DSH better-sidebar

It is sold as an extension base, and its headline number is off from the measurement. DSH better-sidebar is an npm plugin that attaches to the web mode of an agent CLI called DeepSeek Harness and lays a file explorer, editor, terminal, Git and preview onto one screen. 2,575 stars, 250 commits in the 15 days since the repository was created. What the first line of the README boasts is “only a ~325KB core is downloaded at startup” — and measuring the npm artifacts myself, not one of the five published versions was that value: 0.12.1 is 405KB, and nine days later 0.15.0 is 585KB (+45%). The cause is simple. CI has no size gate. But that is not why this piece stays — the same gate I used to judge someone else’s repository for leaking a number was, on my side, outside version control.

An extension base is not the sidebar — it is the two doors that hand the sidebar to someone else

One package splits into a server half and a browser half. The server half registers a /sidebar/api/* JSON dispatch table (file tree, read, write, search; 12 Git operations; subagent status), media / HTML preview / lazy chunk routes, and a terminal WebSocket. Every request scopes the working directory by sessionId, but for the first paint before a session is attached there is a three-step fallback that descends all the way to the process cwd, so “every feature is isolated per session” is not an accurate description. The browser half mounts two workbenches — right and bottom — as portals, cuts each panel with a recursive split tree, then saves the layout in localStorage under a per-session key and, on restore, runs structural validation, viewport clamping and reissue of duplicate pane ids. The terminal is not an imitation but a real shell, and switching it on injects 8 tools into the model so the agent drives a long-lived PTY.

Item	Measured 2026-08-22
Scale	227 files 2.4MB · tests 82 files 672KB · 21 design documents
Dependencies	27 runtime · 15 of the 18 peers are prereleases
Activity	created 08-07 → pushed 08-21. 250 commits · 17 releases / 11 days
Bus factor	1 — top contributor 217 commits (87%), second 7
PRs	68 merged · 35 closed · 78 open (open PRs exceed every merge ever made)

What I paid for here was nearly reading “22 contributors” as distribution. 19 of them have one or two commits, and most of those are PRs adding their own plugin to a recommendation list. Contributor count had nothing to do with how core development is distributed.

The best thing here is pushing even its own built-ins through the same door as outsiders

Technique	What
Eating its own dogfood	The 7 built-in tabs and 6 viewers pass through the same registerTab/registerFileViewer as outsiders. A comment calls this “eating its own dogfood”
detect → exts, two-stage matching	Viewer choice sniffs content first, extension second. A catch-all match that carries detect is demoted to sniff-only and yields — the structure stops a magic-number viewer from swallowing every file first
dedupeKey	Three hardcoded tab-opening strategies (single, per-path, per-id) absorbed into one function field. It splits “create” from “focus” and fires the callbacks separately
A build purity gate	Importing a package outside the whitelist as a value breaks the build. The message reads “collaborate through cordis services” — one collaboration path, enforced
Degraded mount	A native dependency failing to load does not kill the server. It mounts degraded, the model-facing tools are simply never registered, and a human gets a repair command to paste
Automatic yield on double mount	If the same package comes up under two entry points, duplicate routes fail the whole boot. It switches its own row off not with a runtime singleton but in the loader expression
Side-conversation seeding	Handing a parent log to a child session, it honestly closes the turn that was in flight with a synthetic event and nails down “this is reference context, not the current task.” A test pins that sentence

What I learned is not the technique but where the contract is hung. The purity gate hung the collaboration contract on the build, not on a human comment. As we will see, the same contract on my side was hung on a comment — and it did split once.

Three headline claims all failed, and one defect is not even in the issue tracker

• “~325KB core” — broken. Measuring five tarballs myself: 405 → 424 → 459 → 482 → 585KB. It is 12,283 lines and not even minified, so “that’s the minified figure” is not available as an excuse. A chunk-loader comment in the same repository says “only the ~1MB core bundle”, producing a third number. Meanwhile the “7MB mermaid chunk” measured 6,996,658 B — exactly right. The author is not incapable of measuring; only the numbers with no gate behind them went un-updated.
• “built-in and third-party are fully equal” — broken. The registration path is the same, true, but an external plugin cannot expand a panel. The service exposes no expand or toggle method, so an external tab opens and then hides behind a collapsed panel, and an open issue is alive saying the only remaining means is DOM hacking.
• “the same trust fence as /api” — same function, different scope. Upstream applies that check only to privileged methods; this plugin applies it to every route. So on a 127.0.0.1 connection where Chrome sends an Origin with no port, every sidebar route is 403 while the host-side conversation runs fine. That is over-blocking, not under-blocking.
• HTML preview — the URL is right, the headers kill the assets. Encoding everything into the path rather than the query so relative references do not lose session scope is an excellent design, but the media type table has neither .css nor .js, so everything ships as octet-stream and the same route attaches nosniff. ⚠ The tests create an illusion — they assert URL decoding only and never look at content-type.
• A defect not even in the issues — file writes have no optimistic concurrency check. The save API takes a path and contents, nothing else: no mtime, no hash. On top of that sits the “no file watcher” the repository itself admits. Leave a tab open, have the agent rewrite that file, and the moment the human saves, the agent’s latest contents vanish without a warning — a classic lost update, in exactly the scenario this product exists for.
• “3 platforms supported” — both CI jobs are ubuntu only. And the most painful of the unresolved P0s happen to cluster on Windows. Installing it makes resuming an existing session fail every time, removing it restores things immediately, and the maintainer cannot reproduce it on mac.

What held, for the record. “A real shell” is true, and the outside criticism that “the README still advertises a built-in Office preview” has already been fixed and no longer holds — outside criticism goes stale too. And what I paid for in this section was reading comment density as maturity. This repository has the rare comment culture of writing down why it was not done the other way, and on the first pass I stood that up as evidence of maturity — but a comment pointing at a tab name that does not exist, and a notice that never caught up with an upstream relicensing, are both alive at HEAD. Unverified high-density comments are worth the same as wrong documentation.

The adoption call ended in three lines of measurement, not a review

where dsh returns nothing. No ~/.dsh. Of the 9 global npm packages, zero from that ecosystem. 15 of the 18 peers are plugins from that ecosystem, so the install path itself does not exist. It is not “this does not fit right now” but “there is no host.” That leaves porting techniques — and of four candidates, only one actually survived.

• Bundle size gate → zero gain. The target’s worst defect happens to sit exactly where I had already filled the hole. All 9 limits in scripts/check-weight.mjs judge against dist, each limit carries its derivation in a comment, and scripts/deploy.mjs is four-stage fail-closed: links and anchors → weight → glossary → markers.
• dedupeKey → zero gain. The TC pipeline already does the same three-way open policy off a single ownership marker — create new / idempotent rewrite if it is ours / yield with a suffix if it is someone else’s.
• Degraded mount → rejected. The local model MCP server I meant to attach it to was not registered in any configuration at all. Server name, tool name and slash command diverge three ways, so reviving it would not attach anyway. A takeaway with no place to apply is not a takeaway.
• Automatic yield on double mount → the only takeaway. I opened Loki to see whether a counterpart existed and found zero lockfiles, pidfiles or singleton guards. But ⚠ the spot I pointed at was wrong — the watchdog path that runs every five minutes was already covered by ensure(), and what was not covered is that the login startup entry goes straight to the module rather than through gateway ensure. The minimal fix is one line, not a new preemption protocol.

⚠ And the verification command I wrote down was one I had never actually run. I wrote “check that the process count is 1”, but because of the venv launcher stub a healthy single worker always looks like two — the parent holds 1 thread and 0 sockets, and the child holds the real connection. Implement self-yield that way and a normal worker kills itself on every boot. A fabricated verification procedure is more dangerous than a wrong number — the wrong number you can re-measure, the fabricated procedure leaves behind the belief that you passed.

The yardstick I judged others with was outside version control

git ls-files scripts/lib/ returns an empty result. All five judgement gates — check-links.mjs, check-weight.mjs, check-glossary.mjs, collect-ai-issue.mjs, lib/markers.mjs — are untracked, and git check-ignore answers that they are not ignored: they were not excluded on purpose, they have never once been committed. The header of markers.mjs declares itself the canonical source and even records a recurrence history, yet git log on that file returns nothing. A file that declares itself canonical has no history.

The tracked deploy.mjs hard-depends on those five, so on a fresh clone the very first import dies instantly. Nothing leaks — it is fail-closed, so the deploy simply stops. What does happen is that the grounds on which I killed half the target’s takeaways with “I already have something better” do not reproduce in the repository. The target’s disease was an absent gate; mine is a gate with no provenance, and the second is more expensive. Everyone knows an absent gate is absent; a gate that exists outside the commit makes you believe it is there.

Verdict — and what I did not do

Item	Call	Why · reopen when
Adopt the product	⛔ Rejected	There is no host. Reopen = when I actually start using this harness. Even then, look at the three unresolved P0s first
Size gate · dedupeKey	Zero gain	Already present in a better form. ⛔ Not reopening on this axis
Degraded mount	Rejected	The place to apply it was an unregistered relic. Reopen = when I actually wire that server
Self-yield at boot	Adopted	One line in the startup entry plus atomic preemption for the restart path. Verify by heartbeat record, not process count
Porting the boundary sentence	Held	Per-stage adversarial review already does that job. The real gain is on the “run the verification command once before the call and paste the output” side
Committing the five gates	First priority	Cost near zero. Verify = whether the deploy chain actually passes in a throwaway clone

What I did not do. ① I never once ran this plugin — there is no host so it will not install, which means every call is measured from source, npm artifacts and the issue tracker, and I have never seen the screen. The bundle size is certain, but “what actually comes down at startup” I confirmed in code alone. ② Failure reports from an external install rig I used as corroborating evidence only — it measures by headless boot and this plugin is web-profile only, so the failures may be rig artifacts. ③ The 26 ecosystem plugins I only counted. The listing condition is a single self-declared tag, so that number is not evidence of activity. ④ The one adopted item is still unfixed. This piece goes as far as the call.

The through-line of this piece is this — before doubting someone else’s numbers, check that the tool doing the doubting is loaded in my own repository. If the yardstick used to judge the target does not reproduce on my side, that judgement is not a verdict but an impression.