Shipped fix sample

Three Hermes WebUI ships in one day

Public proof-work record from 2026-05-03: three Dutch AI Agents fixes shipped by maintainer @nesquena in three same-day Hermes WebUI release tags, all credited to @dutchaiagency.

The three same-day ships

v0.50.281 — PR #1536 (codex): allow LM Studio LAN IP endpoints to be classified as custom with regression coverage for the adjacency rule. Maintainer rebased and shipped via #1541; release notes credit (#1536 by @dutchaiagency).
v0.50.284 — PR #1557 (codex): lock-and-re-read stale-stream cleanup for issue #1533 plus deterministic two-thread regression test. See the worked example below.
v0.50.286 — PR #1561 (claude): three-layer lock for the silent password env-var no-op in #1560 (GET surface, POST 409, frontend disable) with 23 regression tests. Maintainer note: “clean, well-scoped UX fix for a longstanding silent-no-op … the three-layer approach is exactly right.”

Worked example: stale-stream race (v0.50.284)

Hermes WebUI had a stale cleanup path that could mutate stream state outside the existing per-session lock. The narrow fix was to take the session lock, re-read active_stream_id inside it, and bail when a concurrent chat start had already registered a new stream.

Scope

Original signal: Hermes WebUI issue #1533 described stale stream cleanup mutating outside the lock.
Patch shape: protect the legitimate mutation path with the existing per-session session lock.
Race guard: re-read active_stream_id under the lock before clearing state.
Test: deterministic two-thread regression coverage for the old clobber window.
Boundary: no broad streaming rewrite and no product-behavior expansion.

Public artifacts (v0.50.284)

Pull request: nesquena/hermes-webui #1557
Maintainer credit: shipped in v0.50.284 and credited to @dutchaiagency
Release tag: Hermes WebUI v0.50.284
Issue closed by release PR: stale-stream race #1533

Verification

The PR included focused regression coverage for the race plus adjacent stale-stream recovery cases:

python -m pytest tests\test_stale_stream_cleanup.py -q
python -m pytest tests\test_stale_stream_cleanup.py tests\test_stale_stream_pending_recovery.py tests\test_stale_empty_session_restore.py -q

Result logged in the PR: 6 focused tests passed, then 10 combined stale-stream tests passed.

Why it matters

This is the kind of proof a buyer can inspect: the failure mode, the lock boundary, the regression test, and the maintainer's shipping signal are all public. The PRs themselves often show as closed rather than merged because the maintainer rebases and lands fixes through release PRs, but the release notes and shipping comments name @dutchaiagency directly. Three same-day ships from one maintainer is the credibility signal we trade on; the worked stale-stream case is one inspectable example of the shape.

Need this shape of work on your public repo? Email a brief or open a public task issue .