Extends _sanitize_messages_for_api() with a two-pass approach:
1. Collect all tool_call_ids declared in assistant messages (handles
both OpenAI 'id' and Anthropic 'call_id' field names).
2. Drop any tool-role messages whose tool_call_id was not declared
by a preceding assistant message.
Strictly-conformant providers (Mercury-2/Inception, newer OpenAI
models) reject histories with orphaned tool results with a 400 error:
'Message has tool role, but there was no previous assistant message
with a tool call.' This can happen when histories are edited, when
switching between providers, or when partial messages are stored.
Adds 13 regression tests covering: valid roundtrip preservation,
multiple tool calls, partial orphan filtering, Anthropic call_id,
edge cases (None tool_calls, missing tool_call_id, non-dict entries).
When a user switches the model via the model picker while a session has
existing messages, a toast now informs them: 'Model change takes effect
in your next conversation'. This prevents confusion when the model
dropdown updates visually but the running conversation continues with
the original model.
Implementation: 4-line addition in modelSelect.onchange in boot.js,
after the existing provider-mismatch warning. Checks S.messages.length
(the reliable in-memory array) and guards showToast with typeof.
Synthesized from PRs #516 (armorbreak001), #517 and #518 (cloudyun888).
Placement follows #518's correct boot.js approach. Reference corrected
from S.session.messages to S.messages (always initialized by loadSession).
4 new tests in test_provider_mismatch.py::TestModelSwitchToast.
Co-authored-by: armorbreak001 <armorbreak001@users.noreply.github.com>
Co-authored-by: cloudyun888 <cloudyun888@users.noreply.github.com>
- sessions.js _formatSourceTag(): return null for unrecognised tags
instead of raw string — prevents legacy 'N/A' values from surfacing
- sessions.js metaBits push: guarded with _stLabel null check so only
known platform labels appear in the session metadata line
- sessions.js [SYSTEM:] title fallback: drop raw s.source_tag middle
term, fall back directly to 'Gateway' for unknown sources
7 new tests in test_issue429.py.
1 updated test in test_sprint40_ui_polish.py (new guarded push pattern).
Closes#429
Bug: the autolink pass stashed <a> tags (via _al_stash) before running,
but did not stash <img> tags. When  was converted to an <img>
tag by the image pass, the subsequent autolink regex matched the URL
inside src="..." and wrapped it in <a href="...">url</a>, producing
src="<a href="...">url</a>" — a completely broken image source.
Fix: extend the _al_stash regex from:
(<a\b[^>]*>[\s\S]*?<\/a>)
to:
(<a\b[^>]*>[\s\S]*?<\/a>|<img\b[^>]*>)
This stashes both <a> and self-closing <img> tags before autolink runs,
then restores them after, so the URL inside src= is never touched.
Adds 7 regression tests in tests/test_issue487b.py.
- test_sprint45.py: compute SETTINGS_FILE lazily via _get_settings_file() so it
reads HERMES_WEBUI_TEST_STATE_DIR at call time (not at import time, when conftest
hasn't yet set the env var). Fixes test isolation across all 1078 tests.
- test_sprint45.py: use auth cookie in teardown when clearing password post-test.
- test_sprint45.py: remove test_synced_version_strings (checks local-patch version).
- static/i18n.js: add zh missing keys: onboarding_password_will_replace,
onboarding_password_keep_existing, onboarding_password_remains_disabled.
- server.py: revert server_version to HermesWebUI/0.50.38 (matches master).
* fix: relax workspace trust boundary to user home directory
The previous restriction required workspaces to be under DEFAULT_WORKSPACE
(/home/hermes/workspace), which blocked all profile-specific workspaces
(~/CodePath, ~/General, ~/WebUI, ~/Camanji, etc.) since each profile uses
a different directory under home.
New boundary: any directory under Path.home() is trusted.
This still blocks /etc, /tmp, /var, /root, /usr and all paths outside the
user's home, while allowing any legitimate workspace under ~/
Also updates test assertions from 'trusted workspace root' to 'outside'
since the new error message says 'outside the user home directory'.
* fix: workspace trust uses home-dir + saved-list, not single ancestor
Three-layer trust model that works cross-platform and multi-workspace:
1. BLOCKLIST: /etc, /usr, /var, /bin, /sbin, /boot, /proc, /sys, /dev, /root,
/lib, /lib64, /opt/homebrew — always rejected, even if somehow saved
2. HOME CHECK: any path under Path.home() is trusted — covers ~/CodePath,
~/hermes-webui-public, ~/WebUI, ~/General, ~/Camanji simultaneously;
Path.home() is cross-platform (Linux ~/..., macOS ~/..., Windows C:\Users\...\...)
3. SAVED LIST ESCAPE HATCH: if a path is already in the saved workspace list,
it's trusted regardless of location — covers self-hosted deployments where
workspaces live outside home (/data/projects, /opt/workspace, etc.)
None/empty → DEFAULT_WORKSPACE (always trusted, validated at startup)
* docs: v0.50.35 release — version badge and CHANGELOG
---------
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
* fix(workspace): restrict session workspaces to trusted roots
* fix: use boot-time DEFAULT_WORKSPACE instead of profile default for trusted workspace root
_profile_default_workspace() reads the agent's terminal.cwd which may differ
from the WebUI's configured workspace root. Use _BOOT_DEFAULT_WORKSPACE (which
respects HERMES_WEBUI_DEFAULT_WORKSPACE for test isolation) to stay consistent
with how new_session() seeds the initial workspace.
* docs: v0.50.34 release — version badge and CHANGELOG
---------
Co-authored-by: hinotoi-agent <paperlantern.agent@gmail.com>
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
* fix: workspace panel close button — no duplicate X on desktop, mobile X respects file preview
Two bugs fixed in the workspace right panel:
1. Duplicate X on desktop (bug): #btnClearPreview (the X icon) was always
visible alongside #btnCollapseWorkspacePanel (the chevron), producing two
close controls at once. Fixed in syncWorkspacePanelUI() — on desktop, the X
is now hidden when no file preview is open (display:none), and only shown
when the user is viewing a file. The chevron remains as the sole close
control in browse mode.
2. Mobile X collapses panel instead of dismissing file (bug): .mobile-close-btn
was calling closeWorkspacePanel() directly, which collapsed the whole panel
even when a file was open. Changed to handleWorkspaceClose(), which already
has the correct two-step logic: clear preview first, close panel only if
no preview is visible.
Files changed:
- static/boot.js: syncWorkspacePanelUI() hides btnClearPreview on desktop
when hasPreview is false, guarded by !isCompact so mobile is unaffected
- static/index.html: mobile-close-btn onclick changed from
closeWorkspacePanel() to handleWorkspaceClose()
- tests/test_sprint44.py: 10 new regression tests
- tests/test_mobile_layout.py: updated test_workspace_close_button_present()
to accept handleWorkspaceClose() as the valid onclick target
* fix: widen test_server_delete_invalidates_index window to 1200 chars
The test extracted a 600-char window starting from the session/delete
handler to check for SESSION_INDEX_FILE. Commit 3cc5839 added session_id
character validation and path traversal guards before the unlink call,
pushing SESSION_INDEX_FILE to ~764 chars from the match — beyond the
600-char limit, causing the test to fail on CI.
Widened the window to 1200 chars, which accommodates any reasonable
amount of guard code before the SESSION_INDEX_FILE.unlink() call.
* docs: v0.50.33 release — version badge and CHANGELOG
---------
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
* fix: delegate all live model fetching to agent's provider_model_ids()
Previously _handle_live_models() maintained its own per-provider logic:
- anthropic, google, gemini returned 'not_supported' (hardcoded exclusions)
- openai-codex had a custom branch (added in v0.50.30)
- openai/copilot had hardcoded base URLs
- other providers fell through to a generic /v1/models fetch
Now the handler delegates entirely to hermes_cli.models.provider_model_ids(),
which is the agent's authoritative resolver:
- anthropic: live fetch via /v1/models with correct API-key or OAuth headers
- copilot: live fetch from api.githubcopilot.com/models with Copilot headers
- openai-codex: Codex OAuth endpoint + ~/.codex/ cache fallback
- nous: live fetch from Nous inference portal
- deepseek, kimi-coding: generic OpenAI-compat /v1/models
- opencode-zen/go: OpenCode live catalog
- openrouter: curated static list (live returns 300+ which is overwhelming)
- google/gemini, zai, minimax: static list (non-standard or Anthropic-compat endpoints)
- any others: graceful static fallback
Also removed the client-side skip guard in _fetchLiveModels() (ui.js) that
blocked live fetching for anthropic, google, and gemini.
The hardcoded model lists in _PROVIDER_MODELS remain as the fallback when
credentials are missing or network is unavailable — they are never shown
when live data is available.
* docs: v0.50.31 release — version badge and CHANGELOG
---------
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
* fix: route openai-codex live model fetch through agent's get_codex_model_ids()
Previously _handle_live_models() grouped openai-codex with openai and sent a
request to https://api.openai.com/v1/models, which returns 403 because Codex
auth is OAuth-based via chatgpt.com, not a standard API key. The live fetch
silently failed and the UI showed only the hardcoded static list.
Now: openai-codex has a dedicated early-exit branch that calls
hermes_cli.codex_models.get_codex_model_ids() — the same path the agent CLI
uses. It resolves models in order: live Codex API (if OAuth token available) >
~/.codex/ local cache > DEFAULT_CODEX_MODELS. This means:
- If the user has a valid Codex OAuth session, the UI gets the exact model list
their subscription provides (e.g. gpt-5.2, gpt-5.3-codex-spark that aren't
in the hardcoded list)
- If the OAuth session is expired, falls back to local ~/.codex/ cache
- Always has DEFAULT_CODEX_MODELS as final fallback
Also: improved label generation for Codex model IDs (GPT-5.4 Mini vs GPT 5 4 Mini).
Added 1 structural regression test.
* docs: v0.50.30 release — version badge and CHANGELOG
---------
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
* fix: correct tool call card rendering on session load
Two bugs caused duplicate/incorrect tool call cards when loading
sessions (especially after context compaction):
1. loadSession() sanitized messages (B9 filter) but did NOT update
the session-level tool_calls array's assistant_msg_idx references.
Since compact() returns only sanitized messages and recomputes
tool_calls with indices into the compacted array, the original
assistant_msg_idx values became stale/misaligned.
2. loadSession() then assigned the broken session-level tool_calls
directly to S.toolCalls. This prevented renderMessages()'s fallback
path (which derives tool_calls from per-message tool_calls using
correct sanitized-array indices) from ever running.
Fix:
- Keep full sanitization loop with index remapping for session-level
tool_calls (in case they're needed by other code paths).
- Instead of assigning broken session-level tool_calls to S.toolCalls,
set S.toolCalls=[] so renderMessages() uses the fallback derivation
from per-message tool_calls, which already have correct indices.
* test: add 8 regression tests for issue #401 tool call index remapping
* docs: v0.50.29 release — version badge and CHANGELOG
---------
Co-authored-by: Frank Song <franksong2702@gmail.com>
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
* fix: expand openai-codex model catalog to match agent DEFAULT_CODEX_MODELS
The _PROVIDER_MODELS["openai-codex"] catalog only listed codex-mini-latest,
so the model dropdown for profiles using openai-codex provider (e.g. CodePath)
showed only that one entry — even when the profile's saved default_model was
gpt-5.4 or another standard Codex model.
Updated to match DEFAULT_CODEX_MODELS from hermes_cli/codex_models.py:
- gpt-5.4
- gpt-5.4-mini
- gpt-5.3-codex
- gpt-5.2-codex
- gpt-5.1-codex-max
- gpt-5.1-codex-mini
- codex-mini-latest (kept, relabeled as 'Codex Mini (latest)')
Also adds 2 regression tests: catalog includes gpt-5.4, display name correct.
* docs: v0.50.28 release — version badge and CHANGELOG
---------
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
* fix: restore mobile chat scrolling and drawer close (#397)
- static/style.css: add min-height:0 to .layout and .main (flex shrink chain fix for mobile scroll)
- static/style.css: add -webkit-overflow-scrolling:touch, touch-action:pan-y, overscroll-behavior-y:contain to .messages
- static/boot.js: call closeMobileSidebar() on new-conversation button onclick and Ctrl+K shortcut
- tests/test_mobile_layout.py: 41 new lines covering all three CSS fixes and both JS call sites
Original PR by @Jordan-SkyLF
* fix: preserve imported session timestamps (#395)
- api/models.py: add touch_updated_at: bool = True param to Session.save(); import_cli_session() accepts created_at/updated_at kwargs and saves with touch_updated_at=False
- api/routes.py: extract created_at/updated_at from get_cli_sessions() metadata and forward to import_cli_session(); use touch_updated_at=False on post-import save
- tests/test_gateway_sync.py: +53 lines — integration test verifying imported session keeps original timestamp and sorts correctly vs newer sessions; also fix: add WebUI session file cleanup in finally block
Original PR by @Jordan-SkyLF
* fix(profiles): block path traversal in profile switch and delete flows (#399)
Master was vulnerable: switch_profile and delete_profile_api joined user-supplied profile
names directly into filesystem paths with no validation. An attacker could send
'../../etc/passwd' as a profile name to traverse outside the profiles directory.
- api/profiles.py: add _resolve_named_profile_home(name) — validates name with
^[a-z0-9][a-z0-9_-]{0,63}$ regex then enforces path containment via
candidate.resolve().relative_to(profiles_root); use in switch_profile()
- api/profiles.py: add _validate_profile_name() call to delete_profile_api() entry
- api/routes.py: add _validate_profile_name() call at HTTP handler level for
both /api/profile/switch and /api/profile/delete (fail-fast at API boundary)
- tests/test_profile_path_security.py: 3 tests — traversal rejected, valid name passes
Cherry-picked commit aae7a30 from @Hinotoi-agent (PR was 62 commits behind master)
* feat: add desktop microphone transcription fallback (#396)
Mic button now works in browsers that support getUserMedia/MediaRecorder but
lack SpeechRecognition (e.g. Firefox desktop, some Chromium builds).
- static/boot.js: detect _canRecordAudio (navigator.mediaDevices + getUserMedia + MediaRecorder);
keep mic button enabled when either SpeechRecognition or MediaRecorder is available;
MediaRecorder fallback records audio, sends blob to /api/transcribe, inserts transcript
into the composer; _stopMic() handles all three states (recognition, mediaRecorder, neither)
- api/upload.py: add transcribe_audio() helper — saves uploaded blob to temp file, calls
transcription_tools.transcribe_audio(), always cleans up temp file
- api/routes.py: add /api/transcribe POST handler — CSRF protected, auth-gated, 20MB limit,
returns {text:...} or {error:...}
- api/helpers.py: change Permissions-Policy microphone=() to microphone=(self) (required to
allow getUserMedia in the same origin)
- tests/test_voice_transcribe_endpoint.py: 87 new lines — 3 tests with mocked transcription
- tests/test_sprint19.py: +1 regression guard (microphone=(self) in Permissions-Policy)
- tests/test_sprint20.py: 3 updated tests for new fallback-capability checks
Original PR by @Jordan-SkyLF
* docs: v0.50.25 release — version badge and CHANGELOG
---------
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
* feat(ui): opt-in chat bubble layout
Closes#336.
Adds a settings toggle that right-aligns user messages and left-aligns
assistant replies. Off by default - the current full-width layout is
friendlier to code blocks and tool output, so bubbles are strictly
opt-in per the maintainer note on the issue.
Wiring follows the existing token-usage / cli-sessions pattern:
- api/config.py: new bubble_layout bool in _SETTINGS_DEFAULTS and
_SETTINGS_BOOL_KEYS, validated + persisted like the rest.
- static/style.css: .bubble-layout gated selectors using :has() to
tag msg-rows by .msg-role.user / .msg-role.assistant without any JS
changes to message creation. User rows get align-self: flex-end,
max-width: 75%, and a row-reverse header; assistant rows flex-start.
A 700px media query widens the max to 92% on narrow screens.
- static/index.html: new checkbox with i18n keys next to the existing
token-usage toggle.
- static/panels.js: loads the setting into the checkbox, saves it
back, and toggles body.bubble-layout immediately on save.
- static/boot.js: applies the class on initial load so refreshed
tabs honor the persisted setting without a flash.
- static/i18n.js: English label + description.
Test suite errors are environmental (test server fails to start on
port 8788 on main as well).
* i18n(es): add Spanish translations for bubble_layout setting
* fix+test: boot.js bubble-layout reset on failure; add 22 tests for issue #336
* docs: v0.50.24 release — version badge and CHANGELOG
---------
Co-authored-by: Matt Van Horn <455140+mvanhorn@users.noreply.github.com>
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
* Add OpenCode Zen and OpenCode Go provider support
The webui model dropdown had no knowledge of these providers.
When hermes_cli detected them as authenticated, they fell through
to the unknown-provider fallback showing wrong models.
Changes:
- Add opencode-zen and opencode-go to _PROVIDER_DISPLAY
- Add model lists for both to _PROVIDER_MODELS
- Add OPENCODE_ZEN_API_KEY and OPENCODE_GO_API_KEY to env-var fallback detection
- Fix custom:* provider IDs (e.g. custom:my-server) displaying raw ID instead of "Custom"
* Add tests for OpenCode provider registration and detection
---------
Co-authored-by: David Case <david.case@shruggr.cloud>
- Read X-Forwarded-For and X-Real-IP before falling back to raw socket IP
- Add HERMES_WEBUI_ONBOARDING_OPEN=1 env var escape hatch for remote servers
- Error message now includes the env var hint
- 18 new tests (TestOnboardingIPLogic + TestOnboardingSetupEndpoint)
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
* fix: persist durable inflight reload snapshots
* fix: remove duplicate loadInflightState stub, update CHANGELOG test count
The stub added in the previous review branch is superseded by the author's
real localStorage-backed implementation in the cherry-picked commit 36051c0.
Remove the duplicate. Update CHANGELOG to 961 tests and document the durable
inflight state feature.
---------
Co-authored-by: Jordan SkyLF <jordan@skylinkfiber.net>
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
* fix: preserve live session output across chat switches
(cherry picked from commit 401e3b643d25e8dad8c06883b478b3c3073f07a5)
* fix: preserve todo state after session reload
(cherry picked from commit 7ee093ba19978af23b79148df2f2347e2f1e5bde)
* fix: preserve live assistant anchor across rerenders
* fix: stream live reasoning and tool progress
* fix: recover inflight session state after reload
* fix: add loadInflightState stub + CHANGELOG v0.50.21
- static/ui.js: add loadInflightState() function (currently returns null —
the typeof guard in sessions.js means reload recovery works via the
else-path attachLiveStream call; this stub satisfies the guard cleanly
and documents the extension point for future localStorage-backed state)
- CHANGELOG.md: v0.50.21 entry; 960 tests (up from 949)
---------
Co-authored-by: Jordan SkyLF <jordan@skylinkfiber.net>
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
* fix: silent errors, stale models, live model fetching (#373, #374, #375)
- api/streaming.py: detect empty agent response (_assistant_added check),
emit apperror(type='no_response' or 'auth_mismatch') instead of silent done
- api/streaming.py: add _token_sent flag so guard works for streaming agents
- static/messages.js: done handler belt-and-suspenders guard for zero replies
- static/messages.js: apperror handler labels 'no_response' type distinctly
- api/config.py: remove gpt-4o and o3 from _FALLBACK_MODELS and
_PROVIDER_MODELS['openai'] (superseded by gpt-5.4-mini and o4-mini)
- api/routes.py: new /api/models/live?provider= endpoint, fetches /v1/models
from provider API with B310 scheme check + SSRF guard
- static/ui.js: _fetchLiveModels() background fetch after static list loads,
appends new models to dropdown, caches per session, skips unsupported providers
Other:
- tests/test_issues_373_374_375.py: 25 new structural tests
- tests/test_regressions.py: extend done-handler window 1500->2500 chars
- CHANGELOG.md: v0.50.19 entry; 947 tests (up from 922)
* fix: SSRF hostname bypass + auth detection operator precedence
1. routes.py: SSRF guard used substring matching (any(k in hostname))
which allows bypass via hostnames like evil-ollama.attacker.com.
Changed to exact hostname matching against a fixed set of known
local hostnames (localhost, 127.0.0.1, 0.0.0.0, ::1).
2. streaming.py: _is_auth detection had a Python operator precedence
bug on the ternary expression. The line:
'AuthenticationError' in type(...).__name__ if _last_err else False
parsed as the ternary absorbing the rest of the or-chain when
_last_err was falsy. Fixed to: (_last_err and 'AuthenticationError' in ...)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* docs: fix v0.50.20 CHANGELOG version number and test count (949 tests)
---------
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: pre-install uv in Docker image + fix workspace dir permissions (#357)
Two fixes for Docker startup reliability:
1. Install uv at build time in the Dockerfile so the container works
without internet access at runtime. The init script now skips the
download when uv is already on PATH.
2. Use sudo mkdir/chown for the workspace directory, matching the
pattern used for /app. Docker auto-creates bind-mount directories
as root, leaving them unwritable by the hermeswebui user.
Fixes#357
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: Docker uv pre-install as root to /usr/local/bin + tests + CHANGELOG
Dockerfile: install uv as root with UV_INSTALL_DIR=/usr/local/bin so it
lands in /usr/local/bin (system PATH) rather than /home/hermeswebuitoo/.local/bin
which the hermeswebui runtime user can't see.
tests/test_issue357.py: 15 structural tests covering Dockerfile uv build-time
install (system-wide, as root, before COPY), init script skip-if-present
logic, and workspace sudo mkdir/chown.
CHANGELOG.md: v0.50.17 entry; 915 tests (up from 900)
---------
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: CSRF check fails behind reverse proxy on non-standard ports
When serving behind a reverse proxy (e.g. Nginx Proxy Manager) on a
non-standard port like 8000, the browser sends
`Origin: https://example.com:8000` but the proxy forwards `Host: example.com`
(without the port). The existing CSRF check compared these as raw strings,
causing all POST requests to be rejected with 403.
This commit:
- Adds `_normalize_host_port()` to properly parse host:port pairs (incl. IPv6)
- Adds `_ports_match()` that treats absent port as equivalent to 80/443
- Adds `HERMES_WEBUI_ALLOWED_ORIGINS` env var for explicitly trusting origins
when port normalization alone isn't sufficient (e.g. port 8000)
- Adds unit tests covering port normalization, allowlist, and rejection cases
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: CSRF port normalization — scheme-aware, allowlist validation, 29 tests (#360)
api/routes.py:
- _normalize_host_port(): parse host:port including IPv6 bracket notation
- _ports_match(scheme, origin_port, allowed_port): scheme-aware — http absent=:80,
https absent=:443; prevents cross-protocol false match (http://host:80 no
longer passes for https://host:443 server)
- _allowed_public_origins(): parse HERMES_WEBUI_ALLOWED_ORIGINS env var;
warn and skip entries missing scheme prefix
- _check_csrf(): extract origin scheme, pass to _ports_match; add origin_scheme
tests/test_sprint29.py: 29 new tests (5 from PR + 24 added in review)
- Unit tests for _normalize_host_port and _ports_match helpers
- Cross-protocol rejection (http vs https default ports)
- Explicit :80 / :443 same-origin pass
- Non-default port rejection
- Bug scenario with/without allowlist
- Comma-separated allowlist
- No-scheme allowlist warning
- Trailing-slash normalization
CHANGELOG.md: v0.50.16 entry; 900 tests total (up from 871)
---------
Co-authored-by: liangxu.5 <liangxu.5@bytedance.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
