`passthrough_anthropic_messages` in `main.py` serializes the full request payload with stdlib `json` before sending it upstream. It is the largest single serialization on that path, since the body carries the whole conversation.
It now goes through `JSONCodec`, which selects orjson when `ENABLE_ORJSON` is set. The result is passed to aiohttp as `data=`, which encodes `str` as UTF-8 and sets `Content-Length` from the encoded bytes. The payload originates from a parsed request dict, so it holds only JSON-native types, and the serialized string is never hashed, compared or persisted.
The remaining stdlib `json` calls in this module are left alone: two are a debug log line and a fixed Ollama unload payload, and one parses an upstream error body.
With `ENABLE_ORJSON` unset, which is the default, `JSONCodec` is stdlib `json` and this call site behaves exactly as before.
`request.app.state.MODELS` is a `RedisDict` when Redis is configured. Unpacking it with `{**pool}` makes Python call `keys()` and then `__getitem__` once per key, which is one HKEYS plus one HGET per model, issued sequentially through a synchronous client. At 200 models that is 201 blocking Redis round trips per call.
`RedisDict.items()` is a single HGETALL, so `dict(pool.items())` fetches the same data in one round trip. `utils/chat.py:184` already does exactly this and carries a comment explaining why; these ten call sites were missed.
They are on the direct-connection branch of the task endpoints (title, tags, follow-up, autocomplete, query generation and the rest), of `chat_completed`, and of context compaction, so they run for background tasks fired on ordinary chat turns.
Behaviour is unchanged. The merged mapping is identical, the explicitly added direct model still overrides any pool entry with the same id, and when Redis is not configured the pool is a plain dict where `dict(d.items())` and `{**d}` are equivalent.
It also closes a race. `RedisDict.set` writes with HSET and then HDELs the stale keys, so a key returned by HKEYS could be deleted before its HGET arrived, raising `KeyError` out of the dict literal and failing the request mid model refresh. The old path could likewise observe a mix of pre- and post-refresh entries. HGETALL is atomic, so the caller now always sees one coherent snapshot.
Fourteen modules import `json` without using it. Ruff flags every one with F401, and a word-boundary search for `json` in each file matches only the import line itself, including inside strings, comments and annotations.
Two exclusions, both deliberate. Migration files are left alone: the import is equally dead there, but those files are frozen history and not worth the churn. `models/chats.py` has the same dead import and is handled in its own change, so it is skipped here to avoid two changes touching the same line.
No behaviour change.
orjson emits U+2028, U+2029 and U+0085 raw, where stdlib `json.dumps` escapes them under its default `ensure_ascii=True`. Python treats all three as line boundaries, so with `ENABLE_ORJSON` set, one of them inside model output splits a `data: {...}` SSE frame in half. Both halves then fail to parse and the delta is dropped with no error.
`utils/middleware.py` reassembles frames with `splitlines()`, so an affected response silently loses content on the direct API path. External clients are exposed as well: httpx's `LineDecoder` reimplements the same line-boundary semantics, so any SDK reading the OpenAI-compatible stream through `aiter_lines` breaks on a raw separator.
The three characters are escaped on the way out of `ORJSONCodec.dumps`. That restores parity with stdlib and fixes every reader at once, rather than patching one consumer and leaving external clients broken. They are the complete set: of the ten code points `splitlines()` treats as boundaries, the other seven are below U+0020, where JSON already forces an escape.
The membership guard is load bearing. Calling `translate` unconditionally costs roughly 1.5 us on a typical SSE chunk against 0.115 us for the serialization it wraps, so it would spend more than orjson saves. The three scans cost about 0.04 us.
Payloads containing none of the three are returned unchanged, byte for byte. With `ENABLE_ORJSON` unset, which is the default, none of this code runs.
U+2028 and U+2029 are common in text extracted from PDFs and word processor documents, so the realistic trigger is a model quoting an uploaded file back to the user.
The Valkey backend serializes chunk metadata on every insert and parses it back on every result row in `get` and `query`. Both directions now go through `JSONCodec`, which selects orjson when `ENABLE_ORJSON` is set.
The stored `metadata_json` field is never matched against as text. `_build_filter_expression` only emits TAG predicates, and the TAG fields are `id`, `hash`, `file_id`, `source` and `knowledge_base_id`; `metadata_json` appears only as a return field that is immediately re-parsed. So rows written with escaped non-ASCII and rows written raw are indistinguishable to every reader, and no migration is needed.
`process_metadata` already stringifies datetimes and strips null bytes and lone surrogates before the write, so the two backends cannot disagree about what is serializable here.
Both read `except` clauses widen from `(json.JSONDecodeError, TypeError)` to `(ValueError, TypeError)`. The codec falls back to engineio's codec, which installs `parse_int=_safe_int` and raises a bare `ValueError` for integer literals longer than 100 characters; the narrower clause would have let that escape and abort a search instead of yielding empty metadata. `json.JSONDecodeError` is a `ValueError` subclass, so this is a strict superset. That removes the module's last use of stdlib `json`, so the import goes with it.
With `ENABLE_ORJSON` unset, which is the default, `JSONCodec` is stdlib `json` and this call site behaves exactly as before.
RAG vector search runs in a thread pool, but then calls `future.result()` on the event loop thread, so the whole worker freezes until every collection answers. Every other user's token stream stops for that long. It's the default retrieval path.
Now `asyncio.gather` over `asyncio.to_thread`, matching what `routers/retrieval.py:2779` already does for the same call.
Measured with 3 queries across 4 collections, 60 ms search, and a second request wanting a turn every 5 ms:
| | before | after |
|---|---|---|
| RAG call | 62.0 ms | 61.2 ms |
| other request's turns | 0 | 7 |
| its worst stall | 62.5 ms | 16.0 ms |
Same results, same order, same `(result, error)` contract. Cancellation now lands mid-search instead of after every thread finishes. Threads move from an unbounded per-call pool to the loop's bounded shared one.
Swap the JSON encoder/decoder used across the backend from stdlib json to
orjson when ENABLE_ORJSON is set — HTTP request bodies, JSONResponse
bodies, upstream provider responses, SSE chunks, and socket.io/Redis
payloads.
The flag defaults to off, in which case the app uses stdlib json and
engineio's codec verbatim, so default behaviour is unchanged.
- json_codec exports JSONCodec (stdlib json or the orjson codec) and
SOCKETIO_JSON (engineio's codec or the orjson codec); call sites import
JSONCodec and stay implementation-agnostic
- apply_orjson_http_json() is a no-op when the flag is off, leaving
starlette's Request.json / JSONResponse.render untouched
- the orjson codec falls back to the stdlib for inputs orjson rejects
(non-str dict keys, ints beyond 64 bits, NaN literals)
- orjson is imported only when the flag is on
- FastAPI(default_response_class=...) is deliberately not used: an
explicit default disables the Pydantic direct-to-bytes fast path for
response_model routes
build_matcher compiled a caller-supplied pattern with Python's backtracking re and ran it over every line of every reachable file, with no timeout, no thread offload and no length caps. is_regex_pattern promotes any pattern containing a metacharacter, and a bare pipe counts, so no explicit regex flag is needed to reach the compiler. The search loop is synchronous inside an async handler, and UVICORN_WORKERS defaults to 1, so the cost lands on every other user of the instance. MAX_GREP_RESULTS bounds how many matches are reported, not how much work is done.
Backtracking cost is exponential in the length of the text being matched, so capping the pattern or the line does not bound it: the subject in the measurements below is 30 characters. `(x|x)*y` against a line of 30 x took 80 seconds, `(a+)+$` against 32 a took 169 seconds, and the same subject with a literal pattern took 0.6 microseconds.
Matching now runs on the regex module, which accepts a per-search timeout that re has no equivalent for. The timeout is the actual bound: regex resolves many classic catastrophic patterns instantly, but not all of them, and `(a|aa)+$` and `(?:a|a)*$` still need it. The budget covers a whole tool call rather than a single search, because a pipeline builds one matcher per segment and a per-search budget would multiply by segment count, and because a per-line timeout would allow timeout multiplied by line count. It is carried in a context variable so one command shares it without threading a parameter through every handler, and it is charged only for time spent inside search(), so database round-trips and other coroutines cannot consume it. Exhausting it raises, and both entry points already render that as an error for the model to read.
Note for anyone tracking search behaviour: re and the regex module define \w, \W and \b differently on non-ASCII text. re follows str.isalnum(), the regex module follows UTS#18, so \w no longer matches superscripts and fractions such as the ones in Nd-adjacent categories, and now does match combining marks. POSIX classes like [[:alpha:]] are interpreted rather than read as a literal set, and \p{...} compiles instead of erroring. Results on ASCII content are unchanged.
regex was already installed as a transitive dependency of nltk, tiktoken and transformers. It is now declared directly, pinned in pyproject.toml and requirements.txt to the version the lockfile already resolves.
* fix: fetch terminal system prompt per request with TTL cache
The system prompt was only fetched once in set_terminal_servers (startup
or connection save) with a 3s timeout, using a synthetic 'system' user.
That snapshot silently stays empty when the fetch races a cold-started
orchestrator instance, and goes stale when instances are reprovisioned
with a changed OPEN_TERMINAL_SYSTEM_PROMPT — recovering only after a
restart or a manual connection re-save.
- Fetch /system during get_terminal_tools with the user's own
credentials via a central TTL-cached method (5 min per server+user;
failures cached 60s so a dead instance doesn't stall every request),
falling back to the cached snapshot.
- Raise the fetch timeout from 3s to 30s so cold-provisioned instances
can answer.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KjnQJNKozp47vTB13pRyYs
* refac: fetch the terminal system prompt per request without a cache
Drop the module-level TTL cache and fetch the system prompt directly in
get_terminal_tools, gathered with the existing uncached per-request cwd
fetch that already follows this pattern. The fetch uses the user's own
credentials and falls back to the set_terminal_servers snapshot, so a
cold or unreachable instance degrades to the previous behaviour instead
of needing an error cache. Also restore the 3s timeout: on the request
path a 30s wait would stall chat completions, and a cold instance is
covered by the snapshot fallback until it warms up.
---------
Co-authored-by: Claude <noreply@anthropic.com>
Classic298