Hue Python SDK

For frozen datasets, local experiments, custom scorers, durable retries and historical rescoring, see Local evaluations.

Python helpers around the official OpenTelemetry trace and log SDKs and OTLP HTTP/protobuf exporters. opentelemetry-api, opentelemetry-sdk and opentelemetry-exporter-otlp-proto-http are accepted as >=1.40,<2; 1.44.0 is the certified lockfile combination and 1.40.0 is tested as the floor in CI. Provider requests run in your application. This package does not proxy model calls or configure global OTel providers.

The distribution is named hue-run on PyPI (import hue_sdk); install hue-run, not hue-sdk, which is not published there. Python 3.10+ is supported by the package contract; recorded validation below identifies the tested runtime.

Documentation · Sign in

Install

pip install hue-run
# Or, in a uv project:
uv add hue-run

Send a trace

In the consuming application, use only public imports:

import os
from hue_sdk import Hue

with Hue(
    api_key=os.environ["HUE_API_KEY"],
    capture_content=True,  # Also works with older releases; False sends metadata only.
) as hue:
    project = hue.validate_project()
    with hue.context(session_id="conversation-42", user_id="observed-user-7"):
        with hue.span("agent.run") as run:
            run.set_input({"question": "What is 2 + 2?"})
            with hue.tool("add") as tool:
                tool.set_input({"a": 2, "b": 2})
                tool.set_output(4)
            run.set_output({"answer": 4})
    if not hue.force_flush():
        raise RuntimeError("Telemetry export failed; inspect Hue export_status.")

capture_content=True, as above, is the recommended choice: trace inspection, evaluations and judges read the inputs, outputs and tool arguments and results it records. Choose False for metadata only when a policy forbids sending that content; see Content and semantic fields.

The SDK uses https://app.hue.run by default. Set base_url only for a different Hue deployment or a local receiver, using an origin without an API suffix. Existing Hue(base_url, api_key, ...) calls remain supported; a bare key passed as the first positional argument raises TypeError pointing at api_key=. The project service key determines the project. The SDK validates the project at GET /api/v1/projects/current when explicitly requested; construction itself does not perform a request. HTTP is permitted only for localhost and loopback IPs. Userinfo, query strings, fragments, paths and redirects are rejected. The key is sent only as Authorization: Bearer …; repr(hue), SDK errors and status counters omit it.

Content and semantic fields

Starting with Python 0.8.0, capture_content defaults to True. Earlier releases require an explicit choice. Set it to True for full traces: prompts and messages, responses and tool inputs/outputs, alongside model, usage, timing and errors. Trace inspection, evaluations and judges read that content, and your redactor and the credential filtering below apply to it. Choose False, metadata-only mode, when your users decline or an application or data policy forbids sending that content to another service.

False makes set_input and set_output omit content before it reaches an OTel queue and makes log_inference emit no record. Explicit JSON null, empty strings and absent content stay distinct when capture is enabled. Exception recording includes the exception type and ERROR status; exception messages and stacks are always excluded by these helpers.

When capture is disabled Hue also strips recognized GenAI, OpenInference, OpenLLMetry, Langfuse and Vercel AI SDK content attributes (including OpenInference retrieval documents, embeddings, reranker documents, prompt-template variables and images, and Langfuse observation and trace inputs, outputs, status messages and model parameters), legacy gen_ai.* message events, log bodies and status descriptions from every record it exports, including spans produced by third-party instrumentors on the same provider. COMPATIBILITY.md lists the exact keys. This setting is still not a blanket PII filter: custom attribute names, span names, session/user identifiers and resource attributes cannot be classified automatically and remain under your control, and other exporters keep their own policy. The server stores received content; there is no automatic telemetry expiry. Delete scoped data explicitly when required by your retention policy.

Hosted tools carry credentials in their definitions, such as the authorization and headers of an OpenAI hosted MCP tool. When content is captured, Hue's export path replaces credential-like fields including authorization, authorization_token, headers, api_key, access_token, x-api-key, and keys ending in token, secret, password, apikey or credential (case-insensitively, ignoring - and _) with "[redacted]" in recorded tool definitions (gen_ai.tool.definitions, ai.prompt.tools, llm.tools.*.tool.json_schema) and in the top-level tools and mcp_servers entries of any JSON object recorded as input.value, output.value or llm.invocation_parameters (as OpenInference records a raw provider request), whether or not it is a provider request; nothing else in that object changes. Parameters named in a JSON Schema properties object keep their schemas, so a tool that takes a headers argument is still described. A definition nested more than 256 levels deep drops its record. Sensitive defaults, constants, examples and enum values under credential-named schema parameters are redacted too. A url or server_url field (named like the credential fields, so serverUrl too) that has userinfo, a query or a fragment loses its userinfo and fragment, and each query value becomes [redacted] under its original name; a URL with none of them is exported as recorded. A scrubbed http, https, ws, wss or ftp URL is serialized as WHATWG URL does (lowercase scheme and host, IDN hosts in Punycode, default ports dropped, dot segments resolved), so for ordinary hosts its text and the tool-catalog digest match the TypeScript SDK's. Internationalized hosts can still differ at the edges: Python refuses an IDN host longer than 1,024 characters and a zero-width joiner or non-joiner anywhere but after a virama, and checks right-to-left labels as the Node.js parser does, which accepts a few that Bun refuses. Other schemes, such as mcp: or file:, keep their earlier handling and can differ from the TypeScript SDK. A value that is not an absolute URL becomes "[redacted]".

With capture_content=False, export removes tool definitions but keeps a summary of them on the same record: hue.tool.names lists each definition's name (Chat Completions function.name, or the type of an unnamed built-in tool such as mcp) in order, and hue.tool.definitions.sha256 is the lowercase hex SHA-256 of the RFC 8785 canonical JSON of the credential-scrubbed definition list, identical to the TypeScript SDK's and unchanged when a credential rotates. Only definitions an instrumentor recorded can be summarized, for example OpenInference without hide_inputs. Recorded content can carry images, documents and audio: see Images, documents and audio for what the SDK uploads and how.

Use redactor=lambda field, value: ... to transform content in supported helpers. It runs synchronously before serialization and export. Return a redacted JSON value; failures omit the field and increment export_status.instrumentation_failures without changing application behavior. It does not inspect arbitrary OTel attributes or logs:

def redact(field, value):
    if isinstance(value, dict):
        return {key: "[redacted]" if key == "email" else item for key, item in value.items()}
    return value

The callback should cover your actual nested input format; this small example is only a top-level dictionary transformation.

Before redaction, helpers copy supported content into detached built-in containers; in-place changes by a redactor cannot change application inputs or results. The input and the redactor's returned value each have a 4 MiB conservative value budget, 64 maximum nesting depth and 65,536 visited values/keys, so a redactor sees content past the value cap and can recognize a secret that crosses it. While values over the cap can be uploaded (see values over the inline limit), a span attribute's content (set_input, set_output) has a 1 GB budget and 1,048,576 values instead and is serialized whole: the redactor sees the whole value, and its answer is what is uploaded. Integers and integer keys are limited to 14,000 bits before decimal conversion. Final serialized content has the 1 MiB UTF-8 JSON value cap (or the value limit Hue's receiver advertises): a value over it is cut to a UTF-8 prefix of its JSON text, as Hue's receiver cuts a value over its own cap, and its key is listed in the span's hue.truncated attribute, so the span still carries the call and the recorded part of its value. Cyclic, nonfinite, unsupported or over-budget content is omitted and counted as an instrumentation failure. See the Python safety boundary for supported types and callback limits.

To store a file your work reads or produces, see Record a file.

Tool/model metadata accepts plain strings; invalid values use stable fallback labels and increment instrumentation failures without invoking custom conversion hooks. Disabled tracing skips metadata validation.

Hue helpers implement the model/tool metadata and inference-message subset of the Development GenAI conventions. They accept caller-supplied JSON without full schema validation. For model helpers, pass the message representation produced by your integration. When authoring your own messages, prefer the shape defined by the GenAI semantic conventions' input messages and output messages JSON schemas. Use either span content or correlated logs for a given input/output, avoiding duplicate copies. Call log_inference inside model(..., provider=...), or pass operation= and provider= explicitly, to include the required inference metadata. log_inference sends a structured body (an explicit None field keeps its key with an empty value, distinct from an absent field) and copies gen_ai.operation.name, gen_ai.provider.name and gen_ai.request.model from the enclosing model() block, or from its operation=, provider= and model= keywords, plus gen_ai.conversation.id from context(), onto the record's attributes when known. A blank or over-long keyword is omitted and counted as an instrumentation failure.

Images, documents and audio

Hue stores the bytes of images, documents and audio apart from the spans and log records that mention them, under the record's trace and named by SHA-256. The record keeps the file's digest in its place and lists the upload in its hue.blobs attribute. Files reach Hue in four ways:

client.media reads a recorded file back. Every upload needs capture_content=True and a project key (never a setup key, hue_setup_…) and follows the redactor: nothing it removes is uploaded, a value it rewrites uploads as rewritten, and an attachment part it rewrites uploads nothing. A record is sent without waiting for its uploads; they run on four background threads within the 128 MiB held for values awaiting upload, and an upload refused before export falls back without failing the export; a file used more than once in a trace is stored once. See Values over the inline limit for the upload protocol, budgets and fallbacks.

Automatically captured media

With capture_content=True and a project key, the SDK finds the files inline in the attributes Hue reads files from, uploads each file's bytes under the record's trace and leaves the file's digest in its place. On a span's own attributes:

• Messages and values: gen_ai.input.messages, gen_ai.output.messages, ai.prompt.messages, the AI SDK's ai.prompt, input.value and output.value (what span() and OpenInference record), traceloop.entity.input and traceloop.entity.output, and Langfuse's langfuse.observation.input, langfuse.observation.output, langfuse.trace.input and langfuse.trace.output.
• Tool arguments and results: gen_ai.tool.call.arguments, gen_ai.tool.call.result, ai.toolCall.args, ai.toolCall.result, tool.parameters, tool_arguments and tool_response.
• Metadata: metadata, langfuse.observation.metadata and langfuse.trace.metadata (whole, or one .<key> per attribute) and ai.telemetry.metadata.<key>. A metadata value is read as JSON when it is an object or array; otherwise it is a file only when the whole value is a data: URL.
• Indexed messages: OpenLLMetry's gen_ai.prompt.<n>.content and gen_ai.completion.<n>.content when the value is the JSON of a parts list (any other content stays text), and OpenInference's flattened llm.input_messages.<n>.message.contents.<m>.message_content.image.image.url (and llm.output_messages.…), whose value is the image's URL.

Inside those values a file is found by its shape: GenAI blob, AI SDK file and image, OpenAI Chat image, file and audio parts and an assistant message's audio, Responses input_image, input_file and computer_screenshot, Anthropic base64 sources, Gemini inlineData and inline_data, Vertex media, Bedrock image and document sources, and any whole string that is a data: URL declaring a media type. A tool call's arguments and a tool result (a tool message's content, tool_result, function_call_output, tool_call_response) are JSON in a string by convention, so the files inside them are uploaded too; a message's own text is never parsed as JSON. Hue's reader looks for files no deeper than ten levels below a value's root (an encoded string's JSON counting as a level), and the SDK inspects no deeper. Hue keeps attribute JSON nested past 32 levels, or holding an integer past 2^53, as text, and an OpenInference input.value or output.value whose record declares input.mime_type or output.mime_type as text/plain is literal text: the SDK leaves such values as recorded. Remote URLs, provider file IDs, S3 references and Langfuse media tokens stay unchanged, and Hue never fetches them. To copy a Langfuse project's traces and the files its media tokens point to into Hue, run the TypeScript SDK's hue import langfuse command (with npx, for an application in any language); the Python SDK has no counterpart.

Each uploaded file is listed in the record's hue.blobs, keyed <attribute>#<JSON pointer>:

• A provider part loses its byte field and carries sha256 and size; the key names the part, for example gen_ai.input.messages#/0/parts/1.
• An OpenAI assistant message's audio loses audio.data and the message carries sha256 and size (gen_ai.output.messages#/0). Its media type comes from audio.format, else from the bytes' WAV, MP3, AAC, FLAC or Ogg signature, else application/octet-stream.
• Raw 16-bit PCM keeps the format its type declares: audio/pcm, audio/x-pcm or audio/L16 with a rate from 8000 to 192000 and 1 or 2 channels is listed as audio/pcm;rate=24000;channels=1 (audio/L16;… for big-endian samples), and OpenAI pcm16 audio as audio/pcm;rate=24000;channels=1, so Hue plays the stored file. Other types are listed without their parameters.
• A whole data: URL string becomes a GenAI blob part, {"type":"blob","mime_type":"image/png","modality":"image","sha256":"…","size":68}, keyed at the string (input.value#/image). A value that is entirely one, such as a metadata attribute or a string input or output recorded as it is, becomes that part's JSON text, keyed <attribute>#: an attribute read as JSON whose text is not JSON is read as that text, as Hue reads it.
• A file inside a JSON-encoded tool string takes the digest form inside the string. Its key goes on through the string, which no JSON pointer can address, so Hue matches it by its digest.
• Text media over 64 KiB of decoded bytes in a provider part keeps a UTF-8-safe 16 KiB prefix in its content field instead, and its key names that field (gen_ai.input.messages#/0/parts/1/content). Smaller text and empty fields stay inline. A text file a GenAI blob, Vertex media or AI SDK part holds as bare base64 takes the digest form.

Only the file's own value, or the members of the part that held it, change: every other byte of the recorded text, its key order, number spellings, escapes and spacing included, is kept, in an encoded string as in the attribute itself, so both SDKs write the same bytes. A record has 64 hue.blobs places: the media inline in its values take them first, in reading order, then its attachments and the files recorded with upload.

Entries use the recorded media type and, when valid, the filename, name or title. A valid name is nonblank after trimming, at most 256 UTF-16 code units and well-formed Unicode, with no C0 characters (U+0000–U+001F) or DEL (U+007F). Invalid names are omitted from the entry without rewriting the original field. Data URLs and fields explicitly encoded as base64 by the provider contract are decoded before hashing. Other content is read as Hue's reader reads it, whatever its media type: content that reads as base64 (a length divisible by four, base64 characters only, as the GenAI conventions and client.media.resolve write a blob or media part's content) is the file's base64 and is decoded, so YWJj hashes as abc; anything else is the file's own UTF-8 text. The digest matches span.record_file's digest for the same bytes.

A file whose upload is refused before its record is exported (the receiver is known to lack the upload route, or a budget ran out) keeps its digest without an entry in the recorded messages (gen_ai.input.messages, gen_ai.output.messages, ai.prompt.messages); in every other attribute it keeps its bytes inline under the ordinary value limit. A file whose upload fails after its record was exported keeps the digest form and the entry the record was exported with, which Hue reads as missing.

Log records upload their media the same way when they are correlated with a span: log_inference's gen_ai.client.inference.operation.details record and any gen_ai.* event with a structured body. The bytes upload under the log record's own trace and the log record lists them in its own hue.blobs, keyed body#<JSON pointer> into the body (body#/gen_ai.input.messages/0/parts/1, or body#/content/1 for a deprecated gen_ai.user.message event) or <attribute>#<JSON pointer> for the attributes above. One step places a log record's files, the attachments in its body included. A body recorded as JSON text stays as it is, and no log value is uploaded whole. A log file whose upload is refused before export is placed as log records were before they uploaded: a file over 64 KiB in a recorded message attribute becomes its digest and any other file keeps its bytes; text still over the value limit is cut inline and listed under hue.truncated. One whose upload fails after export keeps its digest and entry, as on a span. While uploads are available, log_inference records a field over the value limit when it fits the limit with its files as their digest.

Inspection is bounded to 8 Mi UTF-16 code units of text per attribute and 16 Mi per record, a log body's file contents included. Within those bounds, inspection parses the value and decodes each file while the record is queued and again when it is exported. A span's own attribute past either bound that is held whole for upload (it is over the value limit) has its media read in place when it is exported, as a voice agent's audio usually needs: each string in it longer than 16 Ki code units and written without escapes as strict base64 or a base64 data: URL is decoded, hashed and uploaded a 1 MiB window at a time from the recorded text, never parsed or copied out of it, and a file may be up to Hue's 1 GB upload limit. The scan that finds those strings also reads the value a window at a time, so no step of it holds the interpreter for long; a percent-escaped data: URL is parsed with the JSON around it. A long string no file replaces is written back without a second copy of it. The value is placed exactly as one within the bounds: the same digest form, entries and fallbacks. The JSON around those strings, a file written with escapes or inside a JSON-encoded string included, is parsed within the 8 Mi bound; a value past it uses the ordinary value limit (it uploads whole). Each such value is read within that bound on its own, apart from the record's 16 Mi allowance, while it stays held within the 128 MiB for values awaiting upload. Log records keep the bounds. Resources, links and span events never upload inline media: in their recorded messages, media up to 64 KiB stays verbatim and larger media is replaced by its digest, as in 0.9, and log records that cannot upload (not correlated with a span, or while uploads are unavailable) keep the same rule. With capture_content=False, nothing is inspected or uploaded.

Attach a file anywhere

Wrap a file in an Attachment and put it anywhere in a value Hue records: a span() input or output, model() messages, a tool's arguments or result, log_inference() messages, or a span attribute passed to span(attributes=...) or set_attribute.

from pathlib import Path

from hue_sdk import Attachment

png = Path("scan.png").read_bytes()
with hue.span(
    "review contract",
    attributes={"metadata": {"scan": Attachment(png, content_type="image/png")}},
) as span:
    span.set_input(
        {
            "question": "Which clauses changed?",
            "contract": Attachment(path="contract.pdf", content_type="application/pdf"),
        }
    )

Give exactly one of data (bytes, or a base64 string or data: URL) and path (a file on the file system), a content_type and optionally a name. data is read and hashed when the attachment is created and is limited to 25 MiB, the record_file limit: it is held in memory from then until it is uploaded, within the 128 MiB budget that queued spans share for values awaiting upload. A path names a regular file: anything else, such as a named pipe, is refused without waiting on it. A path up to 25 MiB is read the same way. A larger regular file, up to Hue's 1 GB upload limit, is not held: it is hashed a 1 MiB slice at a time when the attachment is created and read again from disk when it is uploaded, each slice hashed as it is sent and the last sent only once the file ended at the size and SHA-256 it was recorded with. Keep such a file in place and unchanged until the client flushes: one that changed, is gone or is no longer a regular file is not uploaded (the request ends before its body is complete, and Hue's upload URL signs the size and SHA-256, so the store refuses other bytes), and it counts in export_status.upload_fallbacks, its part keeping the digest and entry it was recorded with; another reference in the trace to the same bytes held in memory is uploaded instead.

At its position the attachment is recorded as a GenAI blob part that names the file by digest, {"type":"blob","mime_type":"application/pdf","modality":"document","sha256":"<64 hex>","size":48213,"name":"contract.pdf"}, with modality read from the content type (image, audio, video, document or file). The record lists the file under its hue.blobs when it is exported, and the bytes upload in the background, keyed <attribute>#<JSON pointer to the part>: input.value#/contract above, metadata#/scan for the attribute, and body#/gen_ai.input.messages/0/parts/1 for a part in a log_inference() log record, which uploads with the other media in the body. An attribute value that is an attachment, a dict or a list OpenTelemetry cannot hold (not of one primitive type), with or without attachments, is recorded as JSON text, as in the TypeScript SDK; put files under metadata for Hue to list them as the span's metadata files.

With capture_content=False or a setup key no bytes leave the process: a metadata attribute still records the part, without its name, and content values are not recorded at all. A part the redactor removes or rewrites is not uploaded. A part that cannot be uploaded keeps its digest (and its entry, when the upload failed after export) and is counted in export_status.upload_fallbacks, as is one nested more than 10 levels deep in its value, where Hue does not read parts, or one in a value cut to the value limit. An attachment in a value that is still over the value limit with its files as digest parts is uploaded and listed as usual, and client.media.fetch(trace_id=..., sha256=...) and Hue's span read reach it by digest; the value itself uploads whole and keeps only its first 16 KiB inline, so readers do not place the file at its position in the value. Creating an attachment never raises: one without exactly one of data and path, with unreadable or larger data or an invalid content type is recorded as None and counted as an instrumentation issue; an invalid name alone is omitted and counted.

Record a file

Record a file your work read, received or produced with span.record_file. It adds a hue.file event keyed by hue.file.sha256 to the span; pass upload=True to store its bytes in Hue as well:

from pathlib import Path

with hue.span("review contract") as span:
    span.record_file(
        role="input", media_type="application/pdf", data=Path("contract.pdf").read_bytes(),
        name="contract.pdf", upload=True,
    )

upload defaults to False. Supplied data must be at most 25 MiB, as it is held in memory until it is uploaded. Give path instead of data for a file on disk: up to 25 MiB it is read as data is, and a larger regular file, up to Hue's 1 GB upload limit, is hashed a slice at a time now and, with upload=True, read again from disk when it is uploaded, as an Attachment given by path is: it uploads only if it still has the size and SHA-256 it was recorded with. A file over 1 GB, a path that is no regular file, or data and path together, is omitted and counted. Upload requires supplied bytes or a path, content capture and an ordinary project key; the file is listed in the same span's hue.blobs, keyed event:hue.file:<sha256>, when the span is exported, while preserving the event, and uploaded in the background. An export fits every span's events and entries before it lists any of these files, and the background uploads them, four at a time; a file whose event was redacted or shed is not uploaded. Names must be nonblank after trimming, contain at most 256 UTF-16 code units, be well-formed Unicode and contain no C0 characters (U+0000–U+001F) or DEL (U+007F). Tabs, newlines, DEL and lone surrogates make a name invalid. An invalid name is omitted and counted as an instrumentation failure; the otherwise valid event remains. Larger data and invalid records are omitted and counted as instrumentation failures. Pass sha256 and byte_size when you already have the file's identity.

Read recorded media back

EvaluationClient.media reads a recorded file back by its digest, with a key that may read recorded content (Read or Read and write; a trace snapshot's copy needs Read and write):

import json
import os
from hue_sdk.evals import EvaluationClient

client = EvaluationClient(api_key=os.environ["HUE_API_KEY"])

# One file by digest; pass the span that recorded it when you know it.
file = client.media.fetch(trace_id=trace_id, span_id=span_id, sha256=sha256)
print(file.content_type, file.file_name, file.size, len(file.data))

# Every digest part of a recorded value, written back in its own field.
recorded = json.loads(attributes["gen_ai.input.messages"])
resolved = client.media.resolve(recorded, trace_id=trace_id, span_id=span_id)
messages, failures = resolved.value, resolved.failures

fetch asks Hue for a five-minute download link (GET /api/v1/traces/{traceId}/blobs/{sha256}), downloads it without the project key and refuses bytes whose SHA-256 is not the digest. It raises MediaFetchError, whose reason is not_found (Hue holds no such file for the trace), missing (the trace lists it, but it was never uploaded or has been erased), unavailable (still unavailable after one retry that waits out a Retry-After of up to 10 seconds), forbidden (the key cannot read recorded content, or reads a snapshot without Read and write), too_large (over max_bytes, 64 MiB by default, checked before downloading), digest_mismatch or failed (a connection failure, a timeout, a redirect or another unexpected answer). Hue is asked for the link once more after a connection failure or a timeout, never after an unexpected answer; a download ends once no byte has arrived for the client's timeout, at least two minutes, and is not retried. snapshot_id= instead of trace_id= reads a trace snapshot's pinned copy (GET /api/v1/trace-snapshots/{snapshotId}/blobs/{sha256}), which carries no media type or name.

resolve downloads each distinct digest once, four at a time (concurrency, up to 16), and returns a ResolvedMedia with a new value; the input is not changed. With as_="data_url" (the default) each part receives a data: URL where its field takes one (OpenAI image_url and file_data, Responses input_image, input_file and computer_screenshot, AI SDK image and file) and bare base64 where its convention requires it (GenAI blob, OpenAI input_audio and assistant audio, Anthropic source.data, Gemini inlineData, Vertex media, Bedrock source.bytes), so the messages can be sent to the provider again. A GenAI blob or Vertex media part receives base64 whatever its media type, as the GenAI conventions define its content and as Hue reads content that reads as base64; as_="bytes" writes bytes. A resolved part no longer carries sha256 and size. A part whose file cannot be read stays as it is and is listed in failures with its JSON pointer and reason. Parts inside JSON-encoded strings, such as tool-call arguments, and Langfuse media tokens, which name files Langfuse holds, are left as they are; fetch reads the former by the digest the part carries.

A snapshot's copies carry no media type, so a part that declares none would get a data: URL of application/octet-stream. Pass the record's hue.blobs entries in any form Hue reads (the JSON strings the attribute holds, dicts, or one string holding the JSON array), and such a part takes the type of the entry with its SHA-256. Text over 64 KiB that a record keeps as its first 16 KiB is never written back: a part whose content field an entry names, holding only that prefix, stays as it is and is listed in failures as prefix_only with the entry's SHA-256, and fetch(trace_id=..., sha256=...) returns the whole file. An entry names a pointer, not a value, so the entries of another attribute with a file at the same pointer are listed too; pass only the entries of the value's own attribute to list its own file.

resolved = client.media.resolve(recorded, snapshot_id=snapshot_id, blobs=attributes["hue.blobs"])

Files in evaluation cases

add_case and add_eval_set_case pin up to 32 input files on a case with files=: artifacts the project already holds, or data and paths the client checks and uploads first. run_experiment downloads a case's pinned files before its execution starts and hands the agent-visible ones to the target as context.files, and rehydrate_media(inputs, files) writes them back into the digest parts of recorded messages copied into the case, so the target can send the recorded model call again with the real image, document or audio. See Case files.

Coming from Langfuse or Braintrust

Existing instrumentation

Pass an existing opentelemetry.sdk.trace.TracerProvider through tracer_provider=provider to add Hue's exporter, and an existing opentelemetry.sdk._logs.LoggerProvider through logger_provider= when the application already owns one. A provider Hue creates for the other signal reuses the borrowed provider's resource, so spans and correlated logs report the same service.name; service_name applies only when Hue creates both providers. Hue's processor then exports every span that ends on that provider, the same default as other OpenTelemetry exporters, plus placeholders for its running Hue and AI spans; wrap the processor if only part of the provider's spans should reach Hue, and forward on_start only for spans whose on_end the wrapper forwards unchanged. Session/user identifiers from hue.context() are stamped on Hue helper spans only. Hue does not call set_tracer_provider. It exposes hue.tracer_provider, hue.tracer and hue.logger_provider for explicit integration. shutdown() closes Hue's processors; borrowed providers and their other processors stay usable. Finish traced work before shutting Hue down: spans ending or external records emitted afterward increment Hue's dropped-record counters, including emissions through a borrowed provider. New Hue helpers after shutdown are no-ops. Do not repeatedly attach Hue clients to one long-lived provider: OTel has no public processor-removal API. Create one client per provider lifecycle.

An instrumentor that accepts tracer_provider can receive hue.tracer_provider; follow that instrumentor's own capture/redaction configuration. OpenInference and other OTel instrumentors are optional dependencies, not implicitly enabled. Hue's export path strips their recognized content attributes when capture_content is False, but configure their own capture controls as well: unrecognized custom keys pass through, and the instrumentor may still send content to other exporters. The optional compatibility group pins OpenAI 3.14.0, OpenInference OpenAI 0.1.60 and its resolved OpenInference instrumentation 0.1.63. A synthetic HTTP streaming response verifies parentage, canonical model/usage attributes and enabled/disabled message capture with TraceConfig(enable_genai_semconv=True, hide_inputs=..., hide_outputs=..., hide_input_messages=..., hide_output_messages=...). This is a tested adapter combination, not a claim about all OpenAI APIs or live-provider compatibility. See the instrumentor's official source. See the Python integration guide for application setup.

Local development without a Hue account

Hue speaks standard OTLP, so any local collector works. Point base_url at a loopback receiver that accepts /api/v1/otlp/v1/traces and /api/v1/otlp/v1/logs (for example an OpenTelemetry Collector otlp receiver with http.traces_url_path and logs_url_path set to those paths, forwarding to Jaeger or the debug exporter) and pass any placeholder api_key; HTTP is allowed for loopback origins. validate_project() and verify_trace() are Hue-only diagnostics and are not available against a generic collector. Set live_spans=False so Hue-specific placeholders do not reach the collector's other destinations.

Export behavior and limits

• Traces go to /api/v1/otlp/v1/traces; correlated logs go to /api/v1/otlp/v1/logs. Both use the official OTLP HTTP/protobuf exporter with gzip-compressed request bodies and a hue-sdk-python/<version> User-Agent ahead of the exporter's own token. The official exporter handles retryable network/service failures; Hue additionally honors 429 and Retry-After within its transport budget. A 429 whose Retry-After outlasts that budget keeps its records queued, within the queue's bounds, and sends them again once the wait has passed; an export holds its records this way for at most 60 seconds in all, and a longer wait loses them, recorded as a failed export issue. A flush or shutdown budget shorter than the hold returns False while the held export continues. There is no proprietary provider transport.
• Exporter work and its HTTP worker run with OpenTelemetry instrumentation suppressed. Hue ignores records emitted within that suppressed scope, preventing HTTP instrumentation and exporter diagnostics from feeding back into its own queues. Application instrumentation resumes outside that scope. If an OpenTelemetry release ever stops exposing its suppression key, Hue warns once at import and keeps exporting: its own queues still ignore its export work, but OTel HTTP instrumentors could then record Hue's export requests on other exporters.
• Requests are measured as Hue's receiver measures them: after gzip against its request limit and after decoding against its decoded limit. Until a response advertises its own limits (Hue-Max-Request-Bytes, Hue-Max-Decoded-Bytes and Hue-Max-Value-Bytes, on every acknowledgement and refusal) these are 1 MiB on the wire, 4 MiB decoded and 1 MiB per value; advertised limits are adopted, clamped to 1–64 MiB, 1–64 MiB and 256 KiB–16 MiB, and never exceeded: a request that limits lowered while it was retried or held no longer fit is split and shed again before it is sent. Batches contain at most 64 records, are packed to 4 MiB before gzip (or the decoded limit, when lower) and are split at record boundaries when they compress worse than the wire limit allows.
• A value over the value cap that is not uploaded (see values over the inline limit), in any record (helper content, a third-party instrumentation's attribute, a log body), is cut to a UTF-8 prefix before enqueue, bytes are replaced by the receiver's marker ({"hue.truncated": true, "hue.truncated_bytes": <size>}), and its key is listed under hue.truncated, never omitted for its size; an inference log's body field over the cap is still omitted and counted, since its prefix is no JSON, unless uploads are available and the field fits the cap with its files as their digest. Tool definitions and recorded requests have their credentials removed from the whole value before that cut. A record still over the receiver's limits sheds its content values, largest first, each replaced by the marker and listed, until it fits; only a record too large without any content value is lost, recorded as a dropped export issue. Records a trace lost (dropped from a queue, too large without their content, or refused for their rate past the hold) are counted on its root span as hue.sdk.dropped_records when the root is exported, so Hue reads the trace as incomplete by that many records.
• A tracer provider Hue creates keeps 2,000 attributes, events and links per span and attributes per event and link (OpenTelemetry's default is 128), unless the environment sets OpenTelemetry's own limit (OTEL_SPAN_ATTRIBUTE_COUNT_LIMIT, OTEL_ATTRIBUTE_COUNT_LIMIT, OTEL_SPAN_EVENT_COUNT_LIMIT, OTEL_SPAN_LINK_COUNT_LIMIT, OTEL_EVENT_ATTRIBUTE_COUNT_LIMIT, OTEL_LINK_ATTRIBUTE_COUNT_LIMIT), which wins. Log records and externally configured providers keep OpenTelemetry's limits. A span whose events or links do not all fit the queue's byte budget keeps its newest ones and counts the rest in its dropped event and link counts, rather than being dropped whole.
• Hue accepts at most 2,000 distinct spans per trace. This is enforced by the receiver across distributed producers; the client cannot guarantee a global count.
• HTTP errors, malformed/non-200 success responses and OTLP partial_success rejected counts cause a failed export status. Partial rejection is not retried wholesale. Receiver error text is not echoed. A warning-only partial-success response with zero rejected records remains successful. A request carrying only live-span placeholders never fails the status.
• force_flush(timeout_millis=30000) drains both processors and returns False for a timeout, failed export, dropped record or instrumentation omission since this client was created. Hue's processors honor the shared deadline and serialize drains of records in a bounded background worker; the caller then waits for pending uploads within the same deadline, outside that worker, so a concurrent flush never waits behind another's uploads to send its own records. Pending exports continue after timeout. export_status exposes cumulative failure counters. shutdown() stops new helpers, drains and closes owned exporters within the caller's wait budget; repeated calls wait for the same shutdown. After a timeout, keep the process alive and call shutdown again to confirm completion. Context-manager exit calls shutdown_safe(timeout_millis=1000); check flush explicitly when an exit code must reflect delivery failure.
• Hue processors use the public OTel interfaces and bound each signal to 2,048 records and 8 MiB of encoded telemetry by default, including in-flight records (max_queue_size, max_queue_bytes). Dropped records and instrumentation omissions make cumulative export_status.ok false. Queue overflow, process termination and sampling can lose telemetry. Flush success reports observed exporter outcomes, not durable local delivery or proof that every application operation was instrumented. The exporter timeout controls individual export/retry operations. A caller timeout does not cancel an HTTP request already in flight; background workers continue until the operation completes.
• A timed-out HTTP worker can retain one encoded request, up to the receiver's request limit, per signal outside the queue counters. Later records remain queued within the configured limits until that worker finishes; shutdown counts any queued records it must discard. Queue bytes are not total process memory.
• A flush timeout releases drain coordination so later flushes can make progress. Both signals share the caller's remaining wait budget, including time spent waiting for another flush. Shutdown's exporter cleanup can continue after the caller returns without holding that coordination lock. Repeated shutdown calls report current failures and drops as well as cleanup completion.

Values over the inline limit

Hue keeps a value inline up to its value limit (1 MiB unless the receiver advertises another in Hue-Max-Value-Bytes). A span's own string attribute larger than that (helper content such as input.value, output.value and tool results, or another instrumentation's attribute), and the files described in Images, documents and audio, are uploaded to Hue apart from their record, up to Hue's upload limit of 1,000,000,000 bytes:

1. Export computes the SHA-256 of the value's bytes (a string as UTF-8, a file as its own bytes) and lists the value; a background thread reserves it with POST /api/v1/otlp/blobs, authenticated with the project key. Hue answers that it already stores the value under the trace, and nothing is uploaded, or with a presigned URL.
2. The bytes go straight to the project's evidence store in one PUT to that URL, streamed with exactly the headers Hue signed and never the project key (over HTTPS, or plain HTTP to a loopback store), and the upload is confirmed with POST /api/v1/otlp/blobs/complete. A PUT the store answers with 412 (the same value uploaded at the same time by another span) counts as stored.
3. Whole string values and large text media keep a UTF-8-safe 16 KiB prefix. Binary media keeps its digest on the part. The record lists the values whose upload export accepted in hue.blobs, at most 64 compact JSON strings with fields ordered key, sha256, size, content_type and optional name: {"key":"output.value","sha256":"<64 hex>","size":5242880,"content_type":"application/json"}. A whole value is listed under hue.truncated as well, since its inline text is a prefix.

Hue stores the value under the trace, so the trace's retention, deletion and erasure cover it. Uploads run apart from export on at most four daemon threads per client, never on the application's thread or the export worker: a record is sent without waiting for its uploads, which never hold, pause or drop it, and an upload may finish before or after its record reaches Hue. Each reservation, PUT and confirmation is tried up to three times on no answer, 408, 429 or 5xx, waiting the answer's Retry-After (at most 60 seconds) or else a jittered 0–1 s, then 0–2 s; other refusals are final. An upload is counted, and lets go of its bytes and its share of the held-value budget, once its bytes are stored; a flush waits for its confirmation's first attempt only, and a retried confirmation holds no upload thread while it waits. Each request has export_timeout_seconds, and each PUT attempt six of them or one second per MiB, whichever is longer; a PUT that runs out of its own time is not tried again. At most 1,024 uploads are pending at once. force_flush(), and a tracer provider's force_flush(), sends the records it drains first and then waits, within its deadline, for the uploads accepted so far, and returns False at once when none of them can be stored before that deadline (each waits out a Retry-After past it), so a provider's processors after Hue's keep the rest of its budget; a flush called while another waits for uploads still sends its own records at once; shutdown() gives pending uploads up to 30 seconds more once its processors stop, then abandons the rest, counted as not uploaded. A processor's own shutdown, as a borrowed provider's shutdown() runs it, sends Hue's records without waiting for uploads, so the provider's other processors are never held behind them: call hue.shutdown() after it to give pending uploads that grace.

A value whose upload is refused before its record is exported (the receiver is known to lack the upload route, the value is over 1 GB, no hue.blobs place is left, or the held-value budget or the 1,024 pending uploads are full) falls back according to its form: whole attributes keep a prefix up to the cap and are listed under hue.truncated; media parts keep their digest without a blob entry; refused opt-in record_file uploads preserve the file event. An upload that fails after its record was exported (Hue or the store refused or failed it, or the network or a deadline did, after its attempts) leaves the record as it was exported: the 16 KiB prefix or digest and its entry, which Hue reads as missing; a later reference to the same bytes in the trace uploads them again, which resolves the earlier entry too. Bytes given in memory to record_file or an Attachment are limited to 25 MiB; a file given by path, a file read in place from a value and a whole attribute to 1 GB. Each cause is recorded in a warning export issue, whose count is the values and trace_ids their traces, and counted in export_status.upload_fallbacks; an upload is counted in export_status.uploaded_values once it is stored. Both counters include whole attributes, eligible binary or large text media, attachments, and opt-in record_file uploads. A warning is not a failure: it does not affect export_status.ok, force_flush() or export_failure_sequence(). A receiver without the upload route (a 404 that is not Hue's, a 405 or 501, or a 200 that is no reservation) is not asked again for the life of the client, so another OTLP receiver meets no retry storm: values queued after it answers keep their fallback form, and only records exported before it answers list values it never stores.

To upload a value whole, the queued copy of its span holds it apart from max_queue_bytes, within a 128 MiB budget of held values (each charged as the interpreter holds it); a recorded message holding eligible media is held the same way, while the queue is charged its copy with the files as their digest. A value the budget cannot hold, or one queued while the receiver is known to lack the upload route, is cut (its files digested) when it is queued, as before, and reported as not uploaded, once per file; a single value larger than the whole budget is held while nothing else is. A value that would be cut into its record's queue, and an Attachment or record_file(upload=True) file the budget would refuse, first takes the place of the oldest accepted uploads no worker has started, which are counted as not uploaded (their records already list them, so Hue reads them as missing): uploads queued behind a stalled receiver never push a later record's values back into max_queue_bytes nor refuse its files; only the records of the four running uploads keep their charge until those end. A file in a log record's body that its queued copy did not hold whole is charged its bytes when export accepts its upload, and keeps its bytes inline when the budget cannot hold them. Nothing waits for room in the budget. A record's held values stay charged until its uploads finish.

Nothing is uploaded with capture_content=False or a setup key (hue_setup_…), nor from links, resources, log records not correlated with a span or in-progress placeholders; a log record uploads only its files and the attachments in its body (see Automatically captured media). Span events upload only through explicit record_file(upload=True, ...); other event values are cut as before. The redactor runs on helper content before it is queued, and the credential scrub of tool definitions and recorded requests runs on the whole value before it is hashed, so a value they rewrite is uploaded as rewritten.

Live spans

OpenTelemetry exports a span only when it ends, so a long agent turn or model call would otherwise stay invisible until it finishes. When a Hue helper span or an AI span is still open at the export worker's next 0.5 s tick, Hue queues a placeholder: a standard OTLP span whose parent is the running span, with its name, kind, start time and current attributes, an end time of 0, hue.span_type = "pending_span" and hue.pending_parent_id (the running span's own parent, omitted for a root). Hue shows the span as running and replaces the placeholder when the finished span arrives. Input set right after entering, such as set_input, is included.

The placeholder then waits for the worker's next export like any queued span: up to 0.5 s until the next tick, then the 1 s export delay when nothing else is queued, so it usually reaches Hue within about 1.5 s of its span starting. It goes out sooner when an export is already scheduled, for example because another span has just ended, and at once for 64 queued records or force_flush(). It goes out later while an earlier export is still in flight, because the worker runs the next tick and starts the next 1 s delay only after that export returns. A placeholder whose span ends before it is exported is not sent, so a short span may send none and appear in Hue only when it finishes.

• Only spans from hue.tracer (every helper) and spans with a gen_ai., ai., llm. or traceloop. attribute at start, or a name starting with ai., are announced. HTTP, database and other framework spans are not.
• Placeholders follow the same capture_content policy as finished spans, so recognized content attributes are stripped from them too when it is False. Tool definitions, system instructions and any value over 64 KiB are left out; the finished span still carries them. The markers are written last, and these reserved keys are removed from finished spans.
• Apart from that policy (capture_content=False strips recognized content keys) and those omissions, a placeholder carries the running span's attributes as they stand when it is queued. redactor= rewrites the value passed to set_input or set_output before it is set on the span, so placeholders carry the redacted value. Any other sensitive attribute, such as custom metadata or a value recorded by an instrumentor, must be scrubbed where it is recorded: the finished span carries it too, so live_spans=False does not keep it from Hue. That option only prevents the placeholder's earlier copy, for example of a value the span overwrites before it ends.
• Placeholders are advisory. At most 1,024 open spans wait to be announced, and each is announced once. A placeholder is queued only while the queue is under a quarter of its record and byte budgets and is skipped silently otherwise, so placeholders use at most a quarter of the queue. While queued they count in queued_trace_records and queued_trace_bytes, but never as dropped or failed telemetry.
• Placeholders travel in their own requests, after an export's finished spans, so a rejection count is always one kind of record's: a finished span's rejection is never credited to a placeholder, and a rejected request of placeholders never fails the export status. This needs a Hue deployment that accepts placeholders; it marks every trace acknowledgement with Hue-Pending-Spans: 1. The first trace acknowledgement without the header, whatever its request carried, comes from a generic collector or an older Hue: the SDK sets export_status.live_spans_rejected (a warning that does not affect ok) and stops announcing spans for the rest of that client's life, so a receiver that acknowledged finished spans first never receives a placeholder.
• Pass live_spans=False to send none. Setup keys (hue_setup_…) never announce spans.
• A processor wrapper that forwards on_start to Hue should also forward on_end for the same spans. Hue builds each placeholder from the running span itself, so code in the wrapper's on_end never runs on it: if the wrapper scrubs attributes, renames the span or drops it there, a placeholder exported while the span is still open is sent anyway, with the span's original name and its attributes after capture_content. A dropped span is then forgotten and its unsent placeholder discarded, but one already sent can leave the span shown as running. Do not forward on_start for spans you scrub, rename or drop, or pass live_spans=False.

Dependencies

The tracing core depends on the official OpenTelemetry packages and requests only. JSON Schema scoring (builtin_scorers.json_schema) runs jsonschema in an isolated process and needs the optional extra; without it builtin_scorers.json_schema raises ImportError and stored schema scorers report SchemaValidatorUnavailable:

pip install 'hue-run[evals]'

Import the bundle as from hue_sdk.evals import builtin_scorers. The builtins name remains as an alias for parity with TypeScript, but it shadows the standard-library module of the same name inside any file that imports it.

See THIRD_PARTY_NOTICES.md for licenses.

Confirm a trace reached Hue

Hue.verify_trace() checks a server receipt for a known trace from a real application request. It does not send a synthetic trace, invoke a model, or flush an exporter. Finish the request, then flush the provider that produced it. If you borrow a provider, call its force_flush() first, then check Hue's force_flush() result before verifying:

# Retain these IDs while your application's instrumented request runs.
# After the request finishes and the relevant providers have flushed:
confirmation = hue.verify_trace(
    request_trace_id,
    expected_span_ids=[request_span_id, model_span_id],
    required_fields=["input", "output", "model", "usage", "session"],
    timeout_millis=10_000,
)
if confirmation.verified:
    print(confirmation.receipt.trace_url)
else:
    # The latest partial receipt, or None if this trace has not appeared yet.
    print(confirmation.receipt)

Only require fields your instrumentation emits and your capture policy permits. For metadata-only capture, omit input and output. The receipt contains presence booleans and counts, not captured content. verified=True confirms the requested trace, every supplied expected span, and every required field; it does not prove that unlisted application operations were instrumented. Missing spans or fields continue polling until the deadline, then return verified=False with the latest receipt. Trace IDs must be nonzero and contain 32 lowercase hexadecimal characters; expected span IDs must be nonzero, contain 16, be unique, and number at most 100. The timeout must be positive and at most 60,000 ms.

Only trace-not-found responses, HTTP 429 and HTTP 503 are retried. Authentication, unsupported receipt endpoints, malformed responses and connection failures raise TraceVerificationError with a safe code and optional status_code; response bodies and keys are omitted. Redirects are rejected. The deadline covers connection, polling and response-body reads. An in-flight network read may finish in the background after the caller times out. This helper requires a Hue deployment that implements /api/v1/traces/{traceId}/receipt.

Supported runtimes and verification

Python 3.10+ is supported. CI tests Python 3.10 and 3.14, source imports and an independently installed wheel. Tests use synthetic loopback HTTP receivers and decode official OTLP protobuf messages to verify trace/log correlation, metadata-only capture, redaction, propagation, existing-provider ownership, authentication failures, redirects, partial rejection, retries, encoded request limits and live-span placeholders, including their downgrade against a receiver without the Hue-Pending-Spans header. Compatibility tests also exercise local evaluations and the optional OpenInference adapter. No live model provider is required for these checks.

See the documentation for integration guidance and troubleshooting for export failures.

Simulated worlds

hue_sdk.environment.EnvironmentClient drives Hue's World API: create a world after the case's execution starts, hand the agent the provider mirror URLs and the world token (never the project key), finish and wait_for_seal before the execution completes, and read the sealed world's evaluator-only evidence.

import os
import subprocess

from hue_sdk.environment import EnvironmentClient, agent_environment, mcp_config_file, world_handoff

client = EnvironmentClient(api_key=os.environ["HUE_API_KEY"])
run = client.create_run(
    idempotency_key=f"execution:{execution_id}",
    environment_version_id=version_id,
    execution_id=execution_id,
    ttl_seconds=600,
    traceparent=f"00-{span.trace_id}-{span.span_id}-01",
    agent_revision="my-agent@1.4.2",
)
world = None
try:
    world = world_handoff(run)
    if world is None:  # the deployment's gateway is off: this run has Hue-native actions instead
        client.finish_run(
            run["id"], idempotency_key=f"execution:{execution_id}:abandoned", status="abandoned"
        )
        raise RuntimeError("this deployment does not serve simulation worlds")
    child = agent_environment(world)  # os.environ minus Hue control-plane credentials, plus the carriers
    with mcp_config_file(world) as path:  # owner-only mcp.json, removed after the block
        subprocess.run(agent_command, env={**child, "MCP_CONFIG": path}, check=True)
finally:
    if world is not None:
        finished = client.finish_run(
            run["id"], idempotency_key=f"execution:{execution_id}:completed", status="completed"
        )
        client.wait_for_seal(run["id"], completing_until=finished.get("completingUntil"))
evidence = client.get_evidence(run["id"], section="ledger")

agent_environment removes HUE_API_KEY, HUE_MCP_KEY and any value carrying a Hue control-plane credential (hue_sk_, hue_setup_, hue_attempt_, hue_mcp_, hue_inv_, hue_install_, hue_at_, hue_rt_, hue_oauth_) anywhere in it, including wrapped, percent-encoded or JSON-Unicode-escaped values, unless include_hue_credentials=True; world tokens and hue_sim_ capabilities pass. For one compatibility release it also sets HUE_MCP_URL, HUE_MCP_TOKEN and HUE_MCP_EXPIRES_AT from the first MCP mirror. Nothing here logs the token. The world has its own date: world["now"] (also HUE_WORLD_NOW in the child's environment) is the world's clock at creation, a trace-built world's recorded start, and world_now() reads it as an aware datetime from a handoff or from the environment. An agent that computes "today", "tomorrow" or newer_than:7d from it lands on the recorded dates however long after the recording the run starts; one that reads the wall clock does not. The client waits Hue's Retry-After on 429 and 503 before retrying, and sends a read, or a mutation keyed by idempotencyKey, again after a connection failure, a timeout or a 408 or 5xx without the header, up to max_attempts (default 4) times; is_transient_api_error names those failures. Refusal errors expose a validated HueEnvironmentError.diagnostic from X-Hue-Diagnostic when the server sends one; the code distinguishes refusals with the same status without exposing response bodies.

Managed targets

Start a frozen dataset run in Hue while your existing agent stays in your application:

import os
from hue_sdk.managed import ManagedTargetHandler, ManagedTargetResult

def target(invocation):
    # Your function consumes unchanged inputs and verified attachment bytes.
    result = run_agent_for_evaluation(invocation)
    return ManagedTargetResult(output=result)

handler = ManagedTargetHandler(
    machine_credential=os.environ["HUE_MANAGED_TARGET_SECRET"],
    target=target,
    tracer=hue.tracer,  # Required with a Hue-owned client: Hue never sets a global tracer.
    flush_telemetry=hue.force_flush,  # Existing client; False keeps telemetry pending.
)

Without tracer, the handler falls back to the global OpenTelemetry tracer, its span is not recorded and every invocation returns uncertain.

Your POST route calls handler.handle(raw_body_bytes, request_headers) and returns its JSON body, status code and headers. Limit request bodies to 1 MiB; use asyncio.to_thread from an async route. The target must honor cancelled and deadline_monotonic. Use a 120-second host limit for the default 90-second callback and 30-second finalization budget.

For generated files, return ManagedOutputFile entries containing actual bytes, filename, content type and an optional primary flag. Managed outputs accept the same 22 content types as TypeScript: PDF, DOCX, PPTX, XLSX, application/json, text/plain, text/csv, image/png, image/jpeg, image/webp, image/gif, image/avif, audio/mpeg, audio/wav, audio/ogg, audio/webm, audio/mp4, audio/flac, audio/opus, audio/aac, video/mp4 and video/webm. Any other output content type, such as audio/pcm, records the invocation as an error before any upload. Give an .mp4 file audio/mp4 or video/mp4, and a .webm file audio/webm or video/webm, to match its content: the extension alone does not say which. The helper claims the invocation, verifies files and saves the outcome without automatically rerunning the agent. See the managed-run guide and full adapter contract for registration, existing-provider flush callbacks and recovery. Local/CI runners remain available.

Serving safely

Use create_hue_safe for best-effort startup and enabled=False for a local kill switch. The safe constructor returns a disabled client and records an instrumentation failure if initialization fails. Disabled helpers execute application work without exporting. Initialize after fork, once per serving process. force_flush_safe(timeout_millis=1000) and shutdown_safe(timeout_millis=1000) return booleans without raising; monitor those results and export_status. Keep strict project validation and receipt/delivery checks out of customer request paths. Never rerun application work to recover telemetry.

Queue bytes bound retained telemetry rather than total RSS. Network waits have a wall-clock caller bound and at most one retained HTTP worker per signal; a stuck OS call can outlive that wait. Arbitrary user hooks and third-party instrumentation retain their own behavior. See production setup for lifecycle examples, failure tests and limitations.