Blog · Pulse

Community Pulse #4

September 7, 2026 · dshbase · a recurring report on the DeepSeek Harness community

This is the fourth edition of Community Pulse, our scan of the official deepseek-ai/deepseek-harness discussion board via the GitHub API. The headline since #2 and #3: the entry-gate hang that dominated August largely gave way to a new failure mode — breaking on upgrade. A fast-moving release train and a board that keeps doubling are both good problems to have, but the session-loss threads say the transition is costing real trust.

The snapshot

The board now holds 5,757 discussions — up from 4,404 when we wrote #2 on August 25, about +31% in under two weeks. The repo sits at ~214.5k stars / 25.3k forks. In our 13-day window (Aug 26 → Sep 7) we counted 158 new threads: General 75 · Ideas 42 · Show Your Plugins 21 · Q&A 20 — and 29 of them carry an explicit bug report in the title. The team's pace matches the traffic: five tagged releases in five days (0.1.2-alpha.3 on Aug 31 through dsh-v0.1.3-alpha.1 on Sep 4, with 0.1.2-rc.1 the current npm latest).

The headline: the upgrade is the new install

August's pain was getting in the door. September's is getting stepped on while you're inside. A small cluster of threads, all within two days of each other, describe the same user-visible failure around 0.1.2-rc.1:

#5860 — the cleanest report: on macOS (Apple Silicon, node 24), npm install -g @deepseek-ai/dsh@latest resolves to 0.1.2-rc.1, after which the Web UI can no longer select workspaces or list sessions. Rolling back to 0.1.1-rc.2 fixes it instantly.
#5818 — session history fails to load after upgrading: the v0 migration inventory refuses release-tagged data, so old sessions go dark.
#5843 — cold-restart regression in rc.1 itself: seeded/forked sessions lose their entire list projection because cachedSnapshot validates identity with inheritedEventCount = 0 and discards a valid checkpoint. #5857 is its sibling: forked sessions fall back to the workspace name after restart.

Put together, these form a clear theme: session persistence is now the ecosystem's risk surface — the same family we first flagged in #1, now amplified by release velocity. For a tool whose core promise is long-running agent state, every regression here reads as data loss to users even when nothing is actually gone. A forward-migration dry-run against archived sessions (and a rollback story in the release notes) would defuse most of this board noise.

Bug radar: tool-call binding, plugin collateral, and local-model streams

Zero-argument tool calls fail at binding. #5851 (6 comments, reproduced on rc.1 Windows): any schema-less tool — built-in get_goal, job_list, or an MCP list_pages registered without params — dies with ToolCallError: binding arguments must be lossless JSON. Anything that touches the argument-binding layer when given is quietly breaking a whole class of plugins.
Plugins taking the host down. The isolation debate (#2603, #2687 from previous pulses) is producing live casualties: #5846 reports mounting dsh-hooks-claude-code paralyzes every tool call; #5844 shows local file plugins making official DeepSeek requests fail with REQUEST_EX…; #5864 is a plugin-tree boot crash-loop when a plugin imports an export the bundled packages don't re-export (same root as #5854's missing assertNever re-export). This is exactly the failure class our daily compatibility CI is built to catch before you do.
Local models and provider edges. #5729 (6 comments): llama.cpp streams keep dropping mid-call. #5773 is nastier in principle: sessions that ran on google-antigravity models become permanently unusable after switching to Claude — provider state is leaking into session history.

Feature radar: temporary chat, and a security policy argument

The most-commented idea of the window is #5765 (23 comments): a Temporary Chat that starts without binding to a project — quick research, one-off questions, archive or promote later. It's the "lower the cost of a first conversation" ask, and the comment count says it's the single most-wanted quality-of-life feature right now.

The sharpest debate is #5835: why block binding to 0.0.0.0 outright instead of warning? The requester wants LAN access from their own devices and experiences the hard ban as treating adults like the lowest-common-denominator user. This is the security-by-default vs. self-host-friendliness tension every agent runtime eventually hits — DSH picked safety, and is now paying the popularity price in threads. Related: #5849 wants the Settings → Models page to stop hiding adapter-registered providers, and #5853 asks for a local auth path for embedded/desktop integrations.

Ecosystem: the count war reversed itself

Funny thing about the "7,000-plugin count war" we wrote up last month: dshbase now catalogs 7,797 entries, every one checked for a bundle manifest, repository liveness, and (for npm ones) daily install verification. The indexes are converging on each other's numbers — which means, as predicted, counts stopped differentiating and verification is the last axis left. If a directory can't tell you whether a plugin survives dsh plugin add today, it's a search engine with extra steps. That's the line we hold on the directory and our audit page.

The discovery layer itself became a plugin genre this window: #5858 (DSH Plugin Manager — develop, deploy, manage), #5865 (community toolbox refreshed for rc.1), and our personal favorite of the fortnight, #5852 dsh-llm-trace-plugin: patches globalThis.fetch to show the literal HTTP bytes between Harness and your LLM provider. When half the board is debugging "why does the model behave weirdly," wire-level observability is the answer.

The adoption signal of the month is #5842: a team that built — and then deleted — their own agent runtime (planning, tool execution, skills, sub-agents, session recovery, sandboxing) to move their enterprise platform, MOVO, onto DSH. Whatever the bugs above cost in goodwill, that post is the argument that the harness-as-commodity thesis is working. 21 "Show Your Plugins" threads in 13 days, meanwhile, is the boring heartbeat of a healthy ecosystem.

What to watch next

Does rc.2 fix session persistence, or work around it? #5860/#5818/#5843 are the trio to track; enterprise migrations like MOVO's start assuming upgrade-safety, and the harness needs to earn that.
Plugin-failure isolation (#5846/#5844/#5864 keeps growing the evidence pile — the proposal threads from earlier pulses now have receipts).
Temporary chat (#5765): cheapest big-win on the board; if it lands in 0.1.3, expect a visible bump in first-time-retention chatter.
The 0.0.0.0 policy (#5835): watch whether the team bends toward opt-in LAN mode with an auth prompt. The answer shapes who self-hosts DSH vs. reaches for a wrapper.

Community Pulse is a recurring report. Have a thread we should track, or a correction? Open an issue on the dshbase repo.

All articles →