Repository navigation
Conversation
…and count usage once Response output spans now carry only response-owned fields, so request instructions and inputs are no longer echoed into exported output even when hide flags are set. Token usage for one Responses call is recorded once, on the request span. Adds regression coverage.
Authorization and API-key headers passed through extra_headers are no longer recorded in span invocation parameters; non-sensitive request parameters are still recorded. Adds regression coverage.
run_stream keeps its async context-manager behavior, FutureAGI content privacy controls apply to captured content, run spans carry canonical content and span-kind attributes, and the default instrumentor installs its tool and model execution hooks. Adds regression coverage.
… input capture Invocation parameters are allowlisted for completion, embedding and image generation, so API keys and headers are never exported. Failed streams end their span, async stream setup keeps provider exceptions, successful streams record token usage, structured-output response_format classes and positional calls are handled, image-generation usage no longer raises, and canonical input keeps the full message list. Adds regression coverage.
AgentChat spans now use the canonical FutureAGI attribute names so hide-input/output settings apply, and input-only team results no longer expose hidden task text as output. Adds regression coverage.
…ody capture converse_stream is instrumented, non-Anthropic streams record model, usage and output when consumed instead of finishing early, and file-like invoke_model bodies are accepted without being consumed. Adds regression coverage.
…keep every Gemini candidate Provider error payloads inside HTTP 200 responses (including streamed failures) are exported as errors with the provider message, and Gemini wire capture keeps all generated candidates. Adds regression coverage.
The instrumentation reads gen_ai span attribute names that first ship in fi-instrumentation-otel 1.0.0, so importing it with an older core fails with an AttributeError. Raise the declared floor so the resolver upgrades the core instead. The lock already pins 1.1.0; only its content hash changes.
The instrumentation reads gen_ai span attribute names that first ship in fi-instrumentation-otel 1.0.0, so with an older core its spans fail with an AttributeError. Raise the declared floor so the resolver upgrades the core instead. The lock already pins 1.1.0; only its content hash changes.
The instrumentation reads gen_ai span attribute names that first ship in fi-instrumentation-otel 1.0.0, so with an older core it fails with an AttributeError. Raise the declared floor so the resolver upgrades the core instead. The lock already pins 1.1.0; only its content hash changes.
- Record the input messages of one-shot calls. - Keep the IDs of tool calls returned in the model output.
The run span is retained for the whole streamed run instead of ending before the streamed steps execute.
…val models and session identity - Trace consumption of generators returned by traced functions. - Serialize custom eval models whose fields are null. - Emit the native session identity alongside the conversation id.
Resource attribute keys use their canonical names.
…uses Sixteen packages declared a core floor below the names their code reads (most use the SpanAttributes.GEN_AI_* names first shipped in 1.0.0), so an environment that already had an older core failed at import or on the first traced call. Raise each floor to the first core that has every name, refresh the content hash of the seven Poetry locks (no version changes), and add a real-core name check to beeai and strands, whose other tests replace fi_instrumentation with mocks.
Requests with a system message referenced a span attribute that no fi-instrumentation-otel release defines, so their span was silently dropped. Remove that attribute (the system message is already recorded in the input messages), require fi-instrumentation-otel 1.0.0 or later, and test the package against the real core, including hidden inputs.
…tup for TypeScript
…rom the global fetch
…ents in TypeScript tracing - Preserve send options and the callback overloads of send(). - Leave stream errors to the caller, never pull a stream ahead of the caller, and end a never-read stream's span when it is collected. - Keep the requested model id and record finish reasons and time to first token. - Record streamed tool arguments, exporting a fixed placeholder instead of a fragment that never formed valid JSON.
…age in TypeScript tracing - A response that arrives with HTTP 200 but carries an error body is returned to the caller unchanged and its span is marked as an error, instead of the traced call throwing a TypeError and dropping the span. This covers chat completions, legacy completions and responses. - Token usage reported for embedding requests is recorded on the span.
Future AGI files every span under its resource's project and refuses a batch whose resources carry no project_name, so eval traces from the reporter never reached Observe. The reporter now sends project_name (the projectName option, else FI_PROJECT_NAME, else vitest-evals) and project_type observe, the same resource keys @traceai/fi-core sends.
- Patched state is kept inside the package instead of being written onto the read-only module namespace, so setting up and disabling tracing no longer throws or logs an error; a flag already present on a module is still honored. - The quick start keeps OpenTelemetry debug logging opt-in.
- The run_stream span carries the agent span kind and the prompt as its input, and once the stream is consumed, the output the caller received, like non-streamed runs; hide inputs and hide outputs are honored and streaming behavior is unchanged.
- Packed binary and ubinary embedding responses are recorded as span vectors with cohere.embedding_type, like int8 and uint8, and are hidden by the embedding-vector privacy setting.
…on-dev fix(fi-instrumentation-otel): trace generator consumption, nullable eval models and session identity
fix(anthropic): keep one-shot call messages and output tool call IDs
fix(smolagents): keep the run span through streaming execution
…-on-dev fix(pydantic-ai): record input and output on streamed agent runs
…-run-on-dev docs: quickstarts for four Python integrations run as written
…ave9-on-dev fix(fi-semantic-conventions): use canonical resource keys
…t-on-dev fix(fi-core): use a full collector endpoint from the environment as given
…on-dev fix(bedrock): keep send options, caller-driven streams and tool arguments in TypeScript tracing
…n-dev fix(cohere): capture integer embedding vectors
…9-on-dev fix(langchain): preserve normalized tool calls and result links
…ag-on-dev fix(langchain): stop printing a TypeError on every ES module run
…n-dev fix(openrouter): mark completed TypeScript calls OK like the other traceAI instrumentors
docs(openrouter): show the install command and Future AGI exporter setup for TypeScript
docs(n8n): install from npm now that 0.1.0 is published
…9-on-dev fix(anthropic): keep TypeScript tracing across re-instrumentation and keep message content
fix(vercel): keep embedding token usage in auto mode
…-on-dev fix(vitest-evals): name the Future AGI project on exported eval traces
fix(cohere): record binary and ubinary embedding vectors
fix(langchaingo): let the example honor FI_BASE_URL
….3.2 and framework fixes)
…onventions 1.3.0 and package fixes)
chore(release): release 28 Python packages (fi-instrumentation-otel 1.3.2 and framework fixes)
chore(release): release 10 npm packages (fi-core 1.2.4, fi-semantic-conventions 1.3.0 and package fixes)
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Brings main up to date with dev so the release workflow publishes the next traceAI versions.
After this merge the release workflow publishes, in dependency order:
Each published version gets its tag and GitHub release. What users get is in each package's CHANGELOG.md.
Pull requests included since the last merge to main (40 plus the two release pull requests): #625, #626, #627, #628, #629, #631, #632, #633, #634, #635, #637, #640, #641, #642, #643, #644, #645, #646, #647, #649, #650, #651, #652, #653, #655, #656, #657, #658, #659, #660, #661, #662, #665, #666, #668, #669, #670, #671, #672, #673.