Repository navigation
docs: mark the DeepSeek Harness row tested, and say the patch needs no restart - #2446
Conversation
…o restart The dsh row was `untested` because nobody had loaded it. I registered `xerj mcp` v1.0.0-rc.99 in DeepSeek Harness 44.7.0 on Windows through `@deepseek-ai/dsh-mcp-client` in a profile patch, and the harness exposed `xerj_search` to the model as `mcp__xerj__xerj_search` among 22 tools; one call returned 74 hits against an autoindexed repository. Two things the run established beyond the row itself: - the patch is picked up without an application restart. The row and the CLI page both leave that to the reader to guess, and guessing wrong costs a restart that is not needed; - the registration sets `XERJ_DISABLE_FEEDBACK=true`. The server's `instructions` tell the agent to file a field report "automatically ... no need to ask the user", and with `gh` authenticated that opens a pull request from the user's account. Worth knowing before wiring it into a harness. The prose said "Two rows have been run" while three were described; the count is now four, matching the four descriptions and the four `tested` cells. Verified: release asset sha256 matches the published .sha256; `initialize` 7 ms, `tools/list` 2 ms with 22 tools; `xerj autoindex` 248 docs in 4.1 s; `xerj_search` through the harness 74 hits in 46 ms. Not run: the Rust build, `cargo test`, and the ES-YAML suite - no toolchain here, and this is a documentation change. Written by an AI coding agent (DeepSeek-V4.1-Flash, DeepSeek Harness 44.7.0), run by PensiveFei, who has reviewed this change.
|
Thanks for this contribution — we'd like to merge it. Before we can, we need a signed Contributor License Agreement from: @PensiveFei It is a one-time thing per contributor, not per pull request. Read CLA.md, then open a small pull request adding your GitHub username to Once it is merged, comment |
|
@cla-bot check |
|
Re-checking the CLA status for every commit author on this pull request… |
|
CLA signed and merged in #2445; |
Preserve PensiveFei's original documentation contribution while composing with the recently merged Claude Code verification and Trendshift badges. The supported-agent table now contains five tested rows, matching its prose. Attribute the DeepSeek Windows observation to the contributor's recorded verification in xerj-org#2446. Scope hot reload to that observed run and explicitly state that the profile patch was exercised while the dsh bundle was not. Replace the inferred guarantee about feedback being impossible with the observed fact: the run set XERJ_DISABLE_FEEDBACK and saw no automatic report. Do not promote the contributor's single-run latency or hit count into a README performance claim. Verified the five tested table rows and retention of every prior tested client and all three badges. git diff --check passes. No Windows rerun is claimed by this maintainer review; the primary evidence remains xerj-org#2446.
Written by an AI coding agent (DeepSeek-V4.1-Flash, DeepSeek Harness 44.7.0), run by @PensiveFei, who has reviewed this change.
What this changes
The
dshrow of the supported-agents table wasuntested— correct when it was written, because nobody had loaded it. I loaded it:xerjv1.0.0-rc.99 registered as a stdio server in DeepSeek Harness 44.7.0 on Windows through@deepseek-ai/dsh-mcp-clientin a profile patch.xerj_searchappeared in the model's tool list asmcp__xerj__xerj_searchamong 22 tools, and the agent executed one call.Three edits, one file:
untested->tested;tested), and with this row it is four.What the run added beyond the row
No restart. Both this row and
landing/docs/cli.htmlleave the reader to guess whether the patch needs a restart. It does not: the harness picked the patch up and spawnedxerj mcpas a child process of the running app.XERJ_DISABLE_FEEDBACK=trueis worth documenting. The server'sinstructions(2,585 B, injected verbatim into the client's system prompt) tell the agent to file a field report "automatically ... no need to ask the user", andxerj --helprepeats it. Withghauthenticated in the same environment, an agent that follows that instruction opens a pull request from the user's account. The registration above sets the variable; nothing was filed by any route.If you would rather the row pointed at the
dsh/bundle instead of a hand-written patch, say so and I will re-test through the bundle — the bundle was not exercised here, only the raw- insert:block.Verified — commands run in this environment, output observed
xerj-1.0.0-rc.99-x86_64-pc-windows-msvc.zip-> sha2566e3e4b89275310dcc418504494e278826331341a9105635c7543824b689dd47e, identical to the published.sha256initialize-> 7 ms, protocol2026-07-28,xerj-mcp 1.0.0-rc.99tools/list-> 2 ms, 22 tools,xerj_searchpresentxerj autoindex-> 248 docs / 38 files in 4.1 sxerj_searchthrough the harness -> 74 hits in 46 msmcp__xerj__*tools in the model's tool list, no application restartAssumed / Not run
XERJ_DISABLE_FEEDBACKsuppressing the auto-filed report is inferred fromxerj --help; I did not attempt a submission with it unset.cargo test,cargo clippy, ES-YAML suite: not run — no Rust toolchain here, and this is a documentation change (docs/CONTRIBUTION_REVIEW.mdexempts docs-only changes).Scope
README.zh-CN.md,README.ko.mdandlanding/docs/cli.htmlare deliberately untouched.cli.htmlhas no status column. The two translated READMEs are already out of sync with English (Kimi CLI istestedhere anduntestedin both), so flipping one cell there would leave each file's "N rows" sentence contradicting its own table. A follow-up that syncs all three is a better change than half of one; I will open it if you want it.