_process_pdf prepends "\n\n[PDF content]:" to extracted text, and two
call sites in document_routes.py stripped it with .lstrip("\n[PDF content]:").
str.lstrip(chars) treats its argument as a *set of characters*, so it keeps
eating into the page text that follows the marker — e.g. a body starting
with "to the board" loses its leading "to" because 't'/'o' are in the
marker's character set. Replace both sites with a shared
strip_pdf_content_marker() helper that uses str.removeprefix.
Deep Research surfaced 'Error: unknown error' whenever every search
provider returned an empty result set without raising (e.g. SearXNG is
reachable but all its engines fail internally). _last_search_error was
only set on exceptions, so the empty-but-no-exception path left it unset
and the caller fell back to 'unknown error'.
Record an actionable reason on that path naming the providers that were
tried, so users can tell it's a search-backend problem rather than a
model problem. The provider-raised path is unchanged.
Re: #344.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
KNOWN_CONTEXT_WINDOWS lists 'o1' (200k) before 'o1-mini' (128k), and
_lookup_known returned on the first substring hit — so "o1-mini" matched
'o1' and reported 200000 instead of 128000. Track the longest matching
key instead, so the most specific entry wins regardless of table order.
`core.middleware.require_admin` grants admin to any request whose
`request.state.current_user == "internal-tool"` — the sentinel meant only
for the in-process tool-loopback path. But the normal cookie auth path
(app.py) sets `current_user` to the raw username, and neither `create_user`
nor the signup route reserved that name. As a result an account literally
named "internal-tool" was silently treated as admin by every
`require_admin`-gated route. With self-service signup enabled this is an
anonymous -> admin privilege escalation.
Reserve the full synthetic-owner set the codebase already special-cases —
"internal-tool", "api", "demo", "system" (see `_SYNTHETIC_OWNERS` in
routes/assistant_routes.py and the matching guards in src/task_scheduler.py
and routes/research_routes.py). "api" collides with the bearer-token owner
sentinel; "demo"/"system" would leave a real account denied an assistant
and inconsistently owner-scoped.
Refuse to create or rename into any reserved name (case/space-normalized),
and reject empty usernames while we're here. Adds a regression test.
Co-authored-by: Claude <noreply@anthropic.com>
`mdToHtml` deliberately stashes literal <details> blocks and <a> tags from
the source text *before* the global HTML-escape pass and restores them
verbatim into the string callers assign to `innerHTML` (e.g. chatRenderer's
`b.innerHTML = ...processWithThinking(text)`). Nothing scrubbed those
fragments, so message/agent content containing
`<details><img src=x onerror=...></details>` or
`<a href="javascript:..." onmouseover=...>` executed arbitrary script in
the authenticated page.
Route both stashed fragments through `sanitizeAllowedHtml()`, which parses
them in an inert <template> (no resource loads, no script execution),
removes script-capable elements, and strips event-handler attributes plus
javascript:/vbscript:/data: URL schemes. Hardening details:
- Compare tag names case-insensitively and drop the SVG/MathML foreign-
content roots. An SVG-namespaced <script> has the lower-case tagName
'script', so an HTML-only upper-case check would miss it — a real bypass.
- Sanitize to a fixpoint (re-parse + re-clean until stable) to blunt
mutation-XSS, where re-serializing/re-parsing reshapes the tree.
Benign anchors and <details> blocks are preserved unchanged.
Verified under jsdom against the obvious vectors plus mutation-XSS probes
(svg/math-namespaced <script>, foreignObject, ns-confusion, comment
breakout, template smuggling): no script/iframe element, event handler, or
javascript:/data: URL survives, and benign markup is kept.
Co-authored-by: Claude <noreply@anthropic.com>
The synchronous llm_call() runs in FastAPI's threadpool (sync route
handlers such as POST /sessions/auto-sort), while llm_call_async() runs
on the event loop. Both mutate the module-level _response_cache,
_host_fails and _dead_hosts dicts, so these are touched from multiple OS
threads concurrently. Two races result:
- _set_cached_response() snapshots 64 keys then deletes them with
`del _response_cache[key]`; if another thread evicts the same key
first, the del raises KeyError mid-eviction. Switched to
pop(key, None).
- _mark_host_dead() does get()+1+set() on _host_fails with no lock, so
concurrent connect failures lose increments and a genuinely dead host
can stay under its cooldown threshold. Guarded the host-health maps
with a threading.Lock (also applied to _is_host_dead / _clear_host_dead
for consistent reads).
Adds tests/test_llm_core_concurrency.py with deterministic regression
tests (phantom snapshot key for the eviction race; a slow-read dict that
forces the lost-update window for the counter). Both fail on the
unpatched code and pass with the fix.
The email reader folds quoted history into <details> summaries via
`_foldSummary()` (static/js/emailLibrary/signatureFold.js), which builds a
sender/date "meta" chip into the summary HTML and assigns it to innerHTML.
The server-side thread parser (`_extract_quote_meta`,
src/email_thread_parser.py) strips tags but then un-escapes HTML entities
and preserves `<...>` patterns, and that raw meta reaches `_foldSummary`
unescaped via `_renderTurnsFromServer` (`t.meta`) — so an inbound email
whose quoted attribution contains `From: <img src=x onerror=...>`
runs script when the victim merely opens the message (stored XSS).
Make `_foldSummary` the single escaping chokepoint: escape `primary` and
`subMeta` with the module's existing `_esc`. The client-side
`_extractQuoteMeta` previously pre-escaped its output, and every consumer
of it routes through `_foldSummary`, so drop that now-redundant escaping to
avoid double-encoding (e.g. "Ben & Jerry" -> "Ben &amp; Jerry").
Verified (jsdom): server-raw and client-extracted malicious metas yield 0
live elements and 0 event-handler attributes; benign "Ben & Jerry" renders
single-escaped.
Co-authored-by: Claude <noreply@anthropic.com>
SurprisedDuck