dark mode, matching the rest of the app, with a toggle to light. Theme is a viewing preference: it is stored per browser and never written into the note, so the file stays portable and two people can view one drawing differently.
• The task board moved out of the toolbar's Basic/Live control into the sidebar. Those are two different things — Basic/Live is how you edit the open file, the board is a namespace-level destination — and sharing one segmented control made all three read as the same kind of choice.

Security

• Replaced github.com/lib/pq with pgx. govulncheck reports seven advisories against lib/pq (GO-2026-6166, 6168, 6170–6173) — a malformed-frame panic, unbounded SCRAM iteration, and GSS authentication completing without mutual proof. All are Fixed in: N/A: v1.12.3 is affected too, so there is no patched release to move to, and lib/pq is in maintenance mode. Only the Postgres driver changes; the connection string, the SQL, and single mode (which never opens a database) are untouched. golang.org/x/text is bumped to v0.39.0 to clear the one transitive advisory pgx introduced. The backend now scans clean.

Fixed (continued)

• The namespace root is a row in the tree. Aiming the create buttons at the selected folder (above) made selecting one a one-way trip — nothing pointed them back at the top level — and left the root with no drop target beyond whatever blank space remained under the tree, which is none once the tree fills the panel. Click the root row to create at the top level, or drop onto it to move something there.

Testing

• Browser regression specs for every fix above, each written to fail on the previous build.
• The E2E stack now enables drawings (ENABLE_EXCALIDRAW). It only set ENABLE_TASK_BOARD, so every drawing spec skipped — including the one pinning the data-loss bug. A regression test that never runs in the gate is not a gate.

v4.2.0 — Drawings, access groups, and per-note authorship

The largest release since v4.0.0, and almost all of it is contributed work. Three new opt-in capabilities — Excalidraw drawings, a centralized Marp theme catalog, and role-based access Groups — plus per-note authorship, a task board grown up enough to run a week on, and an MCP server that can now edit slides, tasks and drawings rather than only note bodies.

Every new capability is off by default. An operator who wants none of them carries no new service, no new required env var, and ~1.8 KB more in the frontend entry bundle; the drawing engine and the deck exporter are code-split and download only when something actually needs them.

One fix here matters more than its size suggests: in the git-native HA topology a move could look like data loss. See below.

Added

• Excalidraw drawings (opt-in — ENABLE_EXCALIDRAW=true). A .excalidraw.md note opens on a full Excalidraw canvas, and any note can embed a drawing read-only with an ![alt](path.excalidraw.md) image embed. The file is Obsidian-compatible: the scene is stored as a JSON block and the drawing's text is mirrored into a ## Text Elements section so it stays searchable and readable, so drawings reuse the same history, restore, comments and search as any note. Off by default; the (large) editor bundle is code-split and only loads when a drawing is opened. Operators can preload organisation-wide shape libraries via EXCALIDRAW_LIBRARIES (Helm excalidraw.libraries), a list of .excalidrawlib URLs. Concurrent editing is last-write-wins with a conflict warning. See docs/excalidraw.md.
• Centralized Marp themes (opt-in — ENABLE_MARP_THEMES=true, on top of ENABLE_MARP). Decks can reference a shared theme by name (theme: <name> in the frontmatter) instead of embedding a large per-deck style: block, so presentation styles are managed and evolve in one place. Themes live in a reserved, hidden namespace (auto-created, git-versioned locally, and never mirrored to a per-workspace git remote), are readable by every deck in any namespace, and are edited by superadmins from a new Marp Themes admin tab. A neutral starter theme is seeded on first start when the catalog is empty.
• Export a Marp deck as a real, standalone presentation. From the deck view, export to a single self-contained .html: a genuine Marp bespoke deck (keyboard/touch navigation, fullscreen and presenter view) that opens offline in any browser. Rendered entirely in the browser with marp-core and wrapped in marp-cli's bespoke player (vendored, MIT) — no server dependency and nothing added to the backend image. Centralized themes are resolved and images inlined, so the file is fully self-contained.
• Role-based access "Groups" (multi mode). A new superadmin-managed Groups admin tab lets you define named groups whose members are mdnest users and/or IdP (OIDC) group IDs, and grant those groups read/write access to namespaces (with the same path scoping as per-user grants). A user's effective access is the union of their own grants and the grants of every group they belong to — directly, or through an OIDC group ID carried in their login token. OIDC-group membership is read from a configurable ID-token claim (OIDC_GROUPS_CLAIM, e.g. groups on Entra ID) and snapshotted at login, so a change at the IdP applies on the member's next sign-in. Note for operators: because the OIDC-group snapshot lives in the session token, removing a user from an IdP group (e.g. offboarding) does not revoke their mdnest access until that token expires — bounded to the SSO session lifetime (12 hours). Direct user membership, by contrast, takes effect immediately (added and revoked live). Each OIDC-group member can carry an optional display label (reference only — matching is always on the group ID). Fully additive and opt-in: with no groups defined and OIDC_GROUPS_CLAIM unset, behaviour is unchanged.
• Per-note authorship attribution (multi mode). A note now carries an internal activity trail — who created it, who last edited it, and everyone who has contributed — surfaced from the note's context menu as an Attribution panel. Every save is recorded in a Postgres-backed note_activity table (migration 014), and contributors are cross-checked against the note's git history so authorship stays accurate even for edits made outside the app. Multi mode only and fully nil-safe: a single-mode install has no user identities to attribute, so nothing is recorded, the /api/note/attribution route is never registered, and the UI entry stays hidden.
• Task relations, filters and a cross-workspace board. Tasks can declare depends-on, blocked-by and related-to relations (rendered on the card and resolvable to other tasks by their stable ref), carry an assignee, and be narrowed with a filter bar over title, tags and assignee (All / Me / Unassigned / a member). A new All workspaces scope aggregates tasks from every workspace you can access, each card showing where it came from — access is enforced per request, and the view serves nothing rather than everything if its namespace filter is ever left unwired. Plus board polish: a mobile layout, task delete, manual refresh, and cards whose title gets its own full-width line.
• Closing a task is blocked while its sub-tasks are unresolved. Checking a parent done, dragging it into a Done column, or saving an edit that moves it there are all refused (HTTP 422) until every sub-step is ticked, with the reason surfaced in the UI rather than failing silently. Enforced in the backend, so it holds for API and MCP callers too, not just the board.
• MCP: subject-level CRUD for tasks, Marp decks and drawings. The MCP server now manages the things inside a note, not just note bodies: full task CRUD (including relations, assignee, the close guard and cross-workspace search), per-slide Marp operations (list/read/edit/delete/move/insert, with slide splitting that ignores a --- inside a fenced code block), and Excalidraw authoring — compile a high-level nodes+edges spec into a real scene with bound labels and connected arrows, then read and edit individual elements with cascade delete and automatic re-flow. Tools are registered based on the backend's enabled features, so a notes-only deployment exposes only the note/tree/search tools.

Fixed

• A move no longer looks like data loss on HA replicas. In the git-native HA topology (stateless app replicas + a single durability writer), app replicas read exclusively from the Redis working set. The writer applied a rename to the durable tree and then evicted both ends from the working set — the source correctly, the destination wrongly — and never re-populated it. Because a replica cannot re-hydrate from the durable tree on a cache miss, and the enqueuing replica had already dropped the source, every read of the moved file (or the moved directory's whole subtree) returned 404 until the next full hydrate: POST /api/move reported success and the note then appeared to be gone. The destination is now re-hydrated from the durable tree after the rename, for both a single file and a moved subtree. Single-box installs were never affected.

• The Marp theme editor's Save button is styled to match the rest of the admin panel (a primary button with a proper disabled state) instead of the default browser control.

Security

• Go 1.26.6 and refreshed golang.org/x/*. The backend image built on go1.26.5, whose standard library carried reachable advisories in crypto/tls, net/http, encoding/asn1 and net; golang.org/x/text v0.38.0 was also reachable, from HandleNoteAt through ReverseProxy into norm.Form. The go directive and the builder image move to 1.26.6, x/net to v0.56.0 and x/text to v0.39.0. govulncheck reports zero vulnerabilities affecting mdnest's code.
• Frontend dependency advisories cleared. v4.1.3 bundled nanoid 3.3.16 (high — GHSA-28wg-ghj8-5hjv, GHSA-2v37-7h3g-55p8) plus moderate advisories in mermaid 11.15.0 and dompurify 3.4.12. All are transitive, none were reachable from a code path mdnest calls directly, but they are gone: nanoid → 5.1.16, mermaid → 11.16.1, dompurify → 3.4.13, and npm audit reports zero for both the frontend and the MCP server. The new Excalidraw dependency pulled its own vulnerable nanoid and lodash-es copies in transitively; those are pinned up via overrides rather than shipped.

v4.1.3 — comments, Marp safety, and four papercuts that made mdnest look broken

Alongside the comment-editing and Marp work, this release clears a batch of reported bugs that shared a shape: mdnest was working correctly but not saying so, so each one read as a malfunction. A DNS failure blamed the server's config file. A namespace added to mdnest.conf simply never appeared. A file created on disk sat invisible in the sidebar until you clicked Refresh. Diagram text couldn't be copied at all.

Added

• Task assignees, filtering, and a cross-workspace view (task board only, still behind ENABLE_TASK_BOARD). Tasks carry an assignee: metadata field picked from the workspace's members; a filter bar narrows the loaded tasks by title, tags and assignee (All / Me / Unassigned / a member); and a new "All workspaces" scope aggregates tasks from every workspace you can access, with each card showing which workspace it came from. The cross-workspace view enforces access per request and serves nothing if its namespace filter is ever left unwired.
• Edit your own comments. The author (and only the author) can now revise a comment's text inline from the comment panel; an "(edited)" marker is shown once a comment has been changed. Resolve/reopen stays open to everyone.
• Resizable comment panel. Drag the panel's left edge to widen or narrow it (persisted per browser), and the comment/reply/edit text areas can be resized vertically.

Changed

• The comment panel is usable in any view. Opening comments no longer forces you out of preview-only into the Live editor — you can review Marp slides and leave general comments side by side. Selection-anchored comments and highlights still require the Live editor.
• The preview now makes room for the comment panel instead of being covered by it, so slides stay fully visible while commenting.
• Comment text preserves line breaks instead of collapsing multi-line comments into a single block.

Fixed

• mdnest login now diagnoses the actual problem and prints a fix you can paste. Four bugs in one code path. An unreachable server was reported as "this server has no SERVER_ALIAS configured" — a claim about an mdnest.conf the CLI never managed to read, which sent people off to edit and rebuild a blameless server; the two cases are now distinguished and the real reason is named (DNS, refused, timeout, TLS). The recovery hint was rebuilt from the raw positional arguments, so it kept whatever bad argument you typed and dropped your token entirely. That hint's placeholder was @<name>, and <name> is a shell redirection, so pasting the suggested fix errored in both zsh and bash — every command mdnest tells you to run now uses literal placeholder words. And an argument that wasn't a URL was saved to disk anyway with only a warning, becoming the default server on a fresh machine so every later command aimed at garbage. The common root cause — an alias written without its leading @, which shifts every argument along — now gets the same command back with that one character fixed.
• The file tree refreshes itself for writes that never went through the API. A file appearing on disk stayed invisible in the sidebar until you pressed Refresh. The tree-changed event is emitted only for changes that arrive through the API, so git-sync pulling another machine's commits (or an editor on the host) produces no event at all — and the polling fallback was skipped entirely whenever the live-collab websocket was connected, treating "the socket is up" as "the tree is current". Exactly the installs running both collab and git-sync got no automatic update at all. The poll now always runs, at 30s instead of 60s, and silently: no spinner or indicator bar for a refresh you didn't ask for.
• Mermaid diagram text can be selected and copied. In the inline preview the rule making labels selectable was attached to a CSS class nothing had applied since click-anywhere-to-expand was replaced by the expand button, so it matched nothing. In the fullscreen viewer it was impossible by construction: the canvas suppresses selection so a drag pans, and every mousedown started that pan — including one on a label. A drag beginning on text now selects instead of panning, and both modes gained a Copy text button that puts the whole diagram's labels on the clipboard, the same affordance code blocks already had.
• Adding a namespace tells you it isn't live yet. Namespaces are Docker volume mounts, so writing a MOUNT_ line only changes the desired state — the running backend keeps serving the mounts it was created with. Nothing surfaced that gap, so an added namespace was just absent from the UI and the API. add-namespace now offers to reload for you (defaulting to yes), and mdnest-server status compares mdnest.conf against what the running container actually serves, reporting either direction: configured but not live, or still served but no longer in the conf.
• Comments no longer vanish when a note is deleted and recreated (STORAGE_BACKEND=git only). A note's identity (its hidden mdnest: marker, which links it to its .mdnest/comments/<id>.jsonl sidecar) is now reconciled on write: a recreated or marker-stripped note recovers the marker the path previously carried in git history, so its comments stay attached instead of being orphaned. A note's marker is also snapped back if an overwrite tries to change it. The default local backend is unaffected by this change — a delete+recreate there still starts a fresh comment thread.
• Marp decks are no longer corrupted by the Live editor. A note whose frontmatter declares marp: true is now always edited as raw text — the Live/WYSIWYG editor is disabled for it, because round-tripping the markdown through the editor's document model rewrote the frontmatter (--- → ***) and slide separators and silently broke the deck on autosave. The Basic editor (with the live slide preview alongside) is forced, and the "Live" toggle is disabled with an explanatory tooltip while a Marp note is open.
• Helm: StatefulSet upgrades no longer stall on an immutable-field diff. volumeClaimTemplates carried the full mdnest.labels set, which includes helm.sh/chart and app.kubernetes.io/version — both change on every release. Because volumeClaimTemplates is an immutable part of the StatefulSet spec, each upgrade became an illegal update to an immutable field: a server-side-apply dry-run failed and Argo CD (or helm upgrade) stalled with a perpetual out-of-sync diff. The claim templates now carry only the release-invariant mdnest.selectorLabels, so upgrades apply cleanly. Chart version bumped to 0.3.1 (template-only change).

• Per-note authorship attribution (multi mode). A note now carries an internal activity trail — who created it, who last edited it, and everyone who has contributed — surfaced from the note's context menu as an Attribution panel. Every save is recorded in a Postgres-backed note_activity table (migration 014), and contributors are cross-checked against the note's git history (blame) so authorship is accurate even for edits made outside the app. Multi mode only and fully nil-safe: single-mode installs have no user identities to attribute, so nothing is recorded, the /api/note/attribution route is not registered, and the UI entry is hidden.

v4.1.2 — mdnest list you can actually read

Patch release fixing GitHub issue #87, reported from Fedora 44. Both halves of that report were the CLI's fault, and both are fixed. Nothing outside the mdnest CLI changed.

Fixed

• A broken python3 install no longer pollutes CLI output (#87). The reporter had a stale matplotlib .pth in ~/.local, which makes every python3 start print a traceback to stderr. Because the CLI shells out to python3 for percent-encoding and JSON parsing, that traceback landed in the middle of mdnest's own output and mdnest list <workspace> looked like it had crashed. Every python3 call now goes through one guarded helper that passes -S (skip site initialisation, so the user's .pth files are never processed) and -E (ignore PYTHON* env vars), drops python's stderr, and reports failure so a python3 that is present but broken degrades to the pure-bash/awk fallback each call site already had, instead of silently returning an empty value. mdnest servers -v and the JSON field parser now fall through the same way, rather than reporting an unreachable server or an empty field when python3 misbehaves.

• mdnest list renders a tree instead of dumping raw JSON (#87). Listing a namespace printed the entire /api/tree payload as one long line of compact JSON — unreadable for a namespace of any size — and mdnest list with no namespace printed a raw JSON array. Namespaces now print one per line, and a namespace or folder prints as a tree with a folders/files count:

  engineering
  ├── Architecture/
  │   ├── decisions/
  │   │   └── 001-storage.md
  │   └── system-overview.md
  └── README.md
  
  2 folders, 3 files
  

  Pass --json (or set MDNEST_JSON=1) to get the exact previous payload, so anything scripted against it keeps working — including the client-side subfolder scoping. The renderer is deliberately awk-only, with no python3/jq tier: awk is on every machine mdnest supports, so a listing looks identical everywhere and a broken python install cannot garble it. Verified byte-for-byte on gawk, mawk and busybox awk.

• The legacy mdnest note read <ns> <path> form returns the note again. It tried a tree lookup first, so for any file that actually existed it printed that file's {"name","type","path"} tree entry instead of its content — the note body only came back for a missing path. The flat mdnest read form was never affected.

Chores

• Lock-file-only dependency bumps to clear advisories published since v4.1.1: undici 7.28.0 → 7.29.0 (frontend, dev-only via jsdom) and @modelcontextprotocol/sdk 1.27.1 → 1.30.0 (pulling hono 4.13.0, ip-address 10.4.0, fast-uri 3.1.5). Both audits report zero vulnerabilities.

Tests

• tests/cli-unit.sh gained two passes that reproduce the issue #87 environment through a PATH shim — a python3 that is noisy-but-working and one that exits non-zero — asserting correct values and a clean stderr in both, plus unit coverage of the tree/namespace rendering (run twice, with and without a parser present, since the two must agree).
• tests/cli-smoke-test.sh now asserts that a listing contains tree connectors and no JSON keys, that --json still returns the scoped payload, and that legacy note read returns a note body. All of it passes in the bare no-python3/no-jq alpine container of tests/e2e-docker.sh.

v4.1.1 — The conflict banner learns whose save it is

Patch release fixing GitHub issue #82.

Fixed

• Editing your own note in multi mode no longer flashes "This file was modified by you. Your changes may conflict." (#82). The v4.0.0 fix remembered the etags this tab's saves produced and suppressed matching file-changed echoes — but it learned each etag from the PUT response, while the backend broadcasts the echo before it writes that response. The echo therefore usually arrived first, found nothing to match, and popped the conflict banner naming the editing user on nearly every autosave — reflowing the document up and down while typing. The suppression now lives in a pure echo-gate module with an in-flight-save window: broadcasts that arrive while this tab's own PUT is outstanding are deferred and re-checked once the save settles, by which point the response has registered its etag and the echo is recognized and dropped. A genuine change by another user — or by the same user through the CLI/MCP — is still delivered after the save settles, and deferred messages are discarded on file switch so they can't replay against the wrong note. Regression-pinned in frontend/src/__tests__/echo-gate.test.js, covering the exact race from the issue, the late-echo ordering the old ring handled, and the remote-change-during-save path.

v4.1.0 — Marp slides, Git Workspaces admin, and the image v4.0.0 forgot

A short cycle on top of v4.0.0, and it starts with an apology: v4.0.0 shipped without its mdnest-mcp-server image, so any Kubernetes install with mcp.enabled=true hit ImagePullBackOff on a tag that was never built. This release publishes it and makes that class of mistake impossible to repeat.

Fixed

• v4.0.0 published no mdnest-mcp-server image (#76). The release workflow's build matrix and the chart had drifted apart: the matrix entry was removed in a hotfix while the MCP server had no Dockerfile, re-added once it did — and then silently lost again when the release branch merged into main (the branch's file matched the merge base exactly, so git took the other side's removal and nothing warned). All three images are published for this release. The workflow now reconciles every image the chart references against the registry and refuses to publish if one is missing, so a release can no longer ship a chart pointing at nothing.
• Version History was broken on every namespace of a git-native HA install. App replicas hold no git tree — only the writer does — so the history endpoints reported "git-sync is not configured" for everything. They now proxy to the writer, exactly as attachments already did. Single-box and writer roles are unchanged.
• Deleting a Git Workspace no longer destroys notes that exist nowhere else. Decommissioning purges a namespace from storage only when a mirror demonstrably holds the current copy (mirroring on, last sync succeeded). Without one, access is revoked and the config removed but the bytes stay put — which matters most for a namespace that came from a MOUNT_, where the purge would have emptied a bind-mounted host directory.
• Mirror status stops claiming "ok" for a workspace that has never synced. Four honest states — off / pending / ok / error — driven by whether a sync actually happened, with push-to-create on a fresh remote classified as a benign pending rather than a red error.
• Personal workspaces no longer appear in the Access Grants and Namespace Admins pickers; they are managed by their owner, not administered by others.
• Deleting a workspace or group now revokes the namespace's grants and namespace-admins instead of leaving them orphaned.

Added

• Marp slide rendering (opt-in — ENABLE_MARP=true). A note whose frontmatter declares marp: true renders as a slide deck in the Preview pane, so the note you drafted an idea in is the deck you present from — no second copy in Slides or PowerPoint to keep in sync. The source stays plain, diff-able, git-mirrored Markdown. Slides render per-slide in a fully sandboxed <iframe sandbox="">, which is what makes it safe to honour the inline HTML real decks use. Off by default; the engine is a lazy chunk, so an install that doesn't want slides carries 1.6 KB, not the engine.
• Git Workspaces admin management — workspace groups, discovery of operator-provisioned sub-projects, and per-group project add/enable/remove, all in the Admin panel.

Notes for operators

• If you are on v4.0.0 with mcp.enabled=true, upgrade to v4.1.0 — the image your chart references now exists. Nothing else about v4.0.0 needs action.
• ENABLE_MARP is plumbed through mdnest.conf / setup.sh as well as the chart, so it works on the single-box path too.

Thanks again to @ecthelion77, who found the missing image, diagnosed it, and sent the fix.

v4.0.0 — Task board, git-native HA, and bring-your-own-repo durability

The first release built substantially with outside contributions, and the largest change to what mdnest can be deployed as since it started. Notes are still plain Markdown files in a git repo you own — that hasn't moved, and most of this release exists to keep it true at scales where it previously wasn't.

A single-box install is unchanged. The generated docker-compose.yml is byte-identical to v3.11.7's, and .env gains exactly one line (ENABLE_TASK_BOARD=false). Everything below is either opt-in or invisible unless you run multi-user mode.

Major version because two changes require operator action on upgrade — see Breaking changes. Huge thanks to @ecthelion77 (Olivier Gintrand), who wrote most of what follows.

Breaking changes

• Superadmins no longer have implicit read access to note content. A superadmin administers every namespace — users, grants, namespace lifecycle — but that authority no longer doubles as ambient access to the notes. Data access now flows through grants for every role. On an existing multi-user install, namespaces a superadmin never held a grant in will disappear from their sidebar, tree and file APIs until they self-grant; the admin surfaces are unchanged and still list every namespace to administer. Single-user mode is unaffected. This is arguably the change that makes multi-user mode trustworthy rather than merely functional: administering a workspace and reading people's notes are different powers.
• The Helm chart runs the backend as a StatefulSet with per-pod volumeClaimTemplates instead of a Deployment plus standalone PVCs (chart 0.1.0 → 0.2.x). Upgrading an existing chart install without action would delete the Deployment-era notes PVC, so the chart now refuses that upgrade and tells you to adopt the volume with persistence.notes.existingClaim: <release>-notes. Fresh installs need nothing.

Added

• Task board (opt-in — ENABLE_TASK_BOARD=true). A per-namespace kanban built on the - [ ] checkboxes already in your notes. No new datastore: a task is a line in a note, the note stays the source of truth, and every board action is a plain line-level edit — so anything you type in a note shows up on the board and vice versa. Tasks can carry an indented detail block (status, due, priority, tags, steps, notes) that remains readable Markdown you wouldn't mind seeing in a diff. Column layout lives in the namespace's .mdnest/board.json sidecar (the same convention as .mdnest/comments, and likewise hidden from the tree and from search). Adds four MCP tools — list_tasks, create_task, edit_task, move_task. Off by default: the routes aren't registered and the UI chunk isn't loaded until you turn it on. See docs/tasks.md.
• Git-native HA — multiple replicas without ReadWriteMany storage. A new git storage backend keeps the same on-disk layout but owns git history in-process (one repo per namespace), replacing the git-sync sidecar and debouncing commits on writer idle so an editing session becomes one commit rather than one per tick. Add REDIS_URL and it splits into N stateless app replicas plus a single writer that owns the git tree, with Redis as the coherence tier. Git stays the durable source of truth — git log, grep and cp -r all still work. Read docs/kubernetes.md before choosing it: on an app replica a save is acknowledged once it's on the queue, not once the writer has committed it, so an acknowledged write can be lost and Redis becomes a durability component that needs AOF. A single box, or the git backend without Redis, has an RPO of zero.
• Per-workspace git remotes (multi mode). Any namespace can mirror to its own repository rather than one operator-wide remote, and a user can point their personal workspace at a repo they control — so "your notes are a git repo you own" survives contact with a multi-team install. Credentials (HTTPS PAT or SSH key) are sealed with AES-256-GCM, never returned by the API, and never placed in argv or a URL; mdnest refuses to store one at all unless a non-default MDNEST_ENCRYPTION_KEY is set. Optional GIT_REMOTE_ALLOWED_HOSTS bounds where they can be used.
• MCP server over streamable HTTP, with optional per-user OAuth 2.1. The bundled MCP server can now be a shared network endpoint instead of a local stdio process, and in OAuth mode each client signs in through your existing SSO so actions are attributed to the real user rather than one shared token. stdio remains the default and a first-class path — with MCP_TRANSPORT unset, no listener is bound and no OAuth code is even imported.
• Opt-in Redis backplane for live collaboration (REDIS_URL), so presence and edits stay in sync across replicas. With it unset the hub is exactly the in-process one it always was — no goroutine, no allocation per event.
• Opt-in SSO user auto-provisioning (SSO_AUTOPROVISION_USERS=true): an unknown but IdP-authenticated email is created as a least-privilege collaborator with no grants, instead of being rejected. Off by default.
• Task-board support in the Helm chart — taskBoard.enabled.

Changed

• API tokens live in PostgreSQL in multi mode (single mode keeps tokens.json, and gains no database). Existing tokens are imported from tokens.json on first start, so an upgrade doesn't invalidate them. This is what finally removed the last ReadWriteMany requirement for running multiple replicas.
• A namespace-scoped storage.Storage interface now sits between the handlers and the filesystem, which is what made the git backend possible without rewriting request logic — and made the handlers testable.
• Note history and namespace listing move through the storage layer, and /api/files/ keeps range requests and conditional GETs (so image caching and media seeking still work).

Fixed

• Ticking a task off in the Live editor now sticks. The box changed, the file didn't, and a refresh brought the tick back — so a card moved to Done couldn't be un-done from the note. The editor's save gate stayed armed until a keypress or a click landed inside the document, and ticking a checkbox produces neither (ProseMirror handles it internally), so every checkbox edit in a freshly opened note was dropped for the whole session. The gate is now scoped to the document injection it exists for. Pinned by a new tests/browser/kanban.spec.js — six specs covering the board, task parsing and both checkbox directions, asserted against the bytes on disk.
• Editing your own note no longer raises a conflict banner or jumps the cursor. The backend fans file-changed to every connection on a note including the tab that just saved; the handler now recognises its own echo. Broadcast latency made this near-constant on a multi-replica deployment. A same-user write from the CLI or MCP still propagates.
• A namespace created for a personal workspace is materialised only once it actually mirrors somewhere, so it can't exist as an undurable directory.
• The task board and the live editor are both lazy-loaded, so neither is on the critical path for someone who doesn't open them.

Security

• Command execution via a crafted git remote URL (unreleased code). remote_url and branch were passed to git as positional arguments, so a value starting with - was parsed as an option — --upload-pack=<cmd> executes <cmd>. Any authenticated user could reach it through their own personal workspace, giving command execution in the writer, the process holding the sealing key and every workspace's credentials. Now blocked at the API boundary and with --end-of-options at the exec site; each layer was verified to stop it alone.
• OAuth authorization codes could be delivered to an attacker-chosen host (unreleased code). /oauth/authorize accepted any HTTPS redirect_uri without checking it against a registered client, and PKCE doesn't help when the malicious client starts the flow. Delivery is now restricted to loopback plus origins explicitly listed in MCP_ALLOWED_REDIRECT_ORIGINS.
• Superadmin implicit read access removed — see Breaking changes.
• API-token list and revoke ownership scoping is now pinned by tests; nothing covered it before.

v3.11.7 — XSS hardening, cross-namespace file leak fixed, Helm chart, build CI

First release with outside contributions. Thanks to @ecthelion77 (Olivier Gintrand) for the sanitization hardening, the /api/files/ authorization fix, the Helm chart, and the CI workflows.

Security

• Rendered markdown, release notes, and mermaid SVG are now sanitized before they reach the DOM. Note bodies are user-authored and shared between users in multi-user mode, and marked passes raw HTML through by design — so a note containing <img src=x onerror=…> or a javascript: link href executed when anyone previewed it. All three injection points now run through a single frontend/src/sanitize.js module (DOMPurify): event-handler attributes and dangerous URI schemes are stripped, and <a target="_blank"> gets rel="noopener noreferrer" so external links can't reach back into the opener. The Preview's own post-passes are unaffected — class, data-*, and task-list checkboxes all survive sanitization.
• A signed-in user could read files from namespaces they had no grant for. GET /api/files/<ns>/<path> — the endpoint that serves uploaded images and attachments — carried its namespace in the URL path rather than the ?ns= query param, so it couldn't use the query-param permission middleware and was registered with authentication only. Any authenticated principal, including an API token, could fetch any file in any namespace by guessing the URL. It now enforces the same per-namespace read check as every other content endpoint. Single-user mode is unaffected.
• google.golang.org/grpc bumped to v1.82.1 for GO-2026-6061 — vulnerabilities in the xDS RBAC authorization engine and the HTTP/2 transport server. It arrives transitively through the Firebase/Google Cloud SDK, and govulncheck confirmed reachable symbols in the built binary, so it's a real exposure rather than an unused-code advisory. Caught by CI on the release PR — the local pre-push hook skips govulncheck when Go isn't installed on the host, which is precisely why the CI check is the authoritative gate.
• Four npm advisories cleared, two of them high severity. postcss ≤8.5.17 (GHSA-r28c-9q8g-f849, arbitrary .map disclosure via source-map auto-loading) and fast-uri (GHSA-4c8g-83qw-93j6, host confusion via failed IDN canonicalization) were both failing the Security Audit on main — and because that audit is a required check with no bypass, the red gate blocked every merge. Also picks up dompurify (GHSA-c2j3-45gr-mqc4) and protobufjs. All transitive, so lockfile-only — no package.json change and no new packages.

Kubernetes

• An opt-in Helm chart for clusters, alongside the usual Docker Compose install. deploy/helm/mdnest deploys the backend, the nginx frontend, and an optional git-sync sidecar as standard Kubernetes resources — no CRDs, no operator, PostgreSQL never bundled. Ingress, TLS, resource limits, probes, PVCs, and a ServiceAccount are all configurable; helm lint, both renders, and kubeconform -strict run in CI. Nothing about the Compose path changed: setup.sh, mdnest.conf, docker-compose.yml, and the Dockerfiles are untouched, and the chart is inert unless you use it. Supported today is single-replica single or multi mode with live collaboration, git-sync, ingress, and TLS. Three options are documented but rejected at install time because their code isn't in this release — storage.backend=s3, collab.redis.*, and mcp.enabled — so the chart fails loudly instead of coming up Ready while writing notes to the wrong place or splitting collaboration state across pods.

CI

• Build and test now run server-side on every push and PR, not just in a local pre-push hook. .github/workflows/ci.yml runs the backend build, go vet, and go test -race; the frontend build and unit tests; the Helm chart lint/render/validate; and a build of the backend and frontend images. The pre-push hook still exists as fast local feedback, but it can be skipped with --no-verify and silently omits govulncheck on a host without Go — so the authoritative gate is CI.
• A release workflow publishes container images and the Helm chart on version tags. v* pushes build and push mdnest-backend, mdnest-frontend, and mdnest-mcp-server to ghcr.io/<owner>/, then package and push the chart as an OCI artifact. Everything is parameterized by repository owner, so a fork publishes under its own namespace with no edits.
• The Security Audit now also runs on PRs into develop. It was scoped to main, so a contribution integrated on develop wasn't scanned until the release PR — which is when the two advisories above were found, well after the code had landed. main's required checks are unchanged; this only moves the signal earlier.

Bug fixes

• Mermaid diagrams rendered as blank boxes once SVG sanitization was added. DOMPurify's SVG profile doesn't allow <foreignObject>, and mermaid renders every flowchart node label inside one — so sanitizing deleted the text of every label while the boxes and arrows still drew. Nothing threw and no console error appeared; the diagram simply looked empty. foreignObject is now allowed, which doesn't weaken the sanitizer: scripts, iframes, objects, embeds, forms, and every on* handler inside it are still stripped, and each of those is pinned by a test.
• The Helm chart shipped pointing at the contributor's fork. Its default image repositories, home/sources, maintainer, and README install command all referenced a third party's registry and repo, so helm install from a checkout of this repo pulled someone else's images. Its appVersion was also frozen at the previous release, so a source install requested stale image tags.

Testing

• Regression coverage at the layer that would have caught each bug. backend/handlers/upload_test.go is the repo's first Go test — it asserts a collaborator granted in one namespace gets 403 on another, that a superadmin reads both, and that single-user mode is unaffected; with the authorization check removed it reports the actual cross-namespace leak. frontend/src/__tests__/sanitize.test.js pins both directions of the sanitizer: foreignObject and its label text survive, six smuggled payloads and inline handlers don't. And the browser suite now seeds a note containing a mermaid flowchart and asserts the label text is visible in Preview — asserting only that shapes rendered would have passed while every label was missing. All three were confirmed to fail with their respective fix reverted.

v3.11.6 — Clearer pitch, reveal-in-tree, instant cross-tab sync, security gate

Docs

• The README and landing page now say what mdnest is in ten seconds. The old README opened with deployment details and a flat wall of features, so a first-time reader couldn't tell what mdnest is or who it's for. It's rewritten in product-style copy: a one-line tagline ("Your notes, on your own server. Open to every device — and your AI."), a screenshot up top, a short What you get list, an Is it for you? self-check, and an honest one-line Obsidian comparison — with the full feature list preserved in a More features section below the fold. The mdnest.dev landing page (hero, meta/OG/Twitter/JSON-LD, and the AI-native/team/git-sync cards) is synced to the same framing.

Features

• Reveal-in-tree. A button jumps the left tree to the currently-open note and highlights it — handy after following a wikilink or searching, when the file is open but not visible in the tree.
• Instant cross-tab tree + sync updates. Creating, renaming, moving, or deleting a note in one browser tab now updates the tree in every other open tab immediately, and git-sync-driven changes land without a manual Refresh — no more stale tree until you reload.

Bug fixes

• The build-details popover closes on an outside click. It previously stayed open until you clicked the ⓘ again; it now dismisses when you click anywhere else, like every other popover.
• The Live editor block handle is no longer clipped behind the tree sidebar. The drag/+ handle in the left margin could render underneath the sidebar on narrow layouts; it now stays fully visible.

Security / CI

• Security Audit is now a required status check before merge to main. The audit (frontend/MCP npm audit, backend govulncheck, shellcheck) already ran on every PR, but wasn't required — so a PR could merge with a failing check (that's how a stdlib vuln once slipped onto main). It's now enforced server-side on the main-branch ruleset with no bypass. The pre-push hook mirrors the govulncheck gate locally, and the ruleset is captured as importable JSON so the gate is managed as code.

v3.11.5 — Obsidian wikilinks

Features

• Obsidian-style [[wikilinks]] are now first-class links. Vaults imported from Obsidian lean on [[wikilinks]], which mdnest previously rendered as plain text. Now [[target]], [[target|alias]], [[target#heading]] and [[#heading]] render in the preview as internal links that open the note in-app with no page reload. The href still carries the #ns/path form, so middle-click and open-in-new-tab keep working, and a same-note [[#heading]] link just scrolls the preview to that heading. Targets resolve Obsidian-style against the namespace tree — an exact path first (with or without the .md suffix), then a case-insensitive basename match with a shortest-path tiebreak; an unresolved target renders as a muted, non-clickable broken-link span so you can see the note is missing. Relative markdown links to .md files ([text](../notes/other.md)) now navigate in-app too, instead of opening a dead file URL in a new tab.
• Wikilinks are visible and clickable in the Live editor. [[...]] spans get a link-coloured highlight, and Ctrl/Cmd+Click opens the target (plain click still places the caret, so editing is unaffected). This is decoration-only — no schema change — so the stored markdown stays literal [[...]]. Round-trip fidelity is guaranteed: Milkdown's serializer escapes [[ to \[\[ in plain text, so every save is routed through a restore pass that keeps documents byte-identical. Code spans and fenced code blocks are excluded, so [['field' => 'x']] in a PHP snippet stays code, not a link.

Thanks to @lglot (Luigi Lotito) for contributing this feature.

v3.11.4 — git-sync self-healing, fresh-machine CLI, tree memory + local test gate

Bug fixes

• git-sync now converges instead of looping forever on a diverged, dirty tree. When a notes repo was simultaneously ahead and behind its remote and had uncommitted live-editor edits, the old sidecar kept failing the same way every cycle: a git pull --rebase couldn't start against the dirty tree, so it never integrated the remote, the local commit never pushed, and the divergence only grew (one namespace fell 118 commits behind unnoticed). The pull is now merge-only and self-healing — it autostashes the working tree (tracked + untracked) so the merge always starts, distinguishes a real content conflict (keep remote, save local copies as .sync-conflict-*) from a merge that simply couldn't begin (abort cleanly, retry next cycle), and re-applies the stashed edits on top of the merged result (saving a recoverable .mdnest-sync-autostash-*.patch if they collide). Push is now gated on HEAD actually containing the remote, so a non-fast-forward can never spin. Each cycle writes a git-excluded .mdnest-sync-status.json (state/ahead/behind/message), and the bookkeeping files are added to .git/info/exclude so they're never committed.
• A broken background sync is now visible instead of silent. /api/admin/sync-status overlays the daemon's self-reported health (daemonState/daemonMessage/ahead/behind), and the sidebar polls it every 60s. When the daemon reports error, the status bar shows a red ✕, a "Git sync broken — N behind" label with the reason in its tooltip, and a Retry button for admins (runs commit + pull + push). Previously a wedged sync only showed a stale "Synced X ago" date, hiding the failure entirely.
• A fresh multi-mode install is usable out of the box again. The first-run bootstrap seeded its one account with the literal role admin, which since the v3.5.0 three-tier role split means "namespace-scoped admin with no namespaces assigned" — so the only account saw zero namespaces and had no way to grant itself access (the grant dropdowns are themselves role-filtered). Migration 007 only promotes pre-existing admin rows, not ones the seed creates after it runs. The bootstrap account is now seeded as superadmin (global); the existing count == 0 guard keeps this to the very first user, so later invitees are unaffected.
• The mdnest CLI works on a fresh machine without python3. The CLI hard-depended on python3 for URL-encoding note paths, parsing the server version, and scoping list <subfolder> — with no fallback. On a machine without python3, note commands broke and mdnest servers labelled a perfectly reachable server "unreachable" because it conflated a failed fetch with a failed parse. Now urlencode/urldecode have pure-bash fallbacks, JSON fields are read via python3 → jq → a brace-depth-aware awk (so the top-level version is returned, not latestRelease.version), and list <subfolder> scoping falls back to jq then an awk subtree extractor. mdnest servers now separates connectivity from parsing (using curl's exit code) and reports the real reason on failure (DNS/refused/timeout/TLS), with a --connect-timeout so it never hangs. A one-time note suggests installing python3 for the best experience.
• The CLI installer no longer aborts mid-download on a fresh machine. install-cli.sh wrote curl's output straight to /usr/local/bin/mdnest, which fails with curl: (56) Failure writing output to destination when that directory doesn't exist yet or isn't writable. It now downloads to a temp file, sanity-checks it, creates the target directory, and installs atomically — with a ~/.local/bin fallback (plus a PATH hint) when /usr/local/bin can't be used. mdnest update uses the same safe temp-download + atomic-install path, and resolves its own location without readlink -f (unsupported on macOS) or python3.
• Install or update from any branch. install-cli.sh and mdnest update honour MDNEST_BRANCH (default main), so curl -fsSL .../develop/install-cli.sh | MDNEST_BRANCH=develop bash installs the develop build — the installer still handles sudo/mkdir/atomic-install for you (no manual sudo needed).
• Left tree remembers which folders are open, per namespace. A refresh used to re-expand all top-level folders and forget whatever you'd collapsed, and the expand/collapse icons flickered. Expansion is now a per-namespace set persisted in localStorage (restored on load and namespace switch); the open file's ancestors still auto-reveal and search still force-expands matches.
• Folders containing the open file can be collapsed again. A regression made any folder on the path to the currently-open note impossible to collapse — it sprang back open immediately. The tree used to force such folders expanded (containsActive); now the open file's ancestors are added to the persisted expansion set once (so they auto-reveal) but stay freely collapsible.
• Copy Path is now unambiguous, and mdnest:// URIs work in the CLI. Copy Path produced mdnest://@alias/ns/<path> with raw spaces, so a path like 19 Jun 2026.md looked like three tokens to an LLM/shell. Path segments are now percent-encoded (19%20Jun%202026.md), and the CLI's parse_path strips a leading mdnest:// and percent-decodes the namespace + path, so the copied URI is usable verbatim (raw CLI paths containing a literal % are left untouched).
• Every code snippet in Settings has a one-click copy button. The CLI / MCP / API tabs showed install commands, the login line, usage examples, the MCP config JSON, and curl examples as plain blocks you had to hand-select. Each now has a copy icon (with a "Copied!" confirmation) — works on both HTTPS and plain-http LAN installs.

Testing

• Local, end-to-end pre-merge test gate — no CI/remote required. A tiered harness runs before code reaches main: tests/cli-unit.sh (instant pure-function checks, run both with python3 and with it force-disabled — the cheap guard for the fresh-machine regression class); tests/e2e-docker.sh (builds the backend from the working tree, boots a throwaway single-mode instance, and drives the real CLI against it on the host and inside a bare no-python3 container); and tests/e2e-browser.sh (boots the full frontend+backend stack and runs a Playwright browser suite covering login, tree, opening/rendering a note, the Live and Basic editors, search, and note creation). The pre-push hook runs the unit tests on every push and the full Docker + browser suites when pushing toward main. The Docker harness immediately caught the list <subfolder> no-python3 gap fixed above.

v3.11.2 — CLI list/move fixes + prettier update indicator

Bug fixes

• mdnest list <ns/subfolder> now scopes to that subfolder. It used to ignore the deeper path and return the entire namespace tree. Now it returns just that folder (its children) or the file entry, and errors with a non-zero exit on a missing path.
• mdnest move no longer loses content when given a full destination path. A full @alias/namespace/path destination (the style typed for the source) made the server treat @alias/namespace/ as literal folder names and relocate the file to a bogus path — the intended destination then read empty and write/delete returned 404. The CLI now normalizes the destination to a namespace-relative path and rejects cross-namespace moves.
• The "new version available" indicator no longer renders as an oversized cream blob. On a narrow sidebar the old badge wrapped its ↑ and version onto two lines inside a pill, which looked broken. The alert is now folded into the build-details ⓘ: when an update is available the icon turns accent-blue and gently pulses (respecting prefers-reduced-motion), and clicking it opens the popover with the build details plus a tidy "↑ vX.Y.Z available — see what's new" action. Removed the standalone badge.

v3.11.1 — Live editor mermaid sizing + contrast

Bug fixes

• Mermaid text is now always readable, whatever fill the source specifies. Author/AI-written diagrams that set a light node fill (style X fill:#fff8e1, a light classDef, …) rendered as light-text-on-light-fill — invisible. The Live editor was injecting a blanket .nodeLabel { color:#cdd6f4 !important } override that forced every label light, fighting the per-node brightness logic that's supposed to pick contrast. Removed the blanket override so fixMermaidTextColors() is the single authority: each label's color is computed from its own node's fill brightness (dark text on light fills, light text on dark), so contrast holds regardless of the colors the author chose. (Print/export keeps its own light-page palette.)

• Mermaid diagrams now render at a sensible size in the Live editor. Small diagrams ballooned to full width while large/tall ones shrank into a corner — because every rendered SVG had its real dimensions stripped and was forced to width:100%, then classified by width alone (<400px = "small"), so a wide flowchart LR filled the pane (stretched up) and a narrow flowchart TB rendered at its tiny natural width with empty space beside it. Now the SVG keeps its natural width, capped at the container and at a 820px max (width:<natural>px; max-width:min(100%, 820px); height:auto): small diagrams stay small (no stretching), large ones scale down to fit (no shrinking into a corner) and don't sprawl past 820px on wide screens, and the zoom/Fit controls layer on top. The preview box also lost its oversized 200px min-height and 2rem 3rem padding, so a tiny inline diagram no longer sits in a giant empty frame.

v3.11.0 — CLI stdin fixes + smoke-test harness

Added

• Build commit shown next to the version. /api/config now reports a commit field — the short git SHA the backend binary was built from, injected at build time via -ldflags (computed by setup.sh, passed through docker-compose as a build arg, baked into the binary). The sidebar footer renders it as v3.11.0 · <sha>, and mdnest servers shows it as 3.11.0 (<sha>). Because the SHA is compiled into the binary rather than read from config, it can't drift from the running code — so a stale container is now obvious even when the version string hasn't changed (the exact situation where a rebuilt develop still displayed an old version). Falls back to dev for local go build without the ldflag.
• Build-details popover (ⓘ) with build time. A short SHA alone isn't very legible, so the sidebar version now has an info button that opens a small popover showing the version, the commit (linked to its GitHub commit page), and when the running build was produced — a buildTime field newly added to /api/config, baked in via -ldflags alongside the commit (UTC timestamp computed by setup.sh, rendered in local time). Answers "is this the build I just deployed, and when?" at a glance.
• Live editor: reclaimed the left space. Crepe reserved an ~88px left gutter for its hover drag/+ block handles, wasting ~20% of the width on mobile. The handle is now a compact vertical grip sitting flush in a 28px gutter (no empty lane beside it), list indentation is tightened (marker column 24px→20px, marker→text gap 10px→4px) so bullets hug the left, and a Live-toolbar toggle hides the handle entirely for full-width content (16px margin). The slash / menu keeps working with the handle hidden. The state persists per browser and defaults to hidden on touch devices (hover: none / pointer: coarse), shown on pointer devices.
• Live editor: toolbar buttons work reliably on touch + a real Link prompt. Toolbar formatting buttons used mousedown, which doesn't preserve the editor selection on touch devices, so most icons did nothing on mobile; they now use pointerdown (covering touch) and refocus the editor after running, so bold/heading/list/etc. apply to the selection. The Link button now prompts for a URL and applies it as the link href (previously it toggled a link mark with no destination — a dead link).

Bug fixes

• mdnest create now accepts piped stdin via -, like its sibling verbs. Previously create forwarded its positional argument straight to the API, so echo "# Note" | mdnest create @ns/file.md - wrote the literal one-byte string - as the file body — and the API still returned {"status":"created"}. The result was a silently corrupted (near-empty) file reported as a success, which is the most common way automated tooling (scripts, AI agents) corrupted notes: the command "succeeded", nothing retried, and the bad file surfaced much later. create now reads stdin on - (or an omitted arg) exactly like write/append/prepend.
• Robust, TTY-aware stdin handling with a literal-dash guard. All content verbs now route through a shared read_content() helper: - reads stdin and errors with a non-zero exit if nothing was piped (instead of writing a literal -); an omitted arg auto-reads stdin only when piped and never blocks on an interactive terminal. This removes both the TTY-hang footgun and the literal-- corruption path.
• Empty content now fails loudly instead of reporting false success. create/write/append/prepend refuse to issue the API call when no content was supplied, printing a clear message and exiting non-zero rather than creating an empty file and returning ok. (The guard returns success explicitly on the happy path so it is safe under the script's set -e.)

Testing

• New CLI smoke-test harness — tests/cli-smoke-test.sh. 18 end-to-end checks covering every note operation (create/write/append/prepend/read/move/delete/search/list) plus the stdin edge cases above, run against a disposable namespace. It tests the working-tree CLI, creates everything under a unique self-cleaning folder, and exits non-zero on any failure. Run it after any change to the mdnest CLI. A new optional MOUNT_testing_workspace mount (documented in mdnest.conf.sample) gives it a dedicated namespace.

Security / CI

• MCP server: bump hono to clear a high-severity advisory (transitive via @modelcontextprotocol/sdk). npm audit now reports zero vulnerabilities.
• Frontend: bump vitest 2.x → 4.x to clear the vulnerable vite/vite-node/@vitest/mocker/esbuild dev-toolchain chain (a high + a critical). The production build's direct vite was already on a fixed version; only the test runner's pinned vite@5 was affected. Tests still pass (11/11) and vite build is unchanged.
• Backend: run govulncheck in binary mode. govulncheck@latest (v1.4.0) segfaults in source mode (nil pointer deref in vulncheck.vulnFuncs) when analysing code built with the Go 1.26.x toolchain that go.mod pins. Building the binary and scanning it with -mode=binary avoids the crashing source-SSA path while still detecting reachable vulnerable symbols; the local pre-push hook does the same. Also points setup-go's module cache at backend/go.sum to clear a warning.
• Backend: bump Go 1.26.3 → 1.26.4 (go.mod + Dockerfile builder image) to clear two called standard-library advisories surfaced by the now-working govulncheck scan: GO-2026-5039 (net/textproto error escaping) and GO-2026-5037 (crypto/x509 candidate-hostname parsing), both fixed in go1.26.4.

v3.10.2 — Live editor list alignment + "Refresh Now" feedback

Security

• Bump golang.org/x/net v0.53.0 → v0.55.0 to clear GO-2026-5026 (IDNA idna.ToASCII fails to reject ASCII-only Punycode-encoded labels — reachable from the in-app update poller's outbound https://api.github.com/... request via http.Client.Do). Transitive bumps: x/crypto 0.50→0.51, x/sys 0.43→0.45, x/text 0.36→0.37. govulncheck ./... now reports zero vulnerabilities.

Bug fixes

• Live editor: nested bullets / task items no longer appear as floating orphans between sub-rows. Symptom (most visible at depth ≥ 2): a parent item's bullet drifted down to sit halfway through its own nested children, looking like a stray dot sandwiched between two unrelated sub-bullets. Root cause: the v3.10.0 Crepe migration added speculative CSS overrides on the list-item layout — most damagingly align-items: center on .list-item. Crepe's DOM for a list row is [label-wrapper | children], and .children contains the parent's paragraph plus every nested .milkdown-list-item-block. Centering the bullet vertically against that whole stack pushed the parent's bullet to the visual midpoint of all its descendants. The fix strips every speculative list-item override and trusts Crepe's playground defaults for sizing / gap / alignment — the only override left is a single .milkdown-list-item-block p { margin: 0 } rule to neutralise the app-wide p { margin: 0.4rem 0 } that leaked into list rows (Crepe inserts a <div class="content-dom"> between .children and <p>, so the previous .children > p selector was silently missing the paragraph and the leaked margin offset the bullet from the text's optical centre).
• "New version available" banner's Refresh Now button now shows Refreshing… immediately on click. Symptom: a user clicked the button after upgrading their server from v3.9.1 to v3.10.1, "nothing happened" visibly, then the tab appeared frozen and they had to kill the browser. Diagnosed: the click was actually triggering window.location.reload(), but the v3.10+ bundle is heavier than v3.9 (Crepe + Vue + CodeMirror + KaTeX ≈ 340 KB gzipped) and parsing it during the reload can take several seconds on a slow connection or low-memory device — during which the OLD tab stays visually idle because the click handler didn't update any UI before calling reload. Two changes: (a) set a refreshing state immediately on click so the button reads Refreshing… and goes disabled before the reload starts, giving the user an instant signal that the click registered; (b) drop the deprecated true argument from window.location.reload(true) — it was Firefox's non-standard "forceGet" flag that no other browser ever implemented and Firefox itself dropped years ago, so it was already a no-op everywhere but spec-incorrect.

v3.10.1 — Login form: proper password-manager hints + per-server scoping + "Keep me signed in"

New features

• "Keep me signed in" checkbox on the login form. Default ON. When checked the backend issues a 1-year JWT instead of the default 30 days, so users on personal/trusted devices stop getting unexpectedly logged out. Unchecked = 30-day TTL (the previous default) — appropriate for shared / kiosk sessions. The flag threads through every multi-step path (initial login → TOTP verify → forced password change) so the final JWT gets the right TTL regardless of which path the user takes. SSO and Firebase login default to the long-lived TTL since they have no checkbox UI of their own (their IdPs own the "stay signed in" UX). Backend helper jwtTTL(rememberMe) in backend/handlers/auth.go is the single source of truth.

Bug fixes

• Browser password-manager hints on the login form. The username and password inputs were missing name + autocomplete attributes, so browsers couldn't reliably offer to save credentials and couldn't autofill them on return visits. Added autoComplete="username" / autoComplete="current-password" (and "new-password" on the forced-password-change step) so the "Save password?" prompt appears at the right time and re-visits get autofilled.
• Per-server credential scoping (best-effort). When the user runs multiple mdnest installs, browsers were lumping their saved credentials together — typing into one install's login form would try to autofill another install's password. The form's name and id now include the server's SERVER_ALIAS (read from /api/config) plus a hidden server input field. Together those make the form's identity distinct per install for password managers that fingerprint form structure (1Password, Bitwarden, KeePassXC). Browsers' built-in password managers (Chrome, Edge, Safari, Firefox) scope primarily by HTTP origin and use the Public Suffix List for autofill suggestions — they conflate sibling subdomains (brain.example.com and wbrain.example.com both autofill from the example.com record). For full per-install isolation, give each install a different parent domain, or in 1Password/Bitwarden set the saved entry's URL match to "Host" or "Exact".
• Post-login password forms don't trigger save-prompts at the wrong time. Settings → Change Credentials, the Admin → Reset Password modal, and the Admin → Create User form all had bare <input type="password"> with no autocomplete attribute. Browsers would see those after login and offer to save / autofill the wrong credentials (which is what the user was seeing as "weird dialogue prompts inside when logged in"). Added the correct autocomplete="current-password" on current-password fields, "new-password" on every new/confirm field, "username" where applicable, and autoComplete="off" on the wrapping forms — so password managers don't mistake these for login forms after sign-in.
• LICENSE copyright corrected — the MIT LICENSE shipped from the initial release commit with the wrong copyright holder ("Ahsan Nabi Dar") — a template artifact that survived because nobody re-reads LICENSE on each release. Now reads Copyright (c) 2026 Ahsan Amin, matching the actual repo owner.
• docs/security.md "Defense layers at a glance" mermaid diagram renders on GitHub. The diagram failed to render in GitHub's docs viewer (and on mdnest.dev) with Parse error on line 2: ...Expecting 'SQE', 'DOUBLECIRCLEEND', ... got 'PS'. Cause: node labels like net[Network boundary<br/>(loopback, Tailscale, …)] contain unquoted parentheses inside a [...] square-bracket node, which mermaid interprets as a nested round-bracket shape definition. Wrapped the affected labels in "…" (the mermaid escape for "treat literally") — renders correctly on GitHub, on mdnest.dev, and in our own Live editor.
• Update-available banner surfaces within ~minutes of a release, not a day. Two cadence issues: the backend poller was hitting https://api.github.com/repos/<owner>/<repo>/releases/latest every 24 hours, and the frontend only fetched /api/config once per page load — so a long-running tab would never see a new release at all. Dropped the backend interval to 1 hour (still 1/60th of GitHub's unauthenticated rate limit) and changed the existing 60s /api/config poller on the frontend to refresh appConfig.latestRelease (it only updates state when the release version actually changed, so re-renders stay cheap). End-to-end: a new GitHub Release now shows up in the sidebar footer within roughly one hour of being published, no manual refresh needed. Also: the v3.10.0 GitHub Release was published this cycle — before that the API returned 404 from /releases/latest because we'd been pushing tags without publishing Releases. Both CLAUDE.md's release process and the /mdnest-ship skill now require gh release create as Step 11 alongside git push --tags.

v3.10.0 — Live editor migrated to Crepe + per-workspace last-file memory

The big one: the Live editor is now built on @milkdown/crepe — the same component the Milkdown playground uses. The pre-v3.10 hand-rolled @milkdown/core + commonmark + GFM stack is gone. Crepe brings a block-edit handle (drag + + button + slash menu), native SVG task-list checkboxes, KaTeX math, polished tables with column / row controls, link tooltip, and an image-block upload affordance. All four custom plugins from v3.9 (mermaid live edit, comments, in-cell [ ]/[x] checkboxes, clear-empty-block) port forward; the Catppuccin Mocha look is preserved.

This release also rolls up the v3.9.1 changes (paste-handler priority, browser tab title, Vitest scaffolding) — they shipped on the migration branch as the first commit and never got their own tag.

New features

• Block-edit menu — hover the left margin of any block to get a drag handle and + button. The + button opens Crepe's slash menu (Heading 1-6, code block, math, image, hr, table, …). Typing / anywhere in the doc opens the same menu inline.
• Native task-list checkboxes — top-level - [ ] foo items now render as proper SVG checkboxes (clickable to toggle). Replaces the v3.9.2 hand-rolled topLevelTaskCheckboxPlugin which never looked right.
• KaTeX inline + block math — $inline$ and $$block$$ render via KaTeX. Auto-detected as markdown; no special syntax needed beyond the dollar signs.
• Image upload UI — slash-menu → Image inserts a placeholder block; click to upload or paste a URL. Pasting an image from clipboard works the same way (PNG screenshots, etc.). Uploads go to /api/upload and the rendered <img> resolves through a new proxyDomURL that rewrites the bare-filename markdown src into /api/files/<ns>/<dir>/<file>?token=….
• Auth middleware accepts ?token=<JWT> — for browser GETs that can't set an Authorization header (<img> tags, future <a> downloads). Same validation flow as Bearer; both JWT and mdnest_… API tokens accepted. Without this, images upload-but-never-display because /api/files/… requires auth and the browser image fetch had no way to provide it.
• Per-namespace last-opened-file memory — switching workspaces and switching back restores whichever file you had open in that workspace, with its scroll position. Stored in localStorage under mdnest_last_path:<ns>. URL hashes (#ns/path) still win for bookmarks. Stale entries (file deleted, moved, or renamed) are cleaned up automatically as the operations happen, and the note-loading effect's catch handler clears any that slip through.
• Mermaid auto-detect on paste — pasting raw mermaid source (text that starts with flowchart TD, sequenceDiagram, graph LR, etc.) auto-wraps it in a ```mermaidfence and renders. The detector is intentionally strict: pastes that ALSO look like markdown (contain#headings or|` table rows) are routed to the markdown path instead, so a document that just happens to mention "user journey" or "pie chart" doesn't get swallowed into a mermaid block.

Mermaid rendering preserved

The legacy MermaidBlock React component (Preview / Source toggle, zoom, Fit, Copy, fullscreen viewer, "click any label to edit") stays. Crepe's code-mirror feature is kept enabled (its LaTeX feature depends on it) and a composing plugin wraps the code_block nodeView so language=mermaid blocks render via the React component while everything else falls through to Crepe's CodeMirror block. The fallback was important: without it disabling code-mirror crashed the editor on any $…$ math because LaTeX's editor uses CodeMirror internally.

Paste-fidelity contracts (v3.9.1 priority + new ProseMirror bypass)

• data-pm-slice bypass — when the clipboard HTML carries ProseMirror's slice marker (you copied from another mdnest tab / another Milkdown), the custom paste handler returns early without preventDefault(). Milkdown's native parseSlice then reconstructs the doc with full schema fidelity — table cells keep their inline marks (e.g. **Enigma** in the first column), fenced code blocks keep their language tag, link attributes survive. The previous behavior routed everything through htmlToMarkdown → markdownToSlice, which flattens GFM-specific structure.
• v3.9.1 priority preserved for external clipboards — plain-text-that-looks-like-markdown wins over rich HTML, so Obsidian's - [ ] Foo task lists stay task lists.
• Code-block paste fix — when the cursor is inside a code fence, the custom paste handler returns early so multi-line SQL / JS / etc. pastes as plain text into the block instead of being broken into one paragraph per line.

Table-editing polish

• Single-click cursor placement — Crepe's default behavior turned the first click on a cell into a node-selection over the cell content (caret only appeared on the second click). Wrapped Crepe's table nodeView to let plain mousedowns fall through to ProseMirror's normal cursor placement.
• Visible caret in cells — set caret-color: #89b4fa on the editor; browser's auto was being swallowed by the dark cell backgrounds.
• Inline code wraps inside cells — white-space: pre-wrap; overflow-wrap: anywhere on code inside .milkdown-table-block, so long URLs / endpoint paths break instead of clipping.
• Table-level horizontal scroll fallback — overflow-x: auto on .milkdown-table-block so tables wider than the editor pane scroll horizontally instead of pushing the rightmost columns off-screen.
• Drag handle anchors to the table — Crepe's default block-edit filter explicitly rejects tables, so the drag handle skipped past them. Overrode blockConfig.filterNodes to reject nodes whose $pos is inside a table while accepting the table itself, so clicking the 6-dot handle selects (and lets you drag) the whole table. Selected-state outline (2px #89b4fa) added so the selection is visible against the table's own cell backgrounds.
• Mermaid block selection outline — same fix as tables, on .mermaid-live-container.ProseMirror-selectednode, so Delete-from-handle works discoverably.

From v3.9.1 (subsumed into this release)

• Pasted GFM task lists survive their checkboxes when the clipboard carries both text/plain and text/html — the plain-text-that-looks-like-markdown path beats the HTML path.
• Browser tab title includes the server alias — mdnest (srv-ahsan-mini) so multi-tab users can tell servers apart.
• Vitest scaffolding — frontend/src/__tests__/markdown-fixtures.test.js exercises the paste-priority detection. npm test in frontend/ runs the suite; the pre-push hook gates on it.

Bug fixes

• Mermaid container right-sizes for small diagrams — a 3-node flowchart used to stretch a full editor-width container with ~1300px of empty halo on each side. .mermaid-live-block now uses width: fit-content; max-width: 100% so the block hugs the diagram (with the toolbar's natural width as a floor so its buttons never wrap).
• Task-list spacing — Crepe's default flex gap of 10px + the app-wide p { margin: 0.4rem 0 } were leaking ~13px of stacked margin between rows. Reset paragraph margin inside .children, tightened the gap to 6px, center-aligned the marker box on the text line so bullets / 1. / checkboxes sit on the text baseline.
• Heading hierarchy underline — H1 and H2 keep their #313244 bottom border (previous override only added it to H1).

Internal cleanup

• Legacy LiveEditor.jsx deleted. The four shared plugins (commentHighlightPlugin, clearEmptyBlockPlugin, tableCellCheckboxPlugin, LiveToolbar, plus findAnchorMatches and commentHighlightKey) live in frontend/src/components/live-editor-plugins.jsx. LiveEditorCrepe.jsx imports from there.
• VITE_USE_CREPE build flag removed. Crepe is now the only Live editor; ./mdnest-server rebuild (no env var) ships it. Dockerfile ARG and mdnest-server BUILD_ARGS propagation deleted.
• Bundle size — the Live editor chunk is now ~1.1 MB / ~340 KB gzipped (Crepe + Vue runtime + CodeMirror + KaTeX). Lazy-loaded; the main app bundle is unchanged in size for the initial page load.

Notes

• Crepe brings a Vue 3 runtime (~340 KB raw, ~80 KB gzipped) into the lazy chunk. It is contained — no Vue components are mounted in the React tree; Crepe creates its own Vue app inside the editor's contenteditable root. mdnest as a whole remains a React app.
• The plain textarea ("Basic") editor stays. Some users prefer raw markdown editing, and Basic is also the auto-fallback when the Live editor crashes on a specific file (the existing EditorErrorBoundary catches Live-editor exceptions and flips to Basic).

v3.9.1 — Paste-handler priority + browser tab title + test scaffolding

Bug fixes

• Pasted GFM task lists (- [ ] Foo) survive their checkboxes when the clipboard carries both text/plain and text/html (Obsidian, terminals, most modern apps populate both). The Live editor previously checked the HTML branch first; htmlToMarkdown is lossier than markdownToSlice for task lists because the HTML version typically doesn't carry GFM's data-item-type="task" data attribute, so the DOM round-trip dropped the checkbox semantics and the result landed as a plain bulleted list. Reordered: plain text that looks like markdown (any line starting with #, -, *, >, |, `, [, !) now wins over rich HTML when both are present. HTML conversion remains the path for sources that don't ship markdown (Google Docs, Confluence, web pages). The basic textarea editor already had this priority; this aligns the Live editor with it. Extracted the markdown-detection regex into frontend/src/markdown-utils.js so the two paste handlers share one source of truth.

New features

• Browser tab title includes the server alias. Multiple mdnest tabs (different servers) are now visually distinguishable: mdnest (srv-ahsan-mini) instead of plain mdnest. Falls back to mdnest when no SERVER_ALIAS is configured. Reads from the existing /api/config payload — no backend change needed.

Tests

• Vitest scaffolding for markdown roundtrips (frontend/src/__tests__/markdown-fixtures.test.js). Targets the regression patterns that have actually shipped: paste-priority detection, task-list HTML conversion (the v3.8.0 fix verification), and looksLikeMarkdown edge cases. Runs under jsdom so the htmlToMarkdown path (uses browser DOMParser) is exercised. npm test in frontend/ runs the suite; pre-push hook gates pushes on it passing. Eleven tests today; add a fixture before touching the editor next time so the next regression of this shape is impossible to ship.

Notes

• The deeper consolidation work (unifying Live's tableCellCheckboxPlugin with Preview's DOM-walker checkbox path, memoizing the comment-highlight plugin, schema-level task-list cleanups) is documented in the plan file as Tier 3 future PRs — explicitly NOT in this release. The pattern of "small markdown things keep breaking" needs that work, but each piece is its own change with its own blast radius and should ship one at a time.

v3.9.0 — Tree auto-refresh in single mode + host-side token CLI + path-confusion guard

New features

• mdnest update (and mdnest upgrade) — self-update verb on the CLI. Re-fetches the latest mdnest script from upstream and replaces itself in place; safe on Unix because the kernel keeps the running inode alive until the current invocation exits, so the next call picks up the new code. Reports the upgrade path (vX.Y.Z -> vA.B.C) and short-circuits with "up to date" when current matches latest. --force re-downloads regardless. Closes the discoverability gap where the only update path was remembering the install-cli.sh URL and re-piping it into bash.
• mdnest-server create-token <name> — host-side API token provisioning. Generates a token, persists it to the same tokens.json store the web UI uses, and prints just the raw token to stdout so callers can capture it: TOKEN=$(./mdnest-server create-token foo). Same trust model as reset-password — anyone with shell access to the server can mint tokens, which is by design (server shell = full operator trust). In single mode the token is bound to MDNEST_USER for clarity in logs / UI; in multi-mode this CLI exits with a clear error pointing at the web UI's per-user token flow (web-UI tokens bind to the logged-in user, which the CLI can't disambiguate from the host).

Bug fixes

• Tree auto-refreshes in single mode (and multi-mode without ENABLE_LIVE_COLLAB). Previously the tree was kept in sync via the WebSocket tree-changed event, which only fires in multi-mode + live-collab installs. Without it, a file created from the CLI / MCP / git-sync / another browser tab stayed invisible until the user clicked the Refresh button. Now the frontend polls the tree every 60s when no websocket is connected (skipped when the tab is hidden so backgrounded sessions don't burn requests). Costs one tree GET per minute per active session.
• Better error when creating a file at a path that conflicts with an existing directory. POST /api/note?path=foo/bar.md while a directory foo/bar/ already exists at the same level used to silently create a misplaced sibling — agents (and humans) frequently meant "create a file inside foo/bar/" and didn't realize the path syntax was wrong. Now the backend detects this case and returns 409 with a clear hint: a directory named 'foo/bar' exists at this location — to create a file inside it use path 'foo/bar/<filename>.md'. POSTing directly to a directory path also returns a more informative 409 instead of the misleading "file already exists." Same safePath checks; just a smarter conflict message.
• CLI-minted API tokens validate without a server restart. The token store loads tokens.json once at startup and serves all subsequent validations from memory. The new create-token CLI runs in a one-shot container that writes the file but can't update the running server's in-memory map, so newly-minted tokens were rejected with 401 invalid API token until the next rebuild. Fixed in tokens.go: on a hash miss the validator re-reads tokens.json from disk before giving up. Successful validations stay fast (in-memory hit, no I/O); only misses pay the file-read cost. Same logic mirrored in ResolveAPITokenUser for multi-mode user-resolution misses.

Notes

• This release is on release/v3.9.0.

v3.8.0 — Update notifications, version compare, and multi-IP bind

New features

• "Update available" badge with release notes. The backend now polls the GitHub releases API once every 24 hours and includes the latest release (version, name, publish date, full markdown body) on /api/config. When a newer mdnest is published, a small badge appears next to the version in the sidebar footer. Clicking it opens a modal that renders the release notes inline so you can see what actually changed (features, bug fixes, breaking notes) before deciding to update — not just the version number. Per-version "don't remind me" dismissal is saved in localStorage; a newer release re-arms the badge automatically.
• Compare two versions in the History modal. A new "Compare to:" dropdown above the preview pane lets you pick any other commit, or the live "Current version", to diff against the primary selection. Green-tinted lines exist only in the target, red-tinted lines only in the primary — so when you're considering a restore, you can read exactly what would change and decide before clicking Restore. Diff is line-based (LCS); identical inputs short-circuit to "no differences."
• Comma-separated BIND_ADDRESS for multi-IP binding. BIND_ADDRESS=127.0.0.1,100.73.118.115 now works — useful for binding localhost plus a private overlay address (Tailscale, ZeroTier, VPN) without falling back to 0.0.0.0. Previously a comma-separated value was passed verbatim to a single Docker port mapping, so mdnest-server rebuild failed with invalid IP address. Single-IP behavior is unchanged.

Bug fixes

• Sidebar shows the configured username in single-user mode instead of the literal placeholder "User". Previously the sidebar's UserFooter only had a username to display when /api/me populated userInfo — but /api/me is registered only in multi-mode, and App.jsx actively passed null to the sidebar in single mode (userInfo={isMulti ? userInfo : null}). So an admin signed in via MDNEST_USER always saw the generic "User" label. Fixed in App.jsx by decoding the JWT's sub claim client-side (which already carries MDNEST_USER from mdnest.conf) and synthesizing a minimal userInfo for single mode (with is_super_admin: true since the single-mode user implicitly owns everything). The gate at the Sidebar prop is dropped — userInfo flows through in both modes now.
• Long-press on a file/folder on mobile now shows the same context menu as right-click on desktop. TreeNode.handleTouchStart didn't call e.stopPropagation() (its right-click sibling does). The touch event bubbled up to the parent .sidebar-tree, whose own empty-area long-press handler also scheduled a 500ms timer with target=null. Both timers fired, the empty-area one ran after the file-specific one, so the file menu was rendered for an instant and then immediately overwritten with the empty-area "New Note / New Folder" menu. Fixed by adding e.stopPropagation() at the top of handleTouchStart, mirroring handleRightClick. The empty-area handler still fires correctly when the user long-presses on actual blank space below the tree.
• Patched Go stdlib + golang.org/x/net CVEs. The pre-push hook surfaced four govulncheck findings on the v3.8.0 branch: GO-2026-4982 / GO-2026-4980 (XSS via html/template escaper bypasses), GO-2026-4971 (panic in net.Dial/LookupPort on Windows from NUL bytes — reachable via database/sql and exec.Command lookups), and GO-2026-4918 (HTTP/2 transport infinite-loop on a malicious SETTINGS_MAX_FRAME_SIZE — reachable from updates.Checker and the Firebase admin SDK). Fixed by bumping go.mod's go directive from 1.26.2 → 1.26.3 (which forces the patched stdlib via Go's auto-toolchain), pinning the Dockerfile builder image from golang:1.26-alpine (floating) to golang:1.26.3-alpine so production binaries match, and upgrading golang.org/x/net from v0.52.0 → v0.53.0. govulncheck ./... now reports clean.
• Checkboxes now render inside table cells. GFM's standard table_cell schema admits only paragraph+ content — list items (where Milkdown's task syntax lives) are not allowed children, so | - [ ] X | in a cell fell through to plain [ ] text in both the Live editor and the Preview. Rather than widening the schema (which breaks the ProseMirror tables editing plugin's cell-selection / tab-navigation / paste-rule assumptions), added two cooperating layers that operate on literal [ ] / [x] text inside cells: (a) Live mode — a new tableCellCheckboxPlugin ProseMirror plugin scans table_cell/table_header nodes on every transaction, hides each three-character bracket span via inline decoration (width:0; visibility:hidden so the cursor steps cleanly across), and paints an <input type="checkbox" contentEditable=false> widget decoration at the same position; clicking dispatches a replaceWith transaction that flips the underlying text → checked state persists through normal autosave. (b) Preview mode — DOM post-pass walks every td/th text node, replaces matched [ ]/[x] with checkbox elements, indexes them left-to-right top-to-bottom, and on toggle finds the N-th [ ]/[x] literal in the source markdown and rewrites it via the existing onCheckboxToggle callback (which now accepts a colIndex second argument so in-cell toggles target the specific bracket pair, not the surrounding list-item form). Round-trip is invariant: the underlying text in the document remains literal [ ] / [x], so toMarkdown serializes it unchanged. Works in nested lists and blockquotes too — anywhere the literal brackets appear inside a cell.
• CLI writes (and other HTTP-originated changes) now propagate to a same-user browser tab. The Hub in backend/collab/hub.go was tracking presence in a noteKey -> userID -> *Conn map, and BroadcastFileChanged excluded every connection sharing the originator's userID. So mdnest write @ns/path.md from the CLI as user X excluded X's own browser tab from the resulting file-changed event — the tab kept showing the old content until you clicked away and back. Refactored to noteKey -> set of *Conn (set semantics on the connection pointer): Join/Leave operate on *Conn, presence + countUsers dedupe by User.ID for display purposes, and broadcasts split into two helpers — broadcastToOthers(key, exclude *Conn, msg) for WS-triggered relays (cursor / selection / live content) where the originating tab should not echo, and broadcastToAll(key, msg) for HTTP-triggered events (file-changed / tree-changed) where there is no originating *Conn. Side effects: a user can now have multiple tabs open on the same note without one silently displacing the other; "join"/"leave" presence events now fire only on the user's first/last connection so a second tab doesn't double-count.
• Pasted markdown task lists no longer drop their checkboxes. Pasting - [ ] Foo (from Obsidian, a terminal, or anywhere that puts plain markdown on the clipboard) into the Live editor produced bare text lines without bullets or checkboxes. Root cause: the paste handler used @milkdown/utils's insert(md), which wraps doc.content in Slice(content, selection.openStart, selection.openEnd) — when the cursor sat inside a paragraph (the common case) the slice's open ends were inferred as inline, so block-level lists collapsed to plain inline text on insertion. Switched to markdownToSlice(md) (the DOM-round-trip path) which serializes via the schema's toDOM (GFM's listItemSchema renders <li data-item-type="task" data-checked="…">) and re-parses via DOMParser.parseSlice, whose parseDOM rule extracts the checked attr and rebuilds the slice with proper open ends. Task list items survive the paste round-trip with checkboxes intact.
• Live editor crash no longer blanks the entire app. Some markdown content (deeply nested tables, certain HTML, malformed mermaid blocks) can throw inside Milkdown's Editor.make() or its plugin chain. Until now there was no React error boundary around the Live editor, so the exception propagated up to the app root and unmounted the entire tree — the user saw a blank white screen and had to hard-reload. New EditorErrorBoundary wraps the Live mount: on catch it (1) calls onError, which auto-flips editorMode to basic and persists the choice, (2) shows a banner above the editor pane naming the affected file and pointing to the toolbar toggle for re-trying Live, (3) resets itself when the user navigates to a different note (resetKey={ns}/{path}) so a single bad file doesn't lock out Live for the rest of the session. Manually toggling back to Live for the crashed file clears the banner.
• Files created outside the UI now have a one-tap path to show up in the sidebar tree. Files created via the mdnest CLI, the MCP server, or git-sync only auto-propagate to the browser when (a) live-collab is enabled (multi-mode + ENABLE_LIVE_COLLAB=true), AND (b) the user has a file open at the time — the per-file WebSocket is closed when no note is selected, so a tree-changed broadcast can't reach the client. Single-mode users have no WebSocket at all. The toolbar already has a refresh button for the open-file case, but it's hidden when no file is open, so on mobile (where there's no F5 / Cmd-R) the only recovery was a full browser reload. New refresh button added to the sidebar's tree-control bar (the row with expand-all / collapse-all / show-full-names). Always visible, works on touch, calls the same refreshTree path used after rename/move/delete. Spins for ~600ms on click so the action feels responsive even when the network call is fast.

Mobile UX

• Tree is usable on phones again at deep nesting. Three CSS-side improvements behind the existing @media (max-width: 768px) breakpoint: (1) per-level indent reduced from 0.75rem to 0.4rem per depth — at depth 7 that's ~50px of left padding instead of ~92px, giving the label ~40% more room on a 360px-wide sidebar; (2) sidebar width grows from a flat 280px to min(88vw, 360px), using more of a phone's available width; (3) .tree-row gets min-height: 40px and a slightly larger font on phones, hitting the Apple HIG / Material Design minimum touch-target size so siblings don't get fat-fingered. Indent is now driven by a --tree-depth CSS custom property (set inline by TreeNode.jsx) so the breakpoint can override it cleanly. Long folder/file names ellipsize via text-overflow: ellipsis instead of pushing the chevron off-screen; the native title="" tooltip still shows the full name on hover.
• Move files and folders without drag-and-drop. Touch devices have draggable=false on tree rows (long-press is reserved for the context menu, and HTML5 drag-and-drop on touch interferes with scroll), so there was no way to move a file from one folder to another on a phone. New Move to… entry in the context menu opens a touch-friendly destination picker — flat list of folders in the namespace with hierarchy indent, 44px-tall rows, "Move here" confirm. Filters out invalid destinations (the source itself, the source's current parent, any descendant of the source if the source is a folder). Calls the same POST /api/move endpoint desktop drag-and-drop uses, so collision and permission rules are byte-identical across the two paths. Available on desktop too as an accessibility-friendly alternative to dragging. New MoveToModal.jsx component.
• History modal is readable on phones. The desktop layout puts a fixed-240px commit list next to a flex-1 content pane — on a 330px-wide phone modal that collapses the content to ~60px and stairsteps every line of markdown one word at a time. Below 768px we now stack vertically: the commit list takes the top 28vh, the content pane takes 50vh, and the "Compare to:" dropdown wraps to fill width. Modal grows from min(900px, 92vw) to min(640px, 96vw) for a touch more horizontal room. Desktop layout is unchanged.
• Tree loading shows a spinner instead of "No files yet". On slow connections the previous empty-state copy made it look like a namespace was empty mid-fetch. Now: while getTree() is in flight and the tree is empty, the sidebar shows a centered Catppuccin-blue spinner + "Loading…" text. When the tree is already populated and a refresh fires (after rename/move/create/delete/git-sync), a thin animated progress bar slides at the top of the tree area while the refresh lands — the existing tree stays visible underneath. Both are CSS-only animations, no JS overhead.
• Long folder/file names are readable on phones. Two coordinated changes that don't disrupt the tree's visual rhythm: (1) the sidebar slide-over grows from min(88vw, 360px) to min(94vw, 420px) on phones — since the sidebar overlays the editor anyway when open, reserving more room for the tree makes long names fit without compromise. (2) The toolbar's open-file path was white-space: nowrap + ellipsize-from-end, which on narrow toolbars cut the filename (most informative part) and kept the parent folders. Now split into dirname / basename spans: the dir half shrinks and ellipsizes when squeezed, the basename half has flex-shrink: 0 and stays visible. The basename gets a slightly brighter color + medium weight to read as the primary identifier. Full path is on the title="" attribute for desktop hover reveal. Tree labels keep their single-line ellipsis everywhere — no per-row layout jumps.

Notes

• Update check is opt-out, not opt-in. Default-on so most operators learn about security patches; air-gapped or privacy-sensitive installs can set DISABLE_UPDATE_CHECK=true in mdnest.conf (or UPDATE_CHECK_REPO=<owner/repo> to point at a fork). Failures are logged at info level and never block startup. The backend hits GitHub from one IP per server per day — user IPs are never exposed to GitHub.
• Release-notes payload is capped. Release bodies over 8 KB are truncated in /api/config with a "see full release notes on GitHub" hint, to keep the config payload small even if a future release ships with a giant body.
• No new database migration. This release is config-and-UI only on the multi-mode side; no schema changes.

v3.7.0 — In-app version history with restore (single + multi mode)

New features

• Right-click any note → History. Opens a modal listing the most recent 50 commits affecting that file from the per-namespace git-sync repo, newest first. Selecting a commit shows its content as of that point, and Restore this version writes the old content back through the regular save path (so the v3.6.1 empty-overwrite guard, the ETag conflict check, and the websocket file-changed broadcast all run as usual — restoration is not a separate write path that could carry a new class of bugs). Works in both single mode and multi mode; the only requirement is that git-sync is configured for the namespace. If it isn't, the modal says so clearly with a one-line setup hint.
• Multi-user awareness on restore. When a user clicks Restore in a multi-user install with live collab on, the resulting websocket file-changed event now carries reason: "restored" and the restored-from SHA, so other users currently on the same file see a distinct info-coloured banner ("X restored this file to an earlier version (sha)") instead of the usual yellow conflict banner. Their unsaved local changes are preserved until they choose to reload — same UX shape as the existing conflict banner, deliberately a different colour and copy because a restore is an intentional action by another user, not a divergence.
• Backend endpoints (also new). GET /api/note/history?ns=&path= returns [{commit, unix_ts, author, message}] (capped at 50, newest first); GET /api/note/at?ns=&path=&ref=<sha> returns the file's content at a specific commit. ref is required to be a 7-40 char hex SHA — branch names, HEAD~N, and other git ref forms are rejected to keep the surface predictable. Both endpoints are read-only and gated by the same RequireNsAccess middleware that protects GET /api/note. PUT /api/note accepts a new optional ?restore-from=<sha> query parameter that adds the broadcast tagging without changing any safety logic.

Notes

• No file locks. The user explicitly asked whether this should add a per-file lock primitive (acquire-while-editing). Decision: no. The existing optimistic-concurrency model (ETag + the new info banner + the existing presence bar) handles the multi-user restore case cleanly without introducing the stale-lock / lock-takeover / lock-expiration UX surface that locking inevitably brings. If real users hit conflicts the new banner can't mediate, locking can be designed as its own feature later.
• No diff highlighting for now. The History modal shows old content as a plain <pre> rather than a coloured diff. A diff library can be wired in later if there's appetite; the simpler viewer is enough for v3.7.0.
• No --follow for renamed files. Per-file history starts when the file was named what it's named now. If a file was renamed, its pre-rename commits aren't in the modal — fall back to the GitHub UI for that case. Easy to add later.
• Read-only collaborators can browse history and view old content; the Restore button is disabled for them with a tooltip.

v3.6.1 — Stop the Live editor's undo from erasing your notes

New (small UX add)

• Undo and Redo buttons in the Live editor toolbar. Hidden behind keyboard shortcuts before — and on macOS the redo binding is Cmd+Shift+Z, not Cmd+Y, which trips up plenty of people. Now there are two visible buttons (curved-arrow icons) at the start of the toolbar. Same effect as Cmd+Z / Cmd+Shift+Z on the keyboard. The basic textarea editor already uses your browser's native undo/redo; no toolbar buttons there.

Bug fixes (critical — data-loss prevention)

• Pressing Cmd+Z 2-3 times in the Live editor could silently erase a non-empty note. The data-loss path was a chain of four cooperating defects, and any single one of them would have prevented the loss. We've fixed all four. (1) The frontend's content state defaulted to '' and was reset to '' on namespace change / failed load / browser-nav transitions, so for a brief window the Milkdown editor was initialized with empty content even when a real (non-empty) note was about to load. That empty state ended up as a reachable entry in ProseMirror's undo stack — pressing Cmd+Z walked back through your typing and then into that empty load. (2) <LiveEditor> had no key prop, so the same Milkdown instance carried across note switches and Cmd+Z could walk into another note's history. (3) handleContentChange's 800ms debounced autosave fired unconditionally — when the editor briefly held empty content, the autosave dutifully committed empty bytes to disk. (4) The backend's PUT /api/note had no guard against truncating a non-empty file to empty. Fixed in App.jsx (initial state is now null until a note is loaded; setting null during transitions instead of ''; autosave skips when newContent === '' and the loaded content was non-empty; key={ns/path} on <LiveEditor> and <Editor> so each note gets a fresh instance with its own undo stack), LiveEditor.jsx (only mounts when content is a real string, so the empty-during-load transition no longer enters the undo stack), and notes.go (refuses to overwrite a non-empty file with an empty body unless the request explicitly passes ?allow-empty=1 — autosave never does, so the silent-truncation path is closed even if every layer above somehow fails).
• Recovery for already-lost content: if your install runs the optional git-sync sidecar (default cadence 600s), every note has a complete commit history in the per-namespace git repo. Browse the repo on GitHub to find a sync: <UTC-timestamp> commit before the destructive undo, view the file at that commit, and copy the content back. We are surfacing this as an in-app "Version history" button in v3.7.0 so future recovery doesn't require the GitHub UI.

v3.6.0 — Admin password reset (UI + host CLI)

New features

• Superadmins can reset another user's password from the Admin Panel. Each non-superadmin row in Admin → Users now has a "Reset password" button (visible only to superadmins, only when USER_PROVIDER=local). The dialog asks for a new password twice; on submit the target user's must_change_password flag is set so their next login is gated on picking their own password before they can do anything else (the existing forced-change flow in Login.jsx already handled this case for invited users — we just reuse it).
• Resetting another superadmin's password from the UI is intentionally blocked. That would be a lateral-escalation primitive — one compromised superadmin could lock out every other superadmin in a single click. The UI button is hidden for superadmin rows and the /api/admin/reset-password endpoint returns 403 if the target's role is superadmin. The legitimate recovery case (a colleague forgot their superadmin password) is handled by the new host-side CLI below.
• mdnest-server reset-password <email> — host-shell command for resetting any user's password, including superadmins. Prompts for the new password with hidden input (twice), pipes it via stdin to a one-shot backend container so the password never appears in argv or shell history. Validates AUTH_MODE=multi + USER_PROVIDER=local and refuses otherwise. Same must_change_password=true guarantee — the temp password is single-use.

Notes

• No database changes. The must_change_password column has existed since the original multi-mode work, so this release is additive: any existing schema works unchanged.
• Federated providers (firebase, sso) reject the new endpoint — identity is owned by the IdP. Reset there.

v3.5.4 — Fix renamed file vanishing from the sidebar when extension is dropped

Bug fixes

• Renaming a note to a name without a file extension silently hid it from the tree. The sidebar only lists files with a recognized text extension (.md, .txt, .json, .sql, .csv, .yaml, .yml, .markdown) — that's tree.go's textExtensions filter. The rename prompt accepted any string and passed it straight to /api/move, so typing foo while renaming foo.md wrote foo to disk and the file disappeared from the sidebar: still on disk, no error, no warning. Fixed in App.jsx by mirroring doCreateNote's extension-preserving behaviour — when the target is a file and the typed name contains no ., the original extension is auto-appended. Folders are exempt; explicit extension changes (foo.md → foo.txt) still work as before.

v3.5.3 — Fix bogus 409 "modified by another user" on first save of new notes

Bug fixes

• "This file was modified by another user" 409 on the first save of a freshly-created note. Notes carry an invisible <!-- mdnest:UUID --> marker so comments survive renames; the marker is lazy-injected by EnsureNoteID the first time a comments endpoint touches a file. ExtractNoteID returned the body in two different shapes — bytes-as-is when no marker, with a trailing \n normalization when the marker was present — so the ETag computed by getNote before the lazy injection didn't match the ETag computed by updateNote after. The frontend's first autosave hit the conflict path with no actual conflict, the user lost their typed content on refresh. Same defect also fired on every save after the first (since newETag = sha256(body) ignored the same normalization). Fixed in notes.go by adding canonicalForETag, a helper that drops trailing newlines from clean note content. Wrapped around all three ETag call sites (getNote, updateNote currentETag, updateNote newETag) so the hash is identical regardless of whether the marker has been injected yet — the race becomes mathematically irrelevant. Bytes on disk and bytes returned to the editor are unchanged; only the hash input is canonicalized. Genuine conflicts (real concurrent edits) still 409 correctly.

v3.5.2 — Fix empty tree for superadmin in multi mode

Bug fixes

• Superadmin users saw an empty file tree in multi-user mode. The grant filter in the tree handler only bypassed filtering for role="admin" (namespace-admin), not role="superadmin". Since superadmins have no explicit grant rows (they're meant to have implicit full access), filterTreeByGrants stripped every node — returning an empty root. Fixed by adding the "superadmin" role check alongside "admin" in tree.go.

v3.5.1 — Go 1.26 bump for stdlib CVEs

Security

• Bumped Go to 1.26 (was 1.25). Clears five stdlib vulnerabilities flagged by govulncheck against the 1.25 line:
  • GO-2026-4865 — html/template JS context tracking bug (XSS)
  • GO-2026-4866 — crypto/x509
  • GO-2026-4870 — crypto/tls KeyUpdate DoS
  • GO-2026-4946 — crypto/x509 slow policy validation
  • GO-2026-4947 — crypto/x509 slow chain building
• backend/go.mod: go 1.25.0 → go 1.26.2. backend/Dockerfile: golang:1.25-alpine → golang:1.26-alpine (the moving tag tracks the latest 1.26.x patch so future fixes land on rebuild without manual bumps).
• No code changes required for the bump — go mod tidy was a no-op, go build and go vet clean, and the production-style docker compose build --no-cache succeeded against the new image.

v3.5.0 — Namespace-scoped Admin role + SuperAdmin + token access scoping

Breaking changes

• Existing role='admin' users are migrated to role='superadmin' on first startup of v3.5.0. They keep current behaviour — global access to every namespace, every user-management endpoint, every grant. This is a one-shot rename done by migration 007 and only fires when the migration runs the first time. No action required from operators.
• The new role='admin' is namespace-scoped, not global. A user with role='admin' can only manage the namespaces listed for them in the new namespace_admins table — they invite users into those namespaces, manage grants on them, promote co-admins, and trigger git-sync for them. They cannot delete users, change anyone's role, reset 2FA, or sync globally — those are SuperAdmin-only.
• API tokens no longer get a system-wide admin bypass. Pre-v3.5.0, an admin's token bypassed every permission check. Post-v3.5.0 a token resolves to its creator's current scope at request time: superadmin tokens are still global, namespace-admin tokens work only on their owner's admin namespaces, collaborator tokens work only on their owner's grants. Revoking a user's grant immediately revokes their tokens for that namespace too.
• ADMIN_EMAILS now auto-promotes to superadmin (was admin). This preserves the operator-bootstrap intent across the role rename.

Features

• Three-tier role hierarchy. SuperAdmin (global) / Admin (namespace-scoped) / Collaborator (grants only). The model lets a multi-tenant deployment have one superadmin operator and per-team admins who can run their own namespace without seeing other teams' data.
• namespace_admins(user_id, namespace, granted_by, created_at) table. Migration 007 creates it; the PermissionChecker consults it on every role='admin' request to decide whether the namespace is in scope. The hot path is a single-row EXISTS query.
• New endpoints /api/admin/namespace-admins — GET ?ns=<n> lists admins of a namespace, POST {user_id, namespace} promotes (auto-bumps users.role from collaborator to admin, auto-creates a permission='write' grant on /), DELETE ?user_id=<id>&ns=<n> demotes (auto-reverts to collaborator if no other admin namespaces; the auto-grant is left in place so access doesn't disappear by surprise).
• /api/me exposes is_super_admin and admin_namespaces so the frontend can scope the admin panel without an extra round-trip on every page load.
• /api/admin/users is filtered by caller scope. SuperAdmins see all users; namespace admins see only users with grants or namespace_admins entries on their own namespaces (plus self).
• /api/admin/grants is filtered by caller scope. Same model: superadmin sees all, namespace admin sees only their namespaces. Create / update / delete return 403 if the target grant isn't in the caller's admin scope.
• Reset 2FA, delete user, change role are SuperAdmin-only. Promoting between superadmin/admin/collaborator globally requires SuperAdmin. Promoting another user as namespace admin only requires admin scope on the target namespace.
• Sidebar admin scope hint. Namespace admins see a yellow "Admin of: " badge at the top of the admin panel so they know what they're managing.
• New "Namespace Admins" tab in the admin panel. Pick a namespace, see who admins it, promote any non-superadmin user, demote with one click. Visible to anyone with the panel; the backend scopes both reads and writes.

Configurable grant depth

• GRANT_MAX_DEPTH in mdnest.conf caps how deep into a namespace tree an admin can scope a grant. / is depth 0 (always allowed), /foo is 1, /foo/bar is 2. New grants beyond the limit are rejected at POST /api/admin/grants with a 400 explaining the depth and the configured limit. Existing rows are grandfathered — only new INSERTs are checked. Default 3. Set to 0 for no limit. The PathPicker dropdown in the admin UI reads the same value from /api/config and hides too-deep folders so admins can't pick something the API will reject.

Dev-only

• INSECURE_DEV_LOGIN backdoor for local SSO testing. When set to true in mdnest.conf, the backend registers POST /api/auth/dev-login which mints a 30-day session JWT for any existing user by email — completely bypassing the IdP. Identity rules match SSO (no auto-provisioning, blocked users still rejected). Off by default; the route 404s when the flag is unset. The default sign-in page is unchanged (still strict SSO); the bypass is reachable only by manually navigating to /?login=dev. While enabled, every authenticated page renders a sticky red warning banner, and the backend logs a multi-line warning at startup. Strictly for local development — never enable on a non-localhost deployment. New LoginDev.jsx component, devLoginEnabled field on /api/config.

Internal

• New backend/store/namespace_admins.go: NamespaceAdminStore interface + Postgres impl with Add, Remove, IsAdminOf, ListByUser, ListByNamespace, CountByUser. Add is idempotent via ON CONFLICT DO NOTHING so promote re-runs are safe.
• backend/middleware/permission.go: new constructor NewPermissionChecker(grantStore, nsAdminStore) and a hasAdminScope(uc, ns) helper. The three places that used to short-circuit on Role == "admin" now go through it.
• backend/middleware/admin.go: RequireAdmin now means "any admin role"; new RequireSuperAdmin for the global gate. IsSuperAdmin(ctx) helper added.
• backend/handlers/admin.go: every method now scopes through callerCanAdminNamespace / callerAdminNamespaces. ensureNotLastAdmin → ensureNotLastSuperAdmin (only superadmins are deadlock-load-bearing).
• backend/handlers/sync.go takes an nsAdminStore and returns 403 when the caller isn't allowed to sync the requested namespace.
• backend/handlers/tokens.go: listTokens and revokeToken no longer give role=='admin' system-wide visibility — superadmin only. Owners always see / revoke their own.
• backend/store/grants.go: + GetGrant(id) so admin handlers can authorize the action against the target grant's namespace.
• Frontend: App.jsx derives isSuperAdmin and adminNamespaces from /api/me; threads them into AdminPanel. The panel hides global actions (Cycle role, Delete user) for non-superadmins, locks the Invite namespace dropdown to admin scope, adds the Namespace Admins tab. New api.js helpers: adminListNamespaceAdmins, adminAddNamespaceAdmin, adminRemoveNamespaceAdmin. adminInviteUser accepts a namespace.

v3.4.0 — Corporate SSO + Federated Identity

Features

• mdnest-server reload — new lightweight subcommand for config-only edits. Regenerates .env + docker-compose.yml from mdnest.conf and force-recreates backend + frontend (and git-sync if enabled) so they re-read the new env. No image rebuild — ~10s vs rebuild's 60-90s. Postgres and other persistent services are left untouched. Use after editing mdnest.conf (e.g. flipping USER_PROVIDER, adding a MOUNT_*, rotating an SSO secret).
• mdnest-server rebuild always force-recreates app containers. Previously the default rebuild relied on the --no-cache backend build to change the image hash, which usually triggered recreation but could miss conf-only changes that produced an identical binary. Now rebuild always passes --force-recreate to compose up -d for backend + frontend, guaranteeing the new .env is read. Postgres + git-sync still stay running. --full continues to nuke everything for the rare cases that need it.
• Backend Dockerfile: BuildKit cache mounts. /root/.cache/go-build and /go/pkg/mod are now persisted across builds. After v3.4.0 added Firebase Admin SDK + grpc + protobuf, a clean go build was taking 180-235s on a small EC2. With cache mounts the first build is unchanged, but every subsequent rebuild reuses the precompiled package archives → typically 10-30s for source-only changes. The default rebuild also drops --no-cache to take advantage of layer caching too; rebuild --full keeps --no-cache for the paranoid case.
• mdnest-server disables BuildKit attestations. Sets BUILDX_NO_DEFAULT_ATTESTATIONS=1 at the top of the script so every build skips SLSA provenance + SBOM generation. These are designed for images pushed to a registry; we build locally, so they're pure overhead and can hang at "resolving provenance for metadata file" depending on the host's network conditions. Skipping them never affects image content or behaviour.
• Corporate SSO via generic OIDC. New USER_PROVIDER=sso mode (requires AUTH_MODE=multi). Users sign in through your IdP (Google Workspace, Okta, Microsoft Entra, Keycloak, Auth0 — anything that speaks OIDC discovery). Backend uses coreos/go-oidc + oauth2 with PKCE; state/nonce/code-verifier carried in a short-lived HMAC-signed cookie. The IdP owns MFA, so mdnest's local 2FA is skipped in this mode. See docs/sso-setup.md.
• Email-gated sign-in, no auto-provisioning. An SSO sign-in only succeeds if the email is already in the mdnest users table (invited by an admin). Role, grants, and blocked flag stay in Postgres. Rejection paths redirect back with #sso_error=<code> for the frontend to surface.
• Optional SSO_ALLOWED_DOMAINS allowlist for corporate-domain-only sign-in.
• Firebase identity (peer mode). USER_PROVIDER=firebase is also available for teams that prefer Firebase Auth + Firestore-backed shared TOTP. Docs: docs/firebase-setup.md. Chosen mode is exclusive per server; Firebase is not required and carries no overhead when not enabled.
• store.TOTPStore interface. TOTP handlers, login flow, and admin 2FA reset now route through an interface with Postgres and Firestore implementations. Makes the 2FA surface swappable and explicit.
• totp_enabled JWT claim. Populated at login-issue time so the frontend can render "Enable 2FA" vs "Manage 2FA" without hitting the TOTP store on every request. Real 2FA enforcement still runs against fresh state at login.
• ADMIN_EMAILS bootstrap. Comma-separated list in mdnest.conf is reconciled into role='admin' on every startup. Removals are NOT auto-demoted — operator demotes explicitly.
• Profile name + avatar from the IdP. SSO callback now reads the name and picture OIDC claims from the ID token. Avatar is mirrored into a new users.avatar_url column on every login (picture URLs rotate at the IdP). Username is filled in once when the row's value is empty — admin-set usernames are never overwritten. The sidebar renders <img> from avatar_url with a graceful fallback to initials when the image fails to load. New users created by the SQL-INSERT bootstrap path get their real face + name automatically on first sign-in instead of "User" / "?". Migration 006 adds the column; additive, safe in all modes.

Internal

• New backend/sso/ package: OIDC relying-party with PKCE, cookie-based state, domain allowlist, SanitizeFromPath to prevent open-redirect abuse through the post-login from= param.
• New backend/handlers/sso.go wiring two routes: GET /api/auth/sso/start, GET /api/auth/sso/callback. Only registered when ssoClient != nil, so misconfigurations 404 cleanly.
• New backend/firebase/ package: Firebase Admin SDK wrapper + Firestore TOTP store. Only instantiated when USER_PROVIDER=firebase.
• Migration 005: users.firebase_uid TEXT UNIQUE, DROP NOT NULL on password_hash / username, indexes on firebase_uid and email. Additive; safe on local-mode databases.
• Frontend: LoginSSO.jsx for SSO mode, LoginFirebase.jsx for Firebase mode, unchanged Login.jsx for local mode. App.jsx picks the right one from /api/config.userProvider. Hash-fragment token handoff (#sso_token=…) for the SSO callback.
• Settings.jsx hides the "Credentials" tab in both federated modes (no local password to change).
• setup.sh validates SSO / Firebase config at rebuild time, emits env vars into .env, mounts Firebase JSON files when needed.
• Dockerfile: Go image bumped to golang:1.25-alpine (Firebase Admin SDK requires Go 1.25+).

v3.3.1 — Preview crash hotfix + CLI server-alias cleanup

Breaking (CLI)

• No more silent @default alias. mdnest login <url> <token> without an explicit @alias used to create an alias literally named default — which hid which server was which in copy-path URIs. Now the CLI fetches /api/config and uses the server's SERVER_ALIAS automatically. If the server doesn't advertise one, login refuses with an actionable error: either pass @alias explicitly or set SERVER_ALIAS=<name> in the server's mdnest.conf and rebuild.
• @default is rejected as an alias name at login time.
• New mdnest rename @old @new command so users stuck with an existing @default alias can fix it in one step. Updates the default-server pointer if it referred to the old name.
• Existing @default aliases keep working (backward-compat for scripts) but print a one-line deprecation nudge per invocation pointing at rename.

Server

• SERVER_ALIAS soft-required. Backend logs a WARNING on startup if it's unset, and setup.sh prints a warning at rebuild. Not a hard failure (existing installs keep running), but CLI users on unnamed servers have to pass @alias manually until it's set.
• mdnest.conf.sample now ships with SERVER_ALIAS=mdnest uncommented and a comment explaining why.

Fixes

• Preview crash on task lists with nested content — clicking Split or Preview view on a file whose task list contained nested blocks (sub-lists, multi-paragraph items) threw Token with "list" type was not found from marked and took the preview tree down. Root cause: the custom listitem renderer called parseInline on block-level tokens. Fixed by dropping the override entirely — marked v15 already renders GFM task lists as <li><input type="checkbox">, and we re-wire those in the DOM post-pass.
• Preview crash on headings / paragraphs / tables — follow-up regression from the first fix attempt. Passing a plain-object renderer via the per-call marked(src, {renderer}) option replaces the default renderer entirely in marked v15, instead of merging with it. Any token type not explicitly overridden (heading, paragraph, table, blockquote, etc.) crashed with this.renderer.X is not a function. Fixed by switching to new Marked().use({renderer: {...}}), which merges with defaults.

Robustness

• Preview error containment — renderMarkdown is now wrapped in try/catch, and the Preview component is wrapped in a PreviewErrorBoundary. A malformed note (or any future renderer bug) now shows a readable error panel inside the preview pane instead of unmounting the whole app. The boundary auto-resets when the user navigates to a different note, so a single bad file doesn't permanently black out the pane.
• View-mode toggle visible without a file open — previously the Editor / Split / Preview and Basic / Live toggles were hidden when no file was selected. That trapped users in a bad mode after a crash: every file they tried to open re-triggered the same render path. Toggles are now always visible so users can pre-switch to a safe mode before opening the next file.

v3.3.0 — Inline Comments

Features

• Inline comments — select text in the Live editor and attach a comment to it. Commented text gets a persistent bright-yellow highlight so reviewers see what's been discussed at a glance. Highlights do not appear in print or export.
• Threaded replies — each comment can carry a conversation. Click Reply under any active thread to add a message; Enter sends, Esc cancels. Replies stack inside the parent card.
• Comment sidebar — slide-out panel on the right with active and resolved threads. Each thread shows the quoted anchor text, author, relative time, and actions (Go To, Reply, Resolve, Delete).
• Clickable highlights — click yellow text in the editor to open the sidebar and pulse the matching comment card into view.
• Go To with pulsing flash — the Go To button scrolls the commented text into view and plays a ProseMirror decoration flash on it, so the location is obvious even in long documents. Position tracking is done by ProseMirror itself, so scrolls and edits don't desync it.
• Cross-mark anchor matching — highlights work even when the commented selection spans inline marks (bold, italic, inline code, links). The search concatenates every text node with position mapping, rather than walking nodes one at a time.
• UUID-anchored storage — each note carries an invisible <!-- mdnest:UUID --> marker at the bottom, stripped on GET and re-injected on PUT. Comments are stored at <namespace>/.mdnest/comments/<uuid>.jsonl, so moving or renaming a file keeps its comments attached.
• Direct-link loading — comments now load correctly when opening a note via URL hash or browser back/forward, not just when clicked in the tree.
• Requires multi-user + live collab — comments need both AUTH_MODE=multi (for real author identity) and ENABLE_LIVE_COLLAB=true (for the WebSocket hub). Without either, the UI is hidden and the /api/comments route is unregistered.

Fixes

• Floating Comment popup at wrong positions — suppressed when the triggering mouseup/keyup comes from outside the editor (e.g. clicking Go To in the sidebar no longer resurrects the popup).
• Single-user / collab-off mode crash on comment — the comment UI was showing in single-user mode and in multi-user installs that disable live collaboration (ENABLE_LIVE_COLLAB=false), even though the feature requires real user identity and the WebSocket hub. Comments are now gated on liveCollab on both the frontend (no icon, no popup, no sidebar, no API calls) and the backend (/api/comments route is only registered when live collab is on).

v3.2.2 — Responsive Mobile & Stability

Fixes

• Mobile responsive rendering — uses React isMobile state instead of CSS-only for editor/preview switching. At 768px breakpoint, only one wrapper renders (editor OR preview), preventing blank screens and split-view glitches.
• Mobile mobileView sync — syncs with desktop viewMode on first load so preview mode works on mobile.
• False update banner (v1.0) — removed second fallback in api.js that returned version '1.0' when config failed.
• WebSocket status hidden when no file — "Offline" no longer shows when no file is selected.

v3.2.1 — Performance & Stability

Fixes

• Server overload (critical) — GET requests on /api/note triggered BroadcastTreeChanged to all WebSocket clients, causing an infinite loop. Now only broadcasts on mutating requests (PUT/POST/DELETE).
• WebSocket ghost reconnections — switching files left stale onclose handlers that reconnected to the old file, stacking connections. Fixed with connection ID tracking.
• Tree filtering — non-markdown files (Postman JSON, binaries) excluded from tree. Supports .md, .txt, .json, .sql, .csv, .yaml. Files >5MB skipped. Empty directories still shown.
• Removed 15-second tree polling — WebSocket tree-changed events handle tree updates. Eliminated 80+ redundant requests/min with 20 users.
• Note poll reduced — 10s → 60s. WebSocket file-changed is real-time, poll is just a fallback.
• Tree-changed debounce — 1-second debounce prevents rapid-fire tree refreshes from bulk operations.
• PathPicker cache — admin panel directory picker caches tree API for 30s, preventing N duplicate calls.
• ETag conditional (304) — note GET returns 304 Not Modified when content unchanged, saving bandwidth.
• Backend always rebuilt --no-cache — prevents stale Docker cache from deploying old binaries.
• False update banner — no longer shows "v3.2.1 → v1.0" when backend is slow/unreachable.
• iPad viewport — 100dvh accounts for mobile browser bar. View mode toggle visible on tablets.
• Copy path URI — uses mdnest://@alias/namespace/path format for LLM readability.
• WebSocket status text — shows "Live", "Reconnecting", or "Offline" next to the status dot.
• CLI login warning — warns before overwriting default server with a different URL, suggests aliases.

v3.2.0 — Two-Factor Authentication & Account Security

New Features

• Two-Factor Authentication (TOTP) — authenticator app support (Google Authenticator, Authy, 1Password). QR code setup, recovery codes, admin reset.
• Mandatory 2FA — REQUIRE_2FA=true in config forces all users to set up 2FA. Guided setup flow during login with QR code + step-by-step instructions.
• Shared 2FA across servers — export-2fa / import-2fa commands let admins share TOTP secrets across multiple mdnest instances. One authenticator entry for all servers.
• Forced password change — new users must change their password on first login (must_change_password flag).
• Block/unblock users — admin can block users, preventing login with a clear error message.
• Multi-step login flow — password → forced password change → 2FA setup/verify → JWT. Each step shows a clean UI.
• 30-day sessions — JWT expiry extended from 24 hours to 30 days (safe with 2FA).
• Auto-migrate on rebuild — ./mdnest-server rebuild automatically runs database migrations for multi-user mode.

Fixes

• Mermaid text colors — injected SVG <style> override ensures light text on all diagram types. No more black text on load or color toggling on click.
• Mermaid label click — diagram-type agnostic click handler. Works on all mermaid types (sequence, flowchart, class, etc.) by finding nearest <g> group text instead of checking specific CSS classes.
• Mermaid label replace — handles <br/> line breaks at any word boundary via brute-force matching.
• WebSocket stale closure — collab message handler used stale namespace/path from closure, causing one user's saves to disrupt another user's view. Now uses refs for current values.
• Editor mode reset — switching files no longer resets Live mode to Basic. Editor/view mode are global user preferences, not per-file.
• Table cell selection — multi-cell selection now visually highlights in Live editor (blue overlay).
• Table row paste — copying table rows and pasting inside an existing table inserts rows after the current row instead of creating a new table.

Config

• REQUIRE_2FA=true|false — require all users to set up 2FA (default: false)
• TOTP_ISSUER=name — issuer name shown in authenticator app (default: mdnest)

v3.1.8 — Developer Experience & Security

New Features

• Pre-push git hook — verifies frontend/backend compile, npm audit, govulncheck, lock file integrity, and version consistency before every push. Install with ./mdnest-server dev-setup.
• remove-namespace command — lists namespaces, removes config entry and deploy key. Files on disk are NOT deleted.
• Improved add-namespace — two clear paths (GitHub clone or local directory), SSH verification, auto-clone, branch name prompt, never exits on bad input (re-prompts instead), auto-creates subdirectories for non-empty paths.

v3.1.7 — Mermaid Improvements & UX Polish

New Features

• Per-file preferences — each file remembers its view mode (editor/split/preview), editor mode (basic/live), and scroll position in localStorage. Survives page refresh.
• Default to Live editor — new files open in Live editing mode instead of basic textarea.
• Sync status visible to all users — "Synced 5m ago" green dot shown to collaborators, not just admins. Sync trigger button stays admin-only.
• add-namespace command — ./mdnest-server add-namespace walks through creating a namespace: directory, git init, deploy key generation, remote URL setup.

Fixes

• Mermaid color revert on label edit — mermaid.initialize was only in Preview.jsx; Live mode used default pastel theme. Moved to shared mermaid-config.js.
• Mermaid text contrast — smart post-processing detects parent node fill brightness and forces dark or light text for readability.
• Mermaid label click for multi-line labels — labels with <br/> line breaks now correctly detected and replaced in source.
• Mermaid code consolidated — theme config, initialization, and text color fix all in one shared file.
• Refresh icon moved — now appears right after the file path instead of at the end of the toolbar.
• Raw editor paste fix — pasting markdown text no longer wraps it in triple backticks.
• Git-sync fresh repos — first push uses --set-upstream for newly created namespaces.
• Git-sync SSH alias auto-fix — detects host:path format (without git@) and rewrites to [email protected]:path.
• Rebuild force-recreates git-sync — volume-mounted services always restart on rebuild.

v3.1.1 — Critical Save Fix

Fixes

• Live Editor stale onChange (critical) — switching files in Live mode caused 409 conflicts and lost changes. Milkdown's markdownUpdated listener captured onChange once at editor creation, so saves went to the wrong file path after switching. Fixed with onChangeRef that always points to the latest callback.
• MutationObserver phantom saves — Milkdown's async MutationObserver fired markdownUpdated after replaceAll, triggering phantom saves that changed file ETags. Now suppressed until real user interaction (keydown/mousedown).
• Auto-refresh poll race condition — in-flight getNote responses from the previous file could overwrite the new file's state after switching. Now discards stale responses.
• Save timer stale closure — saveTimer was React state (stale in closures), changed to ref. Cleared on file switch.
• Version update banner — active sessions show a blue banner when server is updated, with "Refresh Now" button.
• Browser cache on deploy — nginx serves index.html with no-cache so hard refresh picks up new bundles.

v3.1.0 — Mermaid Zoom & Live Toolbar

New Features

• Mermaid zoom controls — − / + / Fit buttons in the mermaid toolbar. Zoom 20%–300% via CSS transform. Small diagrams render at natural size, large diagrams fill container width.
• Rich text formatting toolbar — Live mode now has a full toolbar: Bold, Italic, Strikethrough, Code, H1/H2/H3, Bullet/Numbered list, Blockquote, HR, Link, Code block, Table, +Row/+Col/-Row/-Col.
• Copy mermaid code — Copy button in mermaid toolbar copies the source code to clipboard.
• Version update banner — when the server is updated, active sessions show a blue banner with current → new version and a "Refresh Now" button. Polls /api/config every 60s.

Fixes

• Live Editor stale onChange (critical) — switching files in Live mode caused 409 conflicts and lost changes. Root cause: Milkdown's markdownUpdated listener captured onChange once at editor creation, so saves went to the wrong file path. Fixed by using a ref that always points to the latest callback.
• Auto-refresh poll race condition — in-flight getNote responses from the previous file could overwrite the new file's state. Now discards stale responses via a poll key check.
• Save timer stale closure — saveTimer was React state (stale in closures). Changed to useRef and cleared on file switch.
• Smart mermaid sizing — uses SVG viewBox dimensions (reliable) instead of width attribute (unreliable). Small diagrams centered at natural size, large diagrams fill container.
• Mermaid fullscreen — was broken because modified SVG (stripped attributes) was passed to viewer. Now stores and passes original unmodified SVG.
• Scroll position on view switch — switching between editor/split/preview modes now preserves scroll position.
• Browser cache on deploy — nginx now serves index.html with no-cache header so hard refresh always picks up new JS bundles.

v3.0.0 — Live Rich Editor

New Features

• Live editor mode — Obsidian-style rich editing powered by Milkdown (ProseMirror). Markdown renders inline as you type: bold shows bold, headings render as headings, lists format in place. Toggle between Basic (plain textarea) and Live mode from the toolbar.
• Interactive table editing — click into table cells to edit. Toolbar buttons to insert tables, add/remove rows and columns. Tab between cells.
• Mermaid inline rendering — mermaid code blocks render as diagrams in-place in Live mode with Source/Preview/Fullscreen buttons. Click any node or edge label to edit it directly on the diagram.
• Clickable checkboxes in edit mode — task list checkboxes work in Live mode without switching to preview.
• Rich paste — paste from Google Docs, Confluence, or any rich source into Live mode and it inserts as parsed markdown nodes (headings render as headings, not # text).
• Scroll sync — editor and preview scroll proportionally in split view.
• Collapsible headings — click the toggle icon on any heading in preview to collapse/expand that section. Expand All / Collapse All buttons in preview toolbar.

Improvements

• Lazy-loaded Live editor — Milkdown only downloads when you switch to Live mode (462KB chunk). Main bundle stays at 311KB for fast initial load.
• Smart backspace — empty headings/blockquotes convert to paragraphs on single backspace in Live mode.
• Text selection in mermaid — can select and copy text from rendered mermaid diagrams in preview mode. Fullscreen expand moved to a hover button.
• Editor mode per view — Live mode preference is separate for editor-only view. Split view always uses Basic mode.
• Heading collapse — only the toggle icon (not heading text) triggers collapse. Expand All properly shows all nested content.
• Copy buttons — headings show a clipboard icon on hover (copies heading text). Code blocks show a "Copy" button on hover (copies code content).
• Table delete controls — separate Del Row, Del Col, Del Table buttons using direct ProseMirror commands (cursor in cell is enough, no need to select).
• Scroll position persistence — each document remembers its scroll position. Switch between documents and your reading position is restored.
• Mermaid label editing for sequence diagrams — participants, messages, and other sequence diagram labels are clickable alongside flowchart nodes.
• Auto-expanding label editor — mermaid label input grows/shrinks with text content.

Dependencies

• Added: @milkdown/core, @milkdown/ctx, @milkdown/react, @milkdown/preset-commonmark, @milkdown/preset-gfm, @milkdown/plugin-listener, @milkdown/plugin-history, @milkdown/plugin-clipboard (all v7.20)
• Existing: marked (preview/basic mode), mermaid (diagrams) unchanged

New Files

• frontend/src/components/LiveEditor.jsx — Milkdown editor wrapper with table toolbar, mermaid node view, paste handling
• frontend/src/components/MermaidBlock.jsx — React component for inline mermaid with Source/Preview toggle and click-to-edit labels

v2.1.0 — Multi-Server CLI + Git Sync Fix

New Features

• Multi-server CLI — manage multiple mdnest servers with @alias paths. mdnest login @work <url> <token>, then mdnest read @work/engineering/docs.md. Single-server users see zero change.
• Flat CLI commands — mdnest read, mdnest list, mdnest search etc. (no more mdnest note prefix needed, though it still works).
• mdnest servers — list all configured servers with versions and reachability.
• Copy Path includes server alias — right-click Copy Path in the web UI gives @work/namespace/path when SERVER_ALIAS is set, directly pasteable into the CLI.
• Collapsible headings in preview — click any heading to fold/unfold the section. Expand All / Collapse All buttons in preview toolbar.
• Git sync status indicator — green dot + "Synced 5m ago" in sidebar header.
• Sync button commits + pushes — pressing sync now does git add + commit + pull + push (was pull-only before).

Fixes

• Git-sync SSH key — the git-sync sidecar now falls back to SSH_KEY_PATH when git-sync/keys/ is empty. One SSH key config works for both the sync button and the auto-sync cycle.
• Mermaid inline sizing — 50% on desktop, 90% on mobile. Removed inline style override.
• Sync button reloads current note — not just the tree.
• Tree arrows bigger and blue — more visible expand/collapse indicators.
• Hard refresh on login — clean state, no stale data.
• Removed "no key" warning — was confusing for users who don't need git pull.

Configuration

• SERVER_ALIAS — optional, sets the @alias used in CLI paths and Copy Path.
• SSH_KEY_PATH — now used by both the backend sync button AND the git-sync sidecar.

v2.0.1 — Patch Release

Fixes

• Drag-drop to ancestor directories — moving items up the tree (e.g. subdir to parent) was blocked by an overly aggressive guard. Fixed.
• SSH key mount for git pull — sync button now supports SSH authentication. Set SSH_KEY_PATH in mdnest.conf pointing to a passphrase-free deploy key.

New Features

• HTML-to-Markdown paste — copy from Google Docs, Confluence, Notion etc. and paste into the editor. Rich content (headings, bold, lists, tables, code blocks) auto-converts to clean Markdown.
• View mode persistence — your Edit/Preview/Split selection is remembered across page reloads (stored in localStorage).
• Mobile toggle restyle — Edit/Preview buttons are now pill-shaped buttons instead of flat tabs.

v2.0 — Multi-User Collaboration

Powerful, privately-hosted Markdown notes — use it the way you like.

mdnest v2.0 transforms the app from a personal note tool into a collaborative workspace for teams, while keeping the default single-user experience unchanged.

Upgrading from v1

If you're running mdnest v1 (single-user), no changes are required. Your existing setup continues to work exactly as before. Multi-user features are opt-in.

To enable multi-user mode:

1. Update your code:

   cd mdnest
   git fetch origin
   git checkout v2.0
   

2. Edit mdnest.conf — add:

   AUTH_MODE=multi
   POSTGRES_PASSWORD=a-secure-password
   

3. Run setup and migrate:

   ./mdnest-server setup      # regenerates docker-compose.yml with Postgres
   ./mdnest-server migrate    # creates database tables
   ./mdnest-server rebuild    # rebuilds and starts everything
   

Your first user (from MDNEST_USER/MDNEST_PASSWORD) becomes the admin automatically.

To also enable live collaboration:

4. Add to mdnest.conf:

   ENABLE_LIVE_COLLAB=true
   

5. Rebuild:

   ./mdnest-server rebuild
   

New Features

Multi-User Mode (E1-E6)

• PostgreSQL-backed user management — optional, only when AUTH_MODE=multi
• Roles — Admin and Collaborator. Admins can invite users and manage access.
• Namespace & directory-level access grants — control who can read or write to which namespaces and subdirectories. write implies read. Grant on / covers the full namespace.
• Permission enforcement — every API endpoint checks access. Collaborators only see namespaces they have grants for.
• Admin panel — manage users (invite, promote/demote, delete) and access grants from the web UI. Accessible via the user avatar menu.
• Frontend permission awareness — read-only mode for view-only grants, write actions hidden when no permission, 403 handled gracefully (no redirect to login).
• Logout button and user identity display in the sidebar.
• /api/config — public endpoint returns auth mode and feature flags so the frontend adapts.
• /api/me — returns current user profile and grants.
• Database auto-migration — tables created automatically on startup. Safe to run on every restart.
• mdnest-server migrate — standalone command for running migrations before starting.

Live Collaboration (E7)

• WebSocket-based presence — see who else has the same note open, with colored avatar dots and usernames.
• Real-time cursor tracking — colored cursor lines show where other users are in the document, with name labels.
• Live content sync — when one user types, others see the changes in real-time (~200ms). When both type simultaneously, each keeps their own content to avoid conflicts.
• Typing indicator — pulsing avatar and "bob is typing..." text in the presence bar.
• ETag conflict detection — GET /api/note returns an ETag, PUT /api/note accepts If-Match. Stale saves return 409 Conflict.
• Conflict banner — when another user saves while you have unsaved changes, a banner appears with a Reload button.
• Auto-reconnect — WebSocket reconnects automatically with exponential backoff on connection drop.
• No external services — everything runs on your server via nhooyr.io/websocket. No Firebase, no Google, no third-party dependencies.

UI Improvements

• Resizable sidebar — drag the right edge to make the project pane wider or narrower (180px–600px).
• SVG file tree icons — replaced emoji icons with crisp SVG icons. Folders with content show blue, empty folders show dashed grey outline with italic name.
• Directory-level share dialog — right-click any folder → "Manage Access" opens a clean dialog to add/remove users with read/write toggles per directory.
• Directory picker for grants — admin panel shows actual folder tree in dropdown instead of free-text path input.
• User-centric grants accordion — admin panel Access Grants tab shows each collaborator as an expandable card with all their directory grants inline.
• Namespace sync button — admin can click the sync icon in sidebar to trigger git pull and refresh the file tree.
• Copy Path — right-click any file or folder to copy its full mdnest path (e.g. growth/docs/readme.md) to clipboard.
• User avatar menu — sidebar footer shows user initials in a circle, click to open dropdown with "Manage Users & Access" and "Sign Out".
• Tree filtering by grants — collaborators only see directories they have access to, not the full namespace tree.
• Mobile improvements — Edit/Preview toggle moved to top, editor fills full screen width, sidebar resize handle hidden on mobile.

Bug Fixes

• Fixed links in preview opening in the same tab instead of a new tab (marked v15 renderer compatibility).
• Fixed WebSocket proxy through nginx (missing upgrade headers).
• Fixed concurrent editing overwriting — remote content only applied when local user is idle.
• Fixed live content sync stopping after first remote update.

Configuration Reference

New settings in mdnest.conf (all optional, defaults preserve v1 behavior):

Docker Changes

When AUTH_MODE=multi, setup.sh automatically adds a postgres service to docker-compose.yml with:

• postgres:16-alpine image
• Health check (pg_isready)
• Persistent volume (mdnest-pgdata)
• Backend depends_on with health condition

API Changes

New endpoints (multi-user mode only):

Changed endpoints:

v1.0 — Self-Hosted Private Knowledge Base

The initial release. A single-user, file-based markdown notes app.

Features

• Markdown editor with live preview, split view, and formatting toolbar
• Mermaid diagrams rendered inline with interactive fullscreen viewer
• Task checkboxes — click to toggle in preview, auto-saves to file
• Image upload — paste or drag images into the editor
• Full-text search with concurrent file reading and cached file index
• Namespace model — mount multiple host directories as separate workspaces
• REST API with JWT and API token authentication
• MCP server for AI agent integration (Claude, Cursor)
• CLI tool (mdnest) for terminal-based note access from any machine
• Git sync — optional auto-commit and push to private repos
• Mobile responsive — works on phone, tablet, desktop
• Docker deployment — multi-stage builds, nginx proxy, alpine runtime
• Private by default — binds to localhost, no cloud, no telemetry
• Tailscale ready — one command for encrypted remote access