The /background feature was fundamentally non-functional as shipped —
two coupled bugs kept results from ever reaching the user:
1. complete_background() was defined but NEVER called. The
_handle_background thread ran _run_agent_streaming and then exited;
no hook signalled the task tracker that the work was done. Every
background task stayed in status="running" forever and
get_results() (which filters to done-only) always returned [].
2. get_results() called _BACKGROUND_TASKS.pop(parent_sid, []) which
removed the ENTIRE list — including tasks still in flight. Even if
bug #1 were fixed, the first frontend poll during a long-running
task would drop the task from the tracker, and
complete_background()'s loop would iterate over an empty list when
the worker eventually finished — the result would still be lost.
Fix:
- api/background.py::get_results now retains running tasks in the
dict; only done ones are popped and returned.
- api/routes.py::_handle_background wraps _run_agent_streaming in an
inline worker (_run_bg_and_notify) that, after streaming completes,
reloads the hidden bg session, extracts the last non-error assistant
message, and calls complete_background(parent_sid, task_id, answer).
Worker also best-effort unlinks the hidden bg session file so
SESSION_DIR doesn't accumulate debris.
- Exception safety: any failure in _run_agent_streaming or the
post-processing path still calls complete_background with a fallback
sentinel so the frontend's polling loop doesn't hang forever.
Added 5 regression tests in tests/test_background_tasks.py:
- running tasks survive get_results polls
- done tasks are returned and removed
- poll → complete → poll round-trip surfaces the answer (this is the
original bug's reproduction path)
- empty parent is cleaned up
- static check: _handle_background's worker calls complete_background
and uses Session.load to extract the answer
Full suite: 2023 passed, 0 failed.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Rebased onto master after #931 (aux title routing) to resolve streaming.py conflict.
All changes from both PRs are cleanly integrated.
2088 tests passing (2065 master + 23 from #931).
Co-authored-by: bergeouss <bergeouss@gmail.com>
- _aux_title_configured(): returns True when provider/model/base_url is set
- _aux_title_timeout(): reads configured timeout, falls back to 15.0s default
- _generate_llm_session_title_via_aux: use_agent_model kwarg preserves old behavior
- Missing llm_invalid_aux fallback now triggers agent-model retry
- 23 new tests in tests/test_title_aux_routing.py — all pass
Co-authored-by: starship-s <starship-s@users.noreply.github.com>
streaming-markdown@0.2.15 preserves arbitrary URL schemes in href/src.
Verified with a Node + jsdom harness:
IN : [click](javascript:alert(1))
OUT: <p><a href="javascript:alert(1">click</a>)</p> ← XSS vector
Confirmed unsafe for: javascript:, vbscript:, data:text/html, file://.
The library uses only safe DOM primitives (createElement/appendChild/
createTextNode — no innerHTML/eval), so <script> tags are escaped as
text, but URL-scheme filtering is absent. The existing renderMd() path
implicitly filtered to http(s) via its regex, so this is a regression
the moment streaming markdown is enabled.
Attack path: agent echoes prompt-injection content containing a
markdown link with javascript: href → smd renders it live → user clicks
during the streaming window → JS executes in webui origin → session
cookie, API calls, etc.
Fix: walk the live DOM after each parser_write (and again after
parser_end) and remove href/src attributes whose scheme isn't on the
safe allowlist (http, https, mailto, tel, and relative/anchor paths).
Blocked anchors keep their text content but lose href; blocked images
lose src and get data-blocked-scheme="1" for debugging.
Harness confirms all 10 tested cases behave correctly — javascript:,
vbscript:, data:text/html, file:// all stripped; https://, /path,
#anchor, mailto:, tel: all preserved.
Added 5 regression tests in TestSmdUrlSchemeSanitization that lock:
- the sanitize helper exists
- the allowlist regex permits https? and forbids javascript/vbscript/data:
- _smdWrite invokes sanitize after parser_write
- _smdEndParser invokes sanitize after parser_end
- the sanitizer covers both <a href> and <img src>
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix: persist onboarding_completed for CLI-configured users on first chat_ready (v0.50.179, #921)
Co-authored-by: bsgdigital
* fix(onboarding): don't 500 the status endpoint if save_settings fails
The #921 persist call `save_settings({"onboarding_completed": True})` in
get_onboarding_status() raises if the settings.json write fails
(read-only filesystem, disk full, permission error). That turns every
/api/onboarding/status call into a 500 until the disk is writable,
which is much worse UX than losing the persistence-across-restart guard.
Wrapped in try/except so persistence becomes best-effort. The function
still sets settings["onboarding_completed"] = True in memory on success,
and `completed` reflects `config_auto_completed` on this request either
way, so the user sees the right state even when the write fails — only
the next-restart protection degrades.
Added regression test that patches save_settings to raise OSError and
asserts the endpoint still returns completed=True without raising.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: nesquena-hermes <nesquena-hermes@users.noreply.github.com>
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat: add PWA support (manifest, service worker, install prompt) (v0.50.178, #911)
Co-authored-by: bsgdigital
Closes#685
* fix(sw): await caches.match() before `|| fallback` so offline HTML actually shows
The offline-navigation fallback was dead code:
return caches.match('./') || new Response('<html>...</html>', ...);
`caches.match()` returns a Promise, and Promise objects are always truthy
in a `||` check — so the `new Response(...)` branch was never taken. On
actual offline, `caches.match('./')` resolves to undefined (no cache hit
for the root), the SW returns undefined, and the browser falls back to
its own default offline page. The custom "Hermes requires a server
connection" HTML was unreachable.
Fix by threading the match through `.then()` so the resolved value (not
the Promise object) feeds the `||`:
return caches.match('./').then((cached) => cached || new Response(...));
Added 13 regression tests in tests/test_pwa_manifest_sw.py covering:
- manifest.json validity + required PWA fields + icon existence
- sw.js cache-version placeholder + API/stream bypass + correct offline
pattern (explicitly rejects the broken `|| new Response` shape so it
can't regress)
- /manifest.json + /sw.js routes serve correct Content-Type,
Cache-Control, Service-Worker-Allowed headers and inject WEBUI_VERSION
- index.html links manifest, registers SW, has iOS PWA meta tags
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: nesquena-hermes <nesquena-hermes@users.noreply.github.com>
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix: correct interleaved streaming order (Text → Thinking → Tool → Text)
During live streaming, tool cards were inserted before their associated
thinking cards instead of after them. The root cause was that
appendLiveToolCard's anchor selector didn't include .thinking-card-row,
so finalized thinking cards were skipped when finding the insertion point.
Changes:
- messages.js: Add segment splitting (segmentStart/_freshSegment) so each
text segment after a tool call renders only its own slice, not the full
accumulated text. Sync thinking card render in reasoning handler to
avoid rAF race with tool events. Guard removeThinking() to preserve
finalized cards when reasoningText is active.
- ui.js: Add .thinking-card-row to appendLiveToolCard anchor selector so
tool cards land after finalized thinking. Add anchor-based positioning
to appendThinking for correct interleaved placement. Clean up empty
spinner-only thinking rows in finalizeThinkingCard. Add 3-dot waiting
indicator (toolRunningRow) after tool cards for visual feedback.
- style.css: Scope blinking cursor to last live-assistant segment only.
Add spacing for toolRunningRow.
* chore: CHANGELOG for v0.50.174
---------
Co-authored-by: bsgdigital <bsgdigital@users.noreply.github.com>
Co-authored-by: nesquena-hermes <nesquena-hermes@users.noreply.github.com>
* fix(renderer): ordered list items always showed 1. — emit value= on each <li> (#886)
Root cause: when LLMs output numbered lists with blank lines between items,
renderMd()'s paragraph-splitter (split(/\n{2,}/)) breaks the markdown into
one chunk per item. The ordered-list regex then wraps each item in its own
<ol>, and since each <ol> restarts at 1, the rendered output is always 1. 1. 1.
Fix: capture the original number from each list line and emit value="N" on
every <li>. The HTML spec guarantees that value= overrides the <ol> counter,
so even items in separate <ol> containers display their correct ordinal.
6 regression tests in tests/test_886_ordered_list_numbering.py.
1958 tests pass.
* chore: add v0.50.173 CHANGELOG entry for ordered list fix
---------
Co-authored-by: Hermes Bedrock Fix <hermes-fixes@local>
Co-authored-by: nesquena-hermes <nesquena-hermes@users.noreply.github.com>
* fix(cancel): preserve partial streamed response on Stop Generation (#893)
* docs(cancel): fix misleading comment — partial message is NOT _error=True
The outer comment block claimed `_error=True so _sanitize_messages_for_api()
strips it from future conversation history`, but the actual append call
sets only `_partial=True` (correctly matching the inner comment six lines
below and the PR description). Updated the outer comment to match reality
so a future reader doesn't try to "fix" the code to match the wrong comment.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: nesquena-hermes <nesquena-hermes@users.noreply.github.com>
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(models): preserve @nous: prefix in settings + fix cross-namespace 404 for Nous (#895#894)
* fix(review): persist bare form for CLI compatibility + picker smart-match
The PR persisted `@nous:anthropic/claude-opus-4.6` verbatim to config.yaml
to make the Settings picker match its dropdown options (which carry the
`@nous:` prefix after #885). That fixes the WebUI picker but introduces a
cross-tool regression: hermes-agent's CLI reads `config.yaml -> model.default`
directly and passes it to the provider API verbatim. For aggregator providers
(Nous is one — see hermes_cli/model_normalize.py `_AGGREGATOR_PROVIDERS`),
`normalize_model_for_provider` is skipped entirely (run_agent.py:887), so
the literal `@nous:anthropic/...` string flows to the Nous API, which rejects
it — breaking every user who runs `hermes` in the terminal right after
saving via WebUI.
Fix the tension at the picker rather than the persistence: the existing
`_findModelInDropdown()` smart matcher already normalises both sides
(lowercase, strip namespace prefix, dashes→dots) so a saved bare
`anthropic/claude-opus-4.6` resolves to the `@nous:anthropic/claude-opus-4.6`
option automatically. Applied this in panels.js via `_applyModelToDropdown()`.
Changes:
api/config.py revert the @-prefix preservation; persist the
resolved bare/slash form (CLI-compatible)
static/panels.js Settings picker uses _applyModelToDropdown()
instead of raw `.value =` so saved bare forms
still select the matching @nous: option
tests test renamed + asserts bare persisted form;
new test locks the smart-matcher contract
This also improves behaviour for a dormant case not flagged in #895: a user
who set their default via `hermes model X` and opens Settings for the first
time used to see a blank picker (bare form vs prefixed options). Now the
smart matcher finds the right option, so the "open Settings → save → bare
form in config.yaml" round-trip is stable for both CLI- and WebUI-origin
saves.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* chore: update CHANGELOG v0.50.171 — bare-form persistence + picker smart-match
---------
Co-authored-by: nesquena-hermes <nesquena-hermes@users.noreply.github.com>
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(settings): show live models in default model picker and apply to new chats (#872)
Two related bugs:
1. Settings > Preferences > Default Model dropdown only showed static models
from /api/models — live-fetched models (e.g. @nous:anthropic/claude-opus-4.7)
were missing. Now calls _fetchLiveModels() on the settings picker too.
2. New chats ignored the saved default model preference — they always used the
chat-header dropdown value (which reflects the previous session's model).
Now newSession() uses the saved default_model and syncs the dropdown.
Extracted _addLiveModelsToSelect() from _fetchLiveModels() so cached live models
can be applied to any <select> element (chat-header or settings picker).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* fix(tests): update live-model prefix tests for _addLiveModelsToSelect extraction
The tests searched for og.dataset.provider, _isPortalFetch, and openrouter
exclusion patterns inside _fetchLiveModels(). These were extracted into
_addLiveModelsToSelect() as part of the #872 fix. Updated regex targets to
check _addLiveModelsToSelect first, falling back to _fetchLiveModels.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* chore: add multi-tab note on window._defaultModel
Clarifies that window._defaultModel is per-page-load and not synced
across browser tabs, following maintainer feedback on #889.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* chore: CHANGELOG for v0.50.170
* chore: trigger PR refresh after rebase
---------
Co-authored-by: fr33m1nd <bergeouss@gmail.com>
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
Co-authored-by: nesquena-hermes <nesquena-hermes@users.noreply.github.com>
* feat(models): live-first model fetching for all OpenAI-compat providers (#871)
The WebUI model picker relied on hardcoded _PROVIDER_MODELS as primary
source for providers like zai, minimax, mistralai, xai, openai-codex,
deepseek, and gemini. These lists go stale — new models don't appear
until someone manually updates the dict.
Add an OpenAI-compat /v1/models fetch fallback in _handle_live_models()
that fires when provider_model_ids() is unavailable or returns []. The
resolution chain is now:
1. hermes_cli.provider_model_ids() (agent's live fetch)
2. Custom providers from config.yaml
3. Direct /v1/models fetch for known OpenAI-compat endpoints
4. Static _PROVIDER_MODELS as last-resort offline fallback
Covers: zai, minimax, mistralai, xai, openai-codex, deepseek, gemini.
Uses urllib (stdlib) — no new dependencies. Static lists remain as
offline fallback so the UI always shows something.
Closes#871
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* refactor(models): address review feedback on live fetch (#892)
Five changes from nesquena-hermes review:
1. Move _OPENAI_COMPAT_ENDPOINTS to module level — avoid dict
reconstruction per request
2. Document urllib blocking behavior — 8s timeout acceptable because
server is threaded and frontend enriches in background
3. Add TODO comment for TTL-based caching follow-up
4. Remove openai-codex from endpoint map — same endpoint as base
openai provider, already covered by provider_model_ids()
5. Restrict API key lookup to provider-scoped and model.api_key only
— remove top-level api_key fallback to prevent cross-provider
key leakage
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
The _ob_stash regex in renderMd() used (<code>[^<]*</code>) which failed
to match <code class="language-sql"> tags (attributes) and couldn't capture
multiline content. Code blocks leaked into the bold/italic pipeline,
corrupting SQL/C# comments into <strong><em> tags and producing <
artifacts.
Replace with (<code\b[^>]*>[\s\S]*?</code>) to handle attributes and
multiline content correctly.
Closes#890
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
fix: Nous static models use @nous: prefix — v0.50.164 (#885)
Follow-up to #854 / PR #870. The previous fix made Nous static IDs
slash-prefixed and added a portal-guard branch to resolve_model_provider().
This tightens the static list to use the explicit @nous: prefix, matching
the format of live-fetched models after ui.js's _fetchLiveModels() portal-
prefix step.
The @provider:model branch in resolve_model_provider() is more explicit and
reliable than the portal-guard fallback. Both static and live-fetched paths
now converge on the same resolver output — and as a side effect, the dedup
check in _fetchLiveModels() now correctly identifies static entries as already
present, eliminating duplicate entries in the dropdown for Nous users.
Verified: all 29 Nous models in the browser dropdown carry @nous: prefix,
routing confirmed correct via resolve_model_provider() for all 4 static IDs,
1941 tests passing.
Closes#854.
fix: correct message ordering after task cancellation — v0.50.163 (#883)
Fixes the message-ordering glitch from #882: clicking Cancel while the
agent is responding could cause a subsequent response to render above
the "*Task cancelled.*" marker.
Root cause: the cancel handler pushed the marker only to local S.messages
without persisting to the server. When the done event fired shortly after
and replaced S.messages from server state, the marker disappeared from
client state while the next response anchored to the server-authoritative
position.
Fix has three parts:
- Server (cancel_stream): append *Task cancelled.* to session.messages
with _error:True + timestamp, then save. _error ensures
_sanitize_messages_for_api() strips it from conversation_history on
the next agent turn, so the LLM never sees it as a prior assistant
turn. Precedent: same flag used for the apperror marker at line 1343.
- Client (SSE cancel handler): fetch /api/session instead of pushing
locally (same pattern as the done handler). Falls back to local push
if the fetch fails.
- Tests: fix test window width for cancel handler (1200→dynamic); add
two regression tests pinning _error flag and _sanitize invariant.
1941 tests passing.
Co-authored-by: piliang <piliang1@jd.com>
Adds GET /api/workspaces/suggest endpoint and autocomplete dropdown in the Spaces panel. Suggestions limited to trusted roots (home, saved workspaces, boot default). Keyboard nav, Tab completion, hidden dir support. Symlink-escape and dotdot-escape invariants locked by regression tests.
Root cause: test_profile_env_isolation.py and test_profile_path_security.py called sys.modules.pop() without restoring, poisoning subsequent tests. Fix: monkeypatch.delitem so pytest auto-restores. Also holds _ENV_LOCK for full I/O cycle in _write_env_file and creates .env at 0600 via os.open. Reviewed by Opus (no independent review needed — test/providers fix only).
User bubble selection contrast fixed via scoped ::selection CSS (closes#877). Also adds missing provider i18n keys to es/de/zh/ru/zh-Hant locales, fixing 3 CI failures that crept in from PR #867.
docker_init.bash now checks /opt/hermes as a fallback alongside the primary path. Warning updated with concrete mount guidance. Volume type notes added to compose files and README.
Two bugs fixed: (1) _PROVIDER_MODELS["nous"] updated to slash-prefixed IDs that Nous API expects. (2) resolve_model_provider() now routes portal provider models through the portal (not OpenRouter) and preserves the full slash-prefixed model ID. 10 regression tests.
Update sprint history table in ROADMAP.md through v0.50.156, fix test count header, add Known Limitations section to BUGS.md, update SPRINTS.md header. Reviewed by Opus — factually accurate, table column alignment fixed.
Breaking: auto_install_agent_deps() is now disabled by default. Set HERMES_WEBUI_AUTO_INSTALL=1 to re-enable. New _trusted_agent_dir() checks ownership and permission bits. Addresses #842 by @tomaioo.
Pass gateway_session_key=session_id to AIAgent from streaming.py so Honcho per-session strategy pins to stable WebUI session ID rather than creating a new Honcho session each turn.
Surfaces providers added via credential_pool in the model dropdown. Ambient gh-cli tokens suppressed. _apply_provider_prefix helper extracted. Ollama Cloud display name + dynamic model list. looksLikeBareOllamaId heuristic tightened. Test isolation fixed.
PR #820 by @starship-s.
Bundles three bug fixes (PRs #847, #848, #849) and updates README test count to 1858.
- v0.50.148: prune stale _index.json ghost rows after session-id rotation (closes#846)
- v0.50.149: side-effect-free GET /api/session model resolution (closes#845)
- v0.50.150: profile switching cookie persist + syncTopbar fix + active indicator state
Three related profile-switching fixes:
- Always persist hermes_profile=default cookie when switching back to default (was being cleared with max-age=0, causing fallback to process-global profile)
- Replace undefined updateWorkspaceChip() with syncTopbar() in the sessionInProgress branch of switchToProfile()
- Make sidebar/dropdown active-profile rendering prefer S.activeProfile client state when available, with safe fallback
Tests: 1854 passing.
Replace _normalize_session_model_in_place() on the GET /api/session read path with a read-only _resolve_effective_session_model_for_display() that returns the effective display model without writing it back to disk or the session index.
Closes#845.
Tests: 1856 passing.
Prune ghost _index.json rows whose backing session file no longer exists, on both incremental index writes and all_sessions() reads. Fixes duplicate session entries after session-id rotation (e.g. context compression). Also pre-snapshots in_memory_ids under a single LOCK acquisition in all_sessions() rather than one per row.
Closes#846.
Review additions: optimised lock pattern in all_sessions() (one LOCK acquisition instead of N). Tests: 1856 passing.
* fix(appearance): font size setting now visibly scales UI text
Root cause: the original CSS override only changed :root{font-size} which
has no effect on the 232+ hardcoded px values throughout style.css. Only
the ~49 em/rem values were affected, which are not the main visible text.
Fix: add explicit px overrides for the key UI surfaces under each
data-font-size attribute selector:
- .msg-body (chat messages) + headings, code, tables
- .session-item, .session-meta (sidebar session list)
- #msg (composer textarea)
- .file-item (workspace file tree)
The :root override is kept so em/rem cascade correctly, but the targeted
element overrides are what actually make the text visibly larger/smaller.
Also: 8 new regression tests lock in the targeted CSS rules so this
cannot silently regress again.
* fix: composer large font was no-op — bump to 18px (default is 16px)
---------
Co-authored-by: nesquena-hermes <nesquena-hermes@users.noreply.github.com>
* fix(ui): echo slash command input as user message in chat (#840)
Slash commands like /skills, /help, /status previously showed only the
assistant response with no user message above it — the conversation
appeared to start from nowhere.
Fix: executeCommand() now returns {noEcho:bool} instead of true/false
(returns null when no command matched). send() in messages.js pushes a
user message bubble before returning when noEcho is false.
Commands with noEcho:true are action-only and don't get echoed:
/clear, /new, /stop, /retry, /undo, /voice, /model, /workspace,
/theme, /usage, /reasoning.
Commands without noEcho (get echoed):
/help, /skills, /status, /title, /compress, /compact, /personality.
16 new tests in test_issue840_slash_echo.py.
* fix(ui): push user message BEFORE running slash handler (ordering bug)
The PR as originally written pushed the user message AFTER the slash
command handler ran. That works correctly for async handlers (the
assistant response lands later, after the user push) but breaks for
sync handlers like cmdHelp which push their assistant response
synchronously:
S.messages = [assistant response, user "/help"] ← reverse order
The chat would render the help content ABOVE the user's own "/help"
input — not what the issue asked for.
Fix: look up the command inline, push the user message first (for
echo-worthy commands), then run the handler. If the handler opts out
(returns false — e.g. /reasoning <level>), pop the user message back
off so the normal send path can add it cleanly when forwarding to the
agent.
Renamed the flow so it's clear we're not calling executeCommand twice
(my first attempt did that by accident). executeCommand() stays as a
public API returning null or {noEcho:bool} — just isn't the only path
send() uses now.
Added 2 regression tests:
- test_send_pushes_user_message_before_running_handler: asserts
the user push appears before the handler invocation in source order.
- test_send_rolls_back_user_push_on_handler_optout: asserts the
S.messages.pop() for the opt-out case.
Also tightened the existing `test_send_checks_noecho_flag` and
`test_send_pushes_user_message_for_echo_commands` tests to look at
the new `_cmd.noEcho` pattern inline (vs the original
`cmdResult.noEcho`). Removed `test_send_uses_null_check_not_truthy`
(obsoleted — the control flow no longer stores the executeCommand
return in a variable).
Full suite: 1767 passed, 0 failures.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(ui): compress/compact noEcho + title/personality confirmation messages
Applied Opus mentor review fixes:
- compress and compact: add noEcho:true (S.messages reset internally causes
user bubble to flicker/disappear without noEcho)
- /title <name>: push assistant confirmation message after rename succeeds
- /personality <name>: push assistant confirmation message after set succeeds
- 4 new regression tests covering the above invariants
---------
Co-authored-by: nesquena-hermes <nesquena-hermes@users.noreply.github.com>
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
