dsh community contribution workflow: from bug reports to evidence-grade replies

August 18, 2026 · View on GitHub

For anyone who wants to consistently publish high-quality contributions in the official dsh Discussions. Everything below comes from the 2026-08-18~19 campaign (7 evidence-grade replies + 1 triage script + 2 fix-day tools); commands tested on Windows / Node 24 / gh CLI.

1. Triage first: who is unanswered?

The official discussion board gets dozens of posts a day. Run the triage script instead of eyeballing:

node dsh-ecosystem/scripts/discussion-triage.mjs --since 24

It fetches the newest 100 posts with comment authors in ONE GraphQL request and prints # / comments / ours(✅) / title, then lists 0-comment candidates at the end. Notes:

--since 24 (hours), --all (no window), --json (machine-readable)
"ours" is decided by the current gh login (gh api user --jq .login); already-answered posts are flagged automatically
Discipline: reply to each thread at most once; when the OP already has a full root cause, reply with verification + increment only

2. Verify before you reply: diff against current main

The last step can be automated: node dsh-ecosystem/scripts/verify-citation.mjs <repo> <file:line> [--sha <HEAD>] — validates the line exists, prints it, and compares HEAD, so you never cite a line you have not read.

For any post that cites source locations, verify first — three steps:

# 1) Shallow-clone the official repo (once)
git clone --depth 1 https://github.com/deepseek-ai/deepseek-harness.git

# 2) Record HEAD and state the verification baseline in your reply
git log -1 --format="%H %s"

# 3) Locate with rg, confirm line by line
rg -n "listArtifacts|spillAll|SIGINT" packages/session/session-persistence-jsonl/src packages/subprocess/subprocess-local/src apps/cli/src

Real campaign results:

Thread	Verification
#675 SQLite torn-tail	OP's cited lines hold line-by-line on current main → added tornFrom's single consumer + regression-test suggestion
#1047 session.list 500	Holds and unfixed rc.6→rc.7 → added two more unisolated throw sites in the same loop
#3190 spill ENOENT	Holds → pointed out the existing spillDisabled+discardSpill degrade pattern and gave a minimal patch
#3195 CTRL_C_EVENT	Holds and unfixed → added SIGBREAK same-channel risk + the Node-layer CREATE_NEW_PROCESS_GROUP constraint

"Verified, still unfixed" is the most useful signal to maintainers: it says the report is credible and worth prioritizing, with an exact baseline.

3. Empirics: run the assertion locally when you can

Beyond source citations, reproduce what can be reproduced. Example: the CJK retrieval thread (#3206) claimed unicode61 is useless for Chinese and trigram hits; verified with Node 24's built-in node:sqlite:

import { DatabaseSync } from 'node:sqlite'
const db = new DatabaseSync(':memory:')
db.exec("CREATE VIRTUAL TABLE t USING fts5(content, tokenize='unicode61')")
db.exec("CREATE VIRTUAL TABLE t2 USING fts5(content, tokenize='trigram')")
db.prepare('INSERT INTO t(content) VALUES (?)').run('索引优化减少Token消耗的句子')
db.prepare('INSERT INTO t2(content) VALUES (?)').run('索引优化减少Token消耗的句子')
console.log(db.prepare('SELECT count(*) c FROM t WHERE t MATCH ?').get('Token消耗').c)  // 0
console.log(db.prepare('SELECT count(*) c FROM t2 WHERE t2 MATCH ?').get('Token消耗').c) // 1

That also pinned an edge the OP's table did not show: 2-char CJK queries (e.g. "句子") return 0 on both lexical arms (trigram needs >= 3 chars). A comment with measured data beats pure reasoning by an order of magnitude.

4. Reply via GraphQL, not REST

Creating discussion comments over REST 404s for ordinary tokens (GET works, POST does not). Use GraphQL:

# get the discussion node id
gh api repos/deepseek-ai/deepseek-harness/discussions/3174 --jq .node_id

{
  "query": "mutation AddComment($discussionId: ID!, $body: String!) { addDiscussionComment(input: {discussionId: $discussionId, body: $body}) { comment { id url } } }",
  "variables": { "discussionId": "D_kwD...", "body": "Markdown body" }
}

# submit via `gh api graphql --input -` so shell escaping cannot mangle the body

5. Family-ize: merge same-root reports and help maintainers prioritize

Two real families from the campaign:

#3190 (spill-dir ENOENT crash) + #3203 (windows-acl sandbox temp dir cleaned) = "external OS temp cleanup breaks a running dsh". Same root: long-lived per-process private %TEMP% dirs have no protection against external cleanup. Unified fix: recreate/degrade at the boundary where the dir is consumed.
#675 (one bad row cascades into deleting valid rows) + #1047 (one bad file wipes the whole list) = "bad-artifact isolation". Recommended a shared regression-test convention: corrupt one artifact, assert the rest still work.

Linking sibling threads explicitly in replies ("#1047 and #675 are the same family") gives maintainers the full picture in one pass.

6. Anti-spam discipline

Prefer 0-comment threads; do not repeat threads that already have solid replies
When the OP ships a full fix proposal, reply as "verification + increment", not a rewrite
Every reply contains only evidenced claims; mark speculation explicitly as unverified
Reply in the thread's language; include line numbers / HEAD / test output so claims are re-checkable

7. Tooling inventory

Tool	Where	Purpose
discussion-triage.mjs	dsh-ecosystem/scripts	nightly reply queue (one request, flags ours)
verify-2763-fix.mjs	dsh-ecosystem/scripts	auto re-scan + re-score + report on the #2763 fix day
full-registry-scan.mjs	dsh-ecosystem/scripts	full-registry supply-chain scan (325 plugins)
restore-publish.ps1	local work/	re-trigger the listing gate after repos age in (step 7/8)

The official awesome-dsh-plugin Submission gate re-runs "aged in" PRs automatically (regate.yml every 6h), so gates turn green even without manual action; an active re-trigger just makes it green at the earliest moment.