Issue #1458 reports persistent-host crashes (≥1/day) when running the WebUI
under launchd KeepAlive on macOS. Root cause: `bootstrap.py` calls
`subprocess.Popen([python, "server.py"], start_new_session=True)`, probes
/health, then exits 0. Under any process supervisor (launchd, systemd,
supervisord, runit, s6), the supervisor sees its tracked PID exit, marks
the program as "completed," and respawns it. The new bootstrap fails to
bind port 8787 (orphaned server still has it), exits non-zero, supervisor
respawns again — loop until the orphan crashes for some other reason and
the next respawn finds the port free.
This PR addresses Bug #1 of the three failure modes tracked in #1458:
the `bootstrap.py` double-fork breaking process supervisors. Bug #2
(state.db FD leak) and Bug #3 (HTTP-unhealthy wedge) remain open under
the same issue — they need diagnosis data before a fix can land.
Changes
-------
1. `bootstrap.py`:
- New `--foreground` argparse flag with help text mentioning launchd /
systemd / supervisord.
- New `_detect_supervisor()` that returns the env var name for any
supervisor it detects: `INVOCATION_ID` / `JOURNAL_STREAM` /
`NOTIFY_SOCKET` (systemd, s6), `XPC_SERVICE_NAME` (launchd),
`SUPERVISOR_ENABLED` (supervisord), or `HERMES_WEBUI_FOREGROUND` for
the explicit user opt-in. Truthy values for the explicit opt-in:
`1` / `true` / `yes` / `on` (case-insensitive).
- `main()` branches on `args.foreground or _detect_supervisor()`:
- **Foreground path:** chdir to `agent_dir or REPO_ROOT`, then
`os.execv(python, [python, server_path])` to replace the bootstrap
process image with the server. The supervisor sees the long-lived
server as the original child. No `wait_for_health` probe — the
supervisor's KeepAlive / Restart=on-failure handles liveness.
- **Default path:** unchanged. Spawn server as detached child via
`Popen + start_new_session=True`, probe /health, return 0. This
still works for interactive `bash start.sh` invocations.
- Resolved env vars (HOST/PORT/STATE_DIR/AGENT_DIR) are now mutated on
`os.environ` directly instead of into a local `env` copy so they
are inherited across `os.execv`.
2. `docs/supervisor.md` (new): runnable launchd plist, systemd .service,
and supervisord conf examples + a diagnostic recipe (`lsof` + ppid
chain) for catching the orphan-loop in production.
3. `.gitignore`: allowlist `docs/supervisor.md` (the directory uses an
opt-in pattern; matches the existing `!docs/docker.md` precedent).
4. `tests/test_bootstrap_foreground.py` (new): 35 regression tests
covering the argparse flag, `_detect_supervisor()` behavior across all
five supervisor env vars, the explicit opt-in's truthy/falsy values,
and `main()`'s execv-vs-Popen routing decision under each input
combination. `os.execv` is monkeypatched in the routing tests — we
pin the structural choice (which call is made, with which args, in
which cwd, with which env) not the post-exec behavior.
Why this scope and no more
--------------------------
Bug #2 (state.db FD leak) lists 5 candidate paths and asks the reporter
for `lsof -p <pid> | sort | uniq -c | sort -rn | head -20` output to
disambiguate. Until that data lands, any "fix" would be speculative —
explicitly out of scope per the contributor-pickup comment on the issue.
Bug #3 (launchd-running, port-listening, HTTP-unhealthy) was added in
@stefanpieter's reply comment. Diagnosis is in flight; no concrete fix
shape yet. Also out of scope.
Running locally end-to-end verifies the behavior:
```
[bootstrap] Starting Hermes Web UI on http://127.0.0.1:8789 (foreground mode: --foreground)
$ pgrep -af 'server.py'
2997632 /home/.../python /tmp/wt-fix-1458/server.py
$ ps -o ppid -p 2997632
2997581 ← bash that ran bootstrap.py — same PID as the original bootstrap
$ ps -p 2997581 -o cmd
... bootstrap.py ... ← but exec'd into server.py
```
The same PID that bash forked for `bootstrap.py` is now `server.py`.
A supervisor watching that PID would correctly observe the long-lived
server. No double-fork.
Verification
------------
- 3811 tests pass (`pytest tests/` — full suite, +51 from this PR plus
master-merge-in)
- All 35 new bootstrap-foreground tests pass
- `bash scripts/run-browser-tests.sh` PASS (HTTP API checks against worktree)
- `bash scripts/webui_qa_agent.sh 8789` PASS (23/23 visual QA)
- Live verified: server starts cleanly under both `--foreground` and
`HERMES_WEBUI_FOREGROUND=1`; PID lineage confirms no double-fork
Closes#1458 (Bug #1 only). Bugs #2 and #3 remain tracked under the
issue.
The JavaScript _normalizeConfiguredModelKey function had the same bug as the
Python _norm_model_id function that was fixed in commit d6164cd. It used
substring(indexOf(':')+1) which only removes the first colon-separated segment,
leaving provider names in the normalized model ID.
For example, '@custom:jingdong:GLM-5' became 'jingdong:glm.5' instead of 'glm.5'.
This caused duplicate Primary badges to appear in the model dropdown when using
custom providers with @provider:model ID format.
Changes:
- Replace substring(indexOf(':')+1) with split(':').pop() to strip all colon prefixes
- Add provider name to badge label for clarity (e.g., 'Primary (jingdong)')
When the webui auth session expires (e.g., after a server restart),
api() returns undefined after redirecting to /login. Previously,
loadSession() and _ensureMessagesLoaded() would dereference the
undefined response and throw, surfacing a confusing 'Failed to load
session' toast while the browser was already navigating away.
Add guards after api() calls that may trigger 401 redirects:
- loadSession(): bail early if data is undefined
- _ensureMessagesLoaded(): return silently if data is missing
- _loadOlderMessages(): return silently if data is missing
This prevents the stuck loading state and unnecessary error toasts
when the user is already being redirected to re-authenticate.
Fixes#1391 (reported as 'Failed to load session' after restart)
The _norm_model_id function was using split(':', 1)[1] which only removed
the first colon-separated segment, leaving provider names in the normalized
model ID. For example, '@custom:jingdong:GLM-5' became 'jingdong:glm.5'
instead of 'glm.5'.
This caused the default model injection check to fail, resulting in a
duplicate 'Default' group being added to the model list even when the
model already existed with a provider prefix.
Changes:
- Use split(':')[-1] to get the last segment after all colons
- Use split('/')[-1] consistently for slash-separated paths
- Replace local _norm lambda with _norm_model_id function call
Fixes duplicate Default group appearing in model dropdown when using
custom providers with @provider:model ID format.
Closes#1442 (server-side _LOGIN_LOCALE missing ja/pt/ko)
Closes#1443 (promote _isImeEnter helper to 6 other Safari Enter guards)
Closes#1446 (glued-bold-heading lift for LLM thinking-block output)
Closes#1447 (markdown heading visual hierarchy in chat messages)
All four issues were filed by the Opus pre-release advisor on the v0.50.264 batch
or by Cygnus via Discord (relayed by @AvidFuturist, May 1 2026). They share a
common shape — narrow, well-scoped, independent of each other, all adding
regression tests.
== #1442: _LOGIN_LOCALE parity (api/routes.py + static/i18n.js) ==
Added entries for ja/pt/ko to the server-side _LOGIN_LOCALE dict that renders
the localized login page BEFORE the JS i18n bundle loads. With v0.50.264
shipping Japanese as the 8th built-in locale, ja/pt/ko users were seeing the
English login page even with their language preference set.
While auditing static/i18n.js for English leakage, also fixed:
- ko: 10 user-facing login/sign-out/password keys still in English
- es: 3 sign-out/auth-disabled keys still in English
Tests: tests/test_login_locale_parity.py (20 tests) — pins both invariants:
(a) every locale in i18n.js LOCALES has a matching _LOGIN_LOCALE entry
(b) every locale's login-flow keys (13 of them) are translated, not English
== #1443: window._isImeEnter promotion ==
PR #1441 fixed the Safari IME-composition Enter race in the chat composer
(`#msg`) by widening the guard from `e.isComposing` to a `_isImeEnter(e)`
helper that combines three signals (isComposing || keyCode===229 ||
_imeComposing flag). Six other Enter-input handlers were left on the original
narrow guard and would still drop IME composition Enters on Safari for
Japanese/Chinese/Korean users.
Promoted the helper to `window._isImeEnter` (defined in static/boot.js) and
replaced the `e.isComposing` guards at all six sites:
- static/sessions.js: session rename, project create, project rename
- static/ui.js: app dialog (confirm/prompt), message edit, workspace rename
The state-free part of the helper (`isComposing || keyCode===229`) handles
Safari's race for any focused input without needing per-input composition
listeners — only `#msg` keeps the local `_imeComposing` flag.
Tests:
- tests/test_issue1443_ime_helper_promotion.py (9 tests) — pins each site
+ verifies no raw `e.isComposing` Enter-guards remain in sessions.js/ui.js
- tests/test_ime_composition.py — alternation regex extended to accept
the windowed helper form (loosen-test-on-shape-change pattern from
v0.50.264 reflection notes)
== #1446: glued-bold-heading lift (static/ui.js renderMd + Python mirror) ==
LLMs in thinking/reasoning mode emit "section headers" glued to the end of the
previous paragraph with no whitespace:
Para 1 text.**Heading to Para 2**
Para 2 text.**Heading to Para 3**
The renderer correctly produces inline `<strong>` per CommonMark, but it looks
like trailing emphasis on the body text rather than a section break. Cygnus
reported this as "Markdown feedback 2 of 3."
Added a single regex pre-pass in renderMd():
s.replace(/([.!?])\*\*([^*\n]{1,80})\*\*\n\n/g, '$1\n\n**$2**\n\n')
Constraints chosen to avoid false positives:
- Trigger only on `[.!?]` IMMEDIATELY before `**` (no space) — almost always
an LLM-glued heading, not intentional emphasis
- Inner text ≤80 chars, no `*` or newline (single-line only)
- Trailing `\n\n` required — preserves "this is **important** to know."
mid-paragraph emphasis untouched
- Position: after rawPreStash restore, before fence_stash restore — fenced
code blocks stay protected (their content is `\x00P` / `\x00F` tokens
when the lift runs)
Mirrored in tests/test_sprint16.py render_md() so both stay in sync.
Tests: tests/test_issue1446_glued_heading_lift.py (17 tests, 5 of which drive
the actual ui.js renderMd via node) — covers all 3 trigger forms (.!?), all 4
preserve-emphasis cases the issue spec'd, fenced/inline code protection,
chained glued headings, source-level position pin, regex shape pin.
== #1447: markdown heading visual hierarchy (static/style.css) ==
Pre-fix sizes in `.msg-body`:
h1 18px, h2 16px, h3 14px (= body), h4 13px, h5 12px, h6 11px
So h3 was indistinguishable from body and h4/h5/h6 were SMALLER than body.
Cygnus's report: "Markdown feedback 3 of 3 — Headings seem to be missing
across the board in Hermes. They're there, but all plaintext."
New sizes:
h1 24px (border-bottom) h2 20px (border-bottom) h3 17px h4 15px
h5 14px (uppercase, tracked) h6 13px (uppercase, tracked, muted)
All headings now `font-weight:700` + `color:var(--strong)` for stronger ink.
h5/h6 use uppercase + letter-spacing for "label-style" affordance instead
of being smaller-than-body.
Synced .preview-md (file preview pane) to match exactly so a markdown file
preview and a chat message render identically. Added missing h4/h5/h6 rules
to .preview-md (it only had h1-h3 before).
Updated data-font-size="small"/"large" h1-h6 overrides to scale
proportionally with the new defaults. Hierarchy preserved at all three
font-size settings.
Tests: tests/test_issue1447_heading_hierarchy.py (9 tests) — pins the size
hierarchy, the bottom borders on h1/h2, the uppercase affordance on h5/h6,
the .preview-md sync, and the small/large override scaling.
== Verification ==
pytest tests/ -q → 3748 passed (+56 new)
bash ~/WebUI/scripts/run-browser-tests.sh → 20 + 11 PASS
bash ~/WebUI/scripts/webui_qa_agent.sh 8789 → 23/23 PASS
Visual confirmation in browser at port 8789:
- Heading hierarchy clearly visible at all 6 levels
- Glued-bold lift produces separate paragraphs as designed
- window._isImeEnter accessible from any module after boot.js
- Login page renders ja/pt/ko strings correctly (curl -s /login)
REQUIRED:
- _fully_unquote_path range(3) -> range(10) — defense-in-depth so quadruple-
encoded .. is rejected by validator instead of slipping through (not
exploitable but contract violation)
- docs/EXTENSIONS.md trust-model callout moved to top of file with explicit
'don't enable in untrusted env / don't point at user-writable dir' guidance
NICE-TO-HAVE (taken since Nathan asked for all fixes big and small):
- URL list cap at _MAX_URL_LIST=32 to avoid pathological rendering
- One-shot WARNING log for rejected URLs (silent drop now visible to admin)
- One-shot WARNING log for URL list truncation
- MIME map: ttf (font/ttf), otf (font/otf), wasm (application/wasm)
5 regression tests in tests/test_pr1445_opus_followups.py pin all invariants.
Opus advisor caught a recoverable footgun in PR #1441's manual flag: if
focus is lost mid-composition (window blur or older Safari WebKit IME
quirk), compositionend may never fire and _imeComposing stays true
until the next full composition cycle. Result: Enter-to-send is
silently broken until page reload — an unrecoverable stuck state for
something that's supposed to be transient.
Add a blur listener that also resets the flag. Cheap belt-and-suspenders
against the stuck state. Adds 1 regression test pinning the listener.
(other Opus findings logged in /tmp/stage-264-brief.md as follow-up
issues: _LOGIN_LOCALE parity for ja/pt/ko, promote _isImeEnter to the
6 other Safari-affected Enter guards in sessions.js + ui.js)
The markdown fence regex /```([\s\S]*?)```/g had no line anchoring. A literal
triple backtick inside code block content (e.g. a regex with ``` in a lookbehind,
or a script that documents fences) terminated the outer fence at the wrong place.
The leaked tail then went through bold/italic/inline-code passes, eating `*`
characters as italic markers and emitting literal </strong> tags into the
rendered output.
CommonMark §4.5 requires that an opening code fence be the first non-whitespace
content of a line (up to 3 spaces of indent allowed) and that the closing fence
also start a line. This patch updates 3 sites + the Python mirror to use that
invariant:
static/ui.js:1559 renderMd() fenced-block stash (assistant messages)
static/ui.js:66 _renderUserFencedBlocks() (user messages)
static/ui.js:2599 _stripForTTS() (TTS speech pre-strip)
tests/test_sprint16.py Python mirror
Pattern: (^|\n)[ ]{0,3}```(?:([\s\S]*?)\n)?[ ]{0,3}```(?=\n|$)
The non-capturing (?:...\n)? group keeps empty fences (```\n```) working;
without it, a body+\n is required and the closing fence on the very next line
no longer matches. The lead group (^|\n) is prefixed back to the stash token
so paragraphs above don't bleed into the <pre> block.
20 regression tests in tests/test_issue1438_fence_anchoring.py cover:
- Cygnus's exact repro from Discord (May 1 2026)
- Inline ``` mid-paragraph (must not open fence)
- Partial/streaming fence with no close (must not eat content)
- Empty fences with and without language tag
- 3-space indented fences (allowed) vs 4-space (not a fence)
- Multiple adjacent blocks
- Bold/italic/inline-code surviving after a fence
- Source-level guards on all 3 patched sites + lead-prefix invariant
Empirical browser verification (live JS, on bug repro):
Before fix: </code></pre>[^\n]<em>|%%[ \t]</em>... ← truncated, italic leak
After fix: <pre><code>...```[^\n]*|%%...</code></pre> ← intact, regex preserved
Tests: 3678 passed (+20 from new test file, was 3658), 0 failures.
Reported-By: Cygnus (Discord)
Relayed-By: @AvidFuturist
Closes#1438
Adds a ja locale entry (828 keys) under static/i18n.js LOCALES,
inserted between en and ru. All existing keys translated to natural
concise Japanese suitable for UI labels, with placeholders ({0}, etc.)
and template literals preserved verbatim.
- _lang: 'ja', _label: '日本語', _speech: 'ja-JP'
- 828 keys (matches en, including the documented duplicate keys
whose JS last-wins semantics are preserved)
- syntax verified with `node -c static/i18n.js`
Tested live on a self-hosted instance; Settings → Language → 日本語
selects the new locale and switches the UI text.
East Asian IMEs (Japanese/Chinese/Korean) use Enter to commit composition.
The existing isComposing guard misses Safari, where the committing keydown
fires after compositionend with isComposing=false. Also track composition
manually and check keyCode===229 for broader coverage.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Fix two-layer bug where `/api/session` returned `context_length=0` for
sessions that pre-date #1318, then the frontend silently fell back to
cumulative `input_tokens` and the 128K JS default, producing nonsense
indicators like "100" capped from "890% used (context exceeded), 1.2M
/ 131.1k tokens used".
Empirical impact: 23 of 75 sessions on dev server rendered >100% before
this fix. #1356 fixed the same symptom on the live SSE path but missed
the GET /api/session load path that older sessions go through.
Two-layer fix:
1. Backend (api/routes.py:1295-1313) — resolve context_length via
agent.model_metadata.get_model_context_length() when the persisted
value is 0. Mirrors api/streaming.py:2333-2342.
2. Frontend (static/ui.js:1269) — drop the cumulative `input_tokens`
fallback. When last_prompt_tokens is missing, render "·" + "tokens
used" (existing !hasPromptTok branch) instead of computing a
percentage from the cumulative total.
10 regression tests in tests/test_issue1436_context_indicator_load_path.py
covering both layers + the empty-model edge case (avoids the 256K
default-for-unknown-model trap that get_model_context_length('') returns).
Verified live: claude-opus-4-7 session with input_tokens=5,226,479 now
renders "·" + "5.3M tokens used" instead of "100" + "3987% used".
Reported by @AvidFuturist.
Closes#1436.
Two unrelated UX bugs, both small surgical fixes with regression tests.
Issue #1432 — "+" button doesn't open new chat during streaming
================================================================
Reported by @Olyno: clicking "+" after sending a first message keeps
redirecting to the same chat instead of opening a new blank conversation,
making parallel chats impossible until the first response finishes.
Root cause:
static/boot.js:691 (and the Cmd/Ctrl+K branch at :844) had an empty-session
guard from #1171 that skipped newSession() when message_count===0:
if(S.session && (S.session.message_count||0)===0){
$('msg').focus(); closeMobileSidebar(); return;
}
But during the first user turn of a brand-new session, message_count is
still 0 server-side because the user message hasn't been merged into
s.messages yet. The guard treated that as "empty" and silently dropped
the click, blocking parallel chats for the entire stream duration.
Fix:
Tighten the predicate to also exclude in-flight state:
if(S.session
&& (S.session.message_count||0)===0
&& !S.busy
&& !S.session.active_stream_id
&& !S.session.pending_user_message){
$('msg').focus(); closeMobileSidebar(); return;
}
Same predicate applied to the Cmd/Ctrl+K handler at :844. The in-flight
signal (active_stream_id || pending_user_message) is the same one
_restoreSettledSession() in messages.js:1081 already uses to decide
whether a session is "settled" — keeping both call sites aligned.
Verified end-to-end: with S.busy=true and pending_user_message set, the
old guard returned `block=true` (= the bug), the new guard returns
`block=false` (= fixed). With a truly empty session (no busy, no pending),
both old and new guards still block — preserving #1171 behavior.
Issue #1423 — Profile name field auto-capitalizes typed values
==============================================================
Self-reported (Mac app, May 1 2026): typing `hello` into the New Profile
"Name" field shows `Hello` after blur/autofill, contradicting the
"Lowercase letters, numbers, hyphens, underscores only" hint right next
to it. The form lowercases on submit so stored data is correct, but the
displayed value during typing is misleading.
Root cause:
static/panels.js:2532 had only autocomplete="off":
<input type="text" id="profileFormName"
placeholder="..." autocomplete="off" required>
Missing three attributes that actually prevent the misbehavior:
- autocapitalize="none" — mobile keyboards (iOS Safari, Android Chrome,
WKWebView in the Mac app) auto-capitalize the first letter without it
- autocorrect="off" — Safari runs autocorrect on blur, can rewrite hello→Hello
- spellcheck="false" — desktop browsers may run spellcheck on blur
Fix:
Add the three attributes to profileFormName. Also added to
profileFormBaseUrl since URLs are similarly bad targets for
autocapitalize/autocorrect. profileFormApiKey is type="password" and
already has correct browser behavior.
Verified end-to-end against the live DOM: openProfileCreate() →
getElementById('profileFormName').getAttribute(...) returns the new
attributes correctly, with required preserved.
Tests
-----
3648 passed, 2 skipped, 3 xpassed (was 3640 — added 8 new regression tests
in test_1432_newchat_and_1423_profile_input.py).
One pre-existing test had to be widened: tests/test_mobile_layout.py
test_new_conversation_closes_mobile_sidebar grabbed only the first 500
chars of the btnNewChat handler block to scan for closeMobileSidebar.
The new comment block pushed closeMobileSidebar past that window even
though both calls are still present. Bumped the window to 1500 chars
and the shortcut-block lines from 12 to 24 to match the multi-line guard.
Closes#1432Closes#1423
Reported by @Olyno (#1432, GitHub)
Three fixes from Opus advisor review of stage-261:
1. CRITICAL: dropdown-survives-resize bug. The composerToolsetsDropdown is a
DOM sibling of composerToolsetsWrap, not a child, so CSS hiding the wrap
does not cascade-hide an open dropdown. If a user opens the dropdown at
composer-footer >= 1100px and then opens the workspace panel (or resizes
the window), the dropdown would stay open without a visible anchor.
Fixed in three places (defense-in-depth):
- resize listener: closes dropdown when chip.offsetParent === null
- _positionToolsetsDropdown: closes if chip hidden (defense-in-depth)
- toggleToolsetsDropdown: early-returns if chip hidden (defense against
future #1431 redesign code that might invoke from elsewhere)
2. MEDIUM: display:flex changed to display:block to match sibling wraps
(.composer-profile-wrap, .composer-model-wrap, .composer-reasoning-wrap
all use the natural block display).
3. Added 3 new regression tests to pin all three guards.
Refs #1431, #1433.
Replaces PR #1433 unconditional JS display:none with a CSS @container query
that shows the chip only at composer-footer widths >= 1100px. JS now clears
inline style instead of setting display:none, so the CSS responsive cascade
is the single source of truth. Also removed inline style=\"display:none\" from
index.html so the CSS base rule provides the default-hidden state.
10 regression tests pin the base hide, wide-container show, narrow-container
hide (520px container query), mobile viewport hide (640px @media), JS does
not force display:none, JS clears inline style, /api/session/toolsets and
the dropdown machinery (toggleToolsetsDropdown, _populateToolsetsDropdown)
are preserved.
Refs #1431, #1433.
