Skip to content

feat(semantic-kernel): traceAI-semantic-kernel on SK native diagnostics (TH-8241) - #213

Merged
nik13 merged 16 commits into
devfrom
feat/th-8241-semantic-kernel-python
Oct 26, 2026
Merged

nik13 merged 16 commits into
devfrom
feat/th-8241-semantic-kernel-python

Conversation

@nik13

@nik13 nik13 commented Oct 5, 2026 •

Copy link
Copy Markdown
Contributor

Summary

New Python package traceAI-semantic-kernel for TH-8241. It turns on Semantic Kernel's own OpenTelemetry diagnostics and maps them to Future AGI span kinds with a span processor. It does not wrap any SK function. Kernel.invoke is untouched, and a test checks that.

  • SemanticKernelInstrumentor().instrument(tracer_provider=..., sensitive=False) sets SK's three ModelDiagnosticSettings objects directly, without env vars (model_diagnostics/decorators.py:34, function_tracer.py:27, agent_diagnostics/decorators.py:28). It works whether it is called before or after the kernel is built. With sensitive=False it also switches off a sensitive flag set in env or .env, and logs a warning.
  • SK creates its tracers from the global provider (decorators.py:65, agent_diagnostics/decorators.py:33, kernel_function.py:51, chat_completion_client_base.py:32). instrument() points those module tracers at the provider you pass, and uninstrument() restores them.
  • Kinds: chat / text_completions → LLM; invoke_agent → AGENT; execute_tool for a call the model asked for → TOOL (it has gen_ai.tool.call.id, or it is a direct child of AutoFunctionInvocationLoop, which covers connectors such as Ollama that send no id); any other function call and AutoFunctionInvocationLoop → CHAIN. The kind goes on both fi.span.kind and gen_ai.span.kind, and fi-collector reads fi.span.kind first.
  • Tokens appear only on model-call spans, as SK emits them. Each trace's input-token sum equals its model calls (contract test). Every gen_ai.* key SK emits passes through unchanged.
  • instrument(config=TraceConfig(...)): with sensitive=True, hide_inputs / FI_HIDE_INPUTS drops tool arguments, agent input messages and input.value, and hide_outputs / FI_HIDE_OUTPUTS drops the output side. Other TraceConfig fields are not applied (README).
  • instrument() does not raise into app startup: a moved experimental SK module is skipped with a WARNING naming it, and a non-SDK provider gets a WARNING and is left unchanged. A SemanticKernelSpanProcessor you installed yourself is never adopted or shut down.
  • Session: SK emits no conversation id, so none is invented. gen_ai.conversation.id is copied to session.id for later SK versions. using_session / using_attributes values are copied onto SK spans when they start.

Deviations from the accepted spec (with evidence)

  • Floor is semantic-kernel>=1.38.0, not 1.44.1. Agent diagnostics first appear in 1.19.0, but the diagnostics surface this package reads is byte-identical only from 1.38.0 to 1.44.1. Up to 1.37.x the operation string was chat.completions and the agent keys were gen_ai.agent.invocation_input/output.
  • Provider routing, which the spec did not cover: see above. Without it no SK span reaches a provider passed to register().
  • Model-call message bodies are Python log records, not span events (decorators.py:375-395, 430-447). So chat spans carry no prompt or completion even with sensitive=True. This is documented, and no log bridge was built. Agent messages and tool arguments/results are span attributes, so they come through when sensitive is on.

Relationship to #200

#200 is an earlier community PR for the same package. This PR keeps its names and layout: the distribution name, SemanticKernelInstrumentor, the examples/basic_agent.py path and the README snippet. The commit has a Co-authored-by trailer for that. None of #200's wrapper code came across. Against SK 1.44.1, every kernel.invoke there raises AttributeError: ... FI_AGENT_NAME with wrapt 1.17.3, and with wrapt 2.5.0 nothing is instrumented. #200 itself is left untouched here.

Tests (head dfc510c, clean tree)

Rerun by the coordinator (tickets/TH-8241/exact-head.txt):

PYTHONPATH="python/frameworks/semantic-kernel:python:python/tests" uv run --no-project --python <py> --prerelease=allow \
  --with pytest --with opentelemetry-api --with opentelemetry-sdk --with opentelemetry-exporter-otlp-proto-http --with wrapt \
  --with requests --with jsonschema --with protobuf --with opentelemetry-proto --with "semantic-kernel==<v>" \
  pytest python/frameworks/semantic-kernel/tests -q -p no:cacheprovider --noconftest -o addopts=''
Python semantic-kernel 1.44.1 semantic-kernel 1.38.0
3.10 54 passed 54 passed
3.11 54 passed 54 passed
3.13 54 passed 54 passed
3.11, OpenTelemetry pinned to the declared floor 1.29.0 54 passed (implementer: 54 passed)

The implementer also ran Python 3.12 (54 passed on both SK versions). SK needs --prerelease, because 1.44.1 depends on azure-ai-agents>=1.2.0b3. ruff check --select E,F,W passes with E501 ignored. The repo has no ruff config, and the new lines match the existing line length.

The contract test runs through the shared harness Receiver from #203. It covers:

  • the collector path, X-Api-Key / X-Secret-Key, and project_name / project_type=observe
  • LLM/TOOL/AGENT/CHAIN kinds and parents in SK's call order
  • the token placement described above
  • session handling, streaming and error status
  • no content with sensitive off, plus a sensitive-on control run
  • a collector-down run where the kernel still returns
  • the example run against the Receiver

No file under java/ is touched. SK is MIT-licensed.

Review

Round 1 ran as separate installed cards on Opus 5.5:

  • pr-reviewer t_6080eed8: APPROVE.
  • pr-verifier t_4349d82a: VERIFIED.

Both were at 31f910c. Their non-blocking items were still fixed before the demo, test-first, in 05c26ce..dfc510c. Run against 31f910c's code, the final tests give 13 failed and 41 passed.

ID Finding Resolution
R1 Id-less auto-invoked tools were CHAIN TOOL when the tool call id is present or the parent is AutoFunctionInvocationLoop; tested with a real SK run and an id-less fake connector
R2 instrument() could raise into startup Warn and skip per module; non-SDK provider warns and changes nothing
R3 Double tokens with a client instrumentor README: use one or the other
R4 Processor ordering after add_span_processor README; a test pins the drop and the recovery
R5 uninstrument() shut down a user's processor instrument() owns only its own processor
R6 Python 3.13 and the OTel floor were untested Both tested; a test-only processor fixed for OTel 1.29.0
N1 TraceConfig hide flags ignored hide_inputs / hide_outputs and their env vars applied with sensitive=True

Round 2: pr-verifier t_6c1b41ee: VERIFIED, ready to merge, at dfc510c (still the branch tip; nothing outside python/frameworks/semantic-kernel changed). All seven items above are resolved, with file:line evidence traced against semantic-kernel 1.44.1 / 1.38.0, fi-instrumentation-otel and OTel 1.29.0. Its regression trace found nothing: uninstrument() still restores SK's module tracers and diagnostics settings, tokens stay on model-call spans only, sensitive=False still strips agent messages and tool arguments/results, and processor ownership survives instrument/uninstrument/instrument. It relied on my test-run results rather than re-running them. Review is complete at dfc510c.

Follow-up for after merge (P3, non-blocking):

  • N2: any direct child of AutoFunctionInvocationLoop is kinded TOOL. So a function that a user's AUTO_FUNCTION_INVOCATION filter calls through context.kernel.invoke, or one called from inside a streaming consumer's async for body (SK yields inside the loop span), shows as TOOL instead of CHAIN. Kind only; no token or content effect. Fix: a README note, or also match gen_ai.tool.name against the loop's sk.available_functions.

Video demo

Narrated terminal demo, 6:40, recorded at head dfc510c (the reviewed head), 1080p H.264/AAC, 8 chapters. sha256 e6516f6eaf532456aee980f9a5574e0b1b8d519041f4d1c100141f2ed8031630.

The video, captions, transcript, chapter list, preview and media check are attached privately to Linear TH-8241 (Future AGI workspace access needed). Every run uses a loopback fake OpenAI-compatible model, placeholder keys and the shared harness receiver: no vendor call and no live fi-collector.

Time Chapter
00:00 The problem
01:00 Head dfc510c: branch, 0 uncommitted files, remote at the same SHA, 7 commits
01:35 Content off (default): one export to /tracer/v1/traces with both auth headers (values not shown) and project_name; 19 spans in 7 traces with LLM / AGENT / CHAIN / TOOL kinds; session.id on the agent run; ERROR status on the broken model and tool; no prompt or tool argument in any key; input tokens 88 on LLM spans = 88 on all spans
03:12 Content on (sensitive=True): agent messages and tool arguments/results exported
03:48 FI_HIDE_INPUTS: prompt and tool arguments dropped again; the city remains only in the tool result (output side)
04:28 The change: kind_for (R1) and the hide-flag key drop (N1)
04:58 Tests: 54 passed on camera, 0 uncommitted files after, plus the logged matrix at this head
05:49 Review and limits

Two disclosed cuts remove frozen waits (5 s in chapter 3, 13 s in chapter 7). Checked: full decode, no silence of 2.5 s or more, no keys or header values in frames, captions or transcript.

Limits

  • The Azure AI Inference connector builds its own diagnostics settings from env at call time (azure_ai_inference_tracing.py:25), so instrument() does not turn it on.
  • SK emits no cost, retrieval, cache or reasoning tokens.
  • hide_input_messages, hide_output_messages and the other TraceConfig fields are not applied (documented).
  • Unexpected errors inside instrument() still roll back and raise.
  • The real Ollama connector was not run; R1 uses a fake connector shaped like its _parse_tool_calls.
  • Cancelling a stream midway is not tested. Only the loopback Receiver was used; there was no real fi-collector and no live model.
  • README only; the docs page is tracked in TH-8103.

Stacked on #203 (shared harness). Not merged.

Linear: TH-8241. approved: Nikhil 2026-10-03 blanket

Verification sweep (2026-10-12)

  • Verified end to end: the real library made real calls, traces were exported to a local Future AGI stack built from dev, and the spans were checked in ClickHouse and in the Observe UI (hierarchy, span types, input/output, model and tokens, sessions, error status, and a secret scan of everything stored).
  • Commits added in this sweep:
    • 9fa1149 feat(semantic-kernel): put model-call prompts and completions on chat spans
    • 9651b34 fix(semantic-kernel): show a single completion as the assistant reply
    • c0eb2dd fix(semantic-kernel): require fi-instrumentation-otel>=1.1.0
  • Merge order: merge test(harness): shared OTLP harness for TH-8103 contract tests #203 first (merge commit), and the README catalog PR docs: list the TH-8103 integrations in the README catalogs #250 last. Everything else is order-independent.

Re-verification (2026-10-16)

Head after re-verification: 525069c (one new commit on top of c0eb2dd). Nothing rewritten.

Versions. Latest stable upstream semantic-kernel 1.45.0 (PyPI, 2026-10-16; openai 3.33.0, fi-instrumentation-otel 1.1.0, Python 3.12), plus 1.44.1 (earlier pin) and 1.38.0 (declared floor, Python 3.10). The package was installed from the branch with pip install (not PYTHONPATH) for the live runs.

Tests. Suite is now 63 tests (61 before this commit's 2 new ones): 1.45.0 py3.12 63 passed; 1.44.1 py3.12 63 passed (61 before the fix); 1.38.0 py3.10 63 passed.

Live E2E against local Future AGI (real SK ChatCompletionAgent + real OpenAI connector to CLIProxy gpt-6-luna, one real tool, using_session; run exactly as the README shows). fi_verify PASS on each; hierarchy invoke_agent (agent) > AutoFunctionInvocationLoop (chain) > chat (llm) / execute_tool Weather-get_weather (tool) / chat (llm); model gpt-6-luna; tokens 380/36 and 429/14 on the two chat spans only.

  • 1.45.0 default: project th8103-rick-semantic-kernel-python-r2, trace fa0837a5-df7b-49b9-8336-f19091dfa7b1
  • 1.45.0 failure (tool raises): ...-r2-fail, trace 7057fe66-c0d1-43f2-a601-83dba337ebfd, tool span ERROR (status text RuntimeError); the agent still answered
  • 1.45.0 sensitive=True: ...-r2-on2, content present (agent in/out 94/101 B, chat in/out 236/180 and 546/68 B, tool 17/46 B)
  • 1.44.1 default: ...-r2-pin, same tree and tokens 381/35, 431/12
    UI screenshot of the failing trace checked by eye (tree, ERROR markers, agent graph).

Finding fixed (privacy). With sensitive=False the exported spans still carried SK's exception events with exception.message and exception.stacktrace (local file paths incl. the home directory/username), and the ERROR status description was RuntimeError: <message>. Captured on the wire with a local OTLP sink: /Users/ x9, Traceback x2, the marker in the message x7. (The fi-collector drops event attributes, so the stored rows only showed the status text, but any other OTLP backend would have received them.) Fix: the processor always strips exception.stacktrace; with sensitive=False it also strips exception.message and cuts the status description to the exception type. After the fix the same capture has 0 occurrences of /Users/, Traceback, username and the message. Two new unit tests; README and CHANGELOG updated.

Privacy proof. Unique marker SKN115882540 in the prompt, the tool result and the tool exception message. A scan of every column of every stored spans row (select *, incl. attributes, resource attrs, events, status message): default run: 0 hits for the prompt, tool-result, message, Traceback, /Users/, username and Paris; failure run: 0 for the message (only the user-supplied session id contains the marker); sensitive=True run: prompt marker in 3 rows and tool-result marker in 2 rows, still no stack trace or path. Tool definitions (the registered function schema) are exported with content off; they are configuration, not user content.

Limitations/notes. First sensitive=True run (-r2-on) lost its export during a ClickHouse MEMORY_LIMIT_EXCEEDED period on the shared stack and was re-run (-r2-on2). Docs page: future-agi/docs draft PR (see below), nav suggestion group Integrations > traceAI, title "Semantic Kernel (Python)".

🤖 Generated with Claude Code

nik13 and others added 10 commits October 3, 2026 18:58
The Node OTLP exporter streams with Transfer-Encoding: chunked and sends
no Content-Length. The Receiver read Content-Length only, so every Node
export came back 400 with zero spans; two TH-8103 children had to put a
de-chunking relay in front of it. Receiver now accepts both framings.

It also records one entry per accepted export (path, lower-cased
headers, flattened resource attributes) via requests(), so a contract
test can assert X-Api-Key/X-Secret-Key and project_name through the
shared harness instead of a private recorder.

Verified: 6 harness tests pass, and the TanStack example's real Node
exporter delivered 4 chunked exports (4 spans, collector path, both
auth headers, project_name/project_type) with no relay.

Refs: TH-8339, TH-8103
…ostics

New package python/frameworks/semantic-kernel (traceAI-semantic-kernel).
SemanticKernelInstrumentor().instrument() turns on Semantic Kernel's own
OpenTelemetry GenAI diagnostics in process and prepends a mapping-only
span processor ahead of the register() exporter. No wrapt, nothing wraps
Kernel.invoke / KernelFunction.invoke, no spans of its own. Python only;
java/ untouched.

Processor: fi.span.kind + gen_ai.span.kind (chat/text_completions -> LLM,
invoke_agent -> AGENT, execute_tool with gen_ai.tool.call.id -> TOOL,
execute_tool without it and AutoFunctionInvocationLoop -> CHAIN),
gen_ai.system -> gen_ai.provider.name, derived total tokens on LLM spans,
gen_ai.conversation.id -> session.id unless set, using_attributes context
copied onto SK spans, promoted token/cost keys moved to
semantic_kernel.usage.* on non-LLM spans, content keys dropped unless
sensitive=True, ERROR status when error.type is set and status unset.
Exceptions are swallowed.

Tests first: 40 passed on semantic-kernel 1.44.1 (py3.11, py3.12) and
1.38.0 (py3.10, py3.11), including a shared-harness contract test (real register() exporter ->
python/tests/harness Receiver, loopback fake OpenAI, placeholder keys).

Deviations from the TH-8241 specs (file:line in installed semantic-kernel
1.44.1 unless noted):
- Floor >=1.38.0, not 1.44.1. Agent diagnostics exist from 1.19.0, but
  the surface read here is identical only from 1.38.0 to 1.44.1 (wheel
  diff); <=1.37.x uses "chat.completions" and gen_ai.agent.invocation_*.
- Settings spike resolved: three module-level ModelDiagnosticSettings()
  (model_diagnostics/decorators.py:34, function_tracer.py:27,
  agent_diagnostics/decorators.py:28), read at call time. instrument()
  assigns their fields; no env vars, works before or after the kernel is
  built. sensitive=False also forces an env-set sensitive flag off.
- Tracer routing, not in the spec: SK's tracers come from the global
  provider (decorators.py:65, agent decorators.py:33,
  kernel_function.py:51, chat_completion_client_base.py:32) and register()
  does not set it. instrument(tracer_provider=...) points those four
  module-level tracer names at the given provider; uninstrument restores.
- Sensitive model content is Python logging records (decorators.py:375-395,
  430-447), not span events; LLM spans carry no input/output even with
  sensitive=True. Agent messages (agent decorators.py:208-227) and tool
  args/results (kernel_function.py:268-269, 283-288) are span attributes.
- Operation strings confirmed (decorators.py:37-38, agent decorators.py:35,
  function_tracer.py:22). execute_tool opens on every KernelFunction.invoke
  (kernel_function.py:264); only model-requested calls carry a tool call id
  (kernel.py:468-471), so function vs tool needs no wrapper.
- SK 1.44.1 emits no conversation/thread id (agent gen_ai_attributes.py:9-16).
  using_attributes only reaches FITracer spans, so the processor copies the
  fi context attributes onto SK spans at start.
- Also sets fi.span.kind (fi-collector converter.go:80 reads it first).
- License re-read: MIT (semantic_kernel-1.44.1.dist-info/licenses/LICENSE).
- The given uv test command does not resolve: semantic-kernel 1.44.1 needs
  pre-release azure-ai-agents>=1.2.0b3. Added
  --prerelease=if-necessary-or-explicit --with 'azure-ai-agents>=1.2.0b3'.

PR #200 reuse (NivedhN160, 7b8aef9): package path, distribution name,
SemanticKernelInstrumentor with instrument()/uninstrument(), the
examples/basic_agent.py path and asyncio.run(main()) shape, and the README
install-then-instrument snippet. No wrapper code, test or poetry.lock.

approved: Nikhil 2026-10-03 blanket
Refs: TH-8241
Co-authored-by: NIVEDH SUNIL <nishikradha@gmail.com>
An auto-invoked kernel function was TOOL only when Semantic Kernel put
gen_ai.tool.call.id on its execute_tool span. Connectors that send no tool
call id (Ollama's _parse_tool_calls builds FunctionCallContent without one)
made every model-requested tool render as CHAIN, contrary to PRD R-03.

The processor now also records, at on_start, spans whose parent is
Semantic Kernel's AutoFunctionInvocationLoop span, and kind_for() returns
TOOL for such an execute_tool span. Functions invoked directly, or called
from inside a tool, stay CHAIN. The pending set is keyed by
(trace_id, span_id), drained on end and bounded.

Tests: synthetic loop/tool/nested spans, and real Semantic Kernel auto
function invocation driven by a fake id-less chat service. README wording
fixed.

approved: Nikhil 2026-10-03 blanket
Refs: TH-8241
instrument() imported every Semantic Kernel module it touches outside its
try block and raised TypeError for a provider without an SDK span-processor
chain, so a future Semantic Kernel that moves an experimental diagnostics
module, or a host that passes a non-SDK provider, failed at startup.

Now each module is imported once; one that cannot be imported (or whose
settings object lost a field) is skipped with a WARNING naming it, and the
rest is still switched, routed and mapped. A non-SDK provider (including
the global ProxyTracerProvider before any provider is set) logs a WARNING
and leaves Semantic Kernel unchanged. uninstrument() stays safe in both
cases. Unexpected errors while applying still roll back and raise.

approved: Nikhil 2026-10-03 blanket
Refs: TH-8241
instrument() adopted a SemanticKernelSpanProcessor the user had already
put on the provider, ignoring the sensitive value passed to instrument(),
and uninstrument() then removed and shut down the user's processor.

instrument() now always installs its own processor (logging a warning
when a user-installed one is present) and uninstrument() removes and
shuts down only that one, by identity. The user's processor stays in the
chain and keeps mapping spans.

approved: Nikhil 2026-10-03 blanket
Refs: TH-8241
instrument(config=TraceConfig(...)) was swallowed by **kwargs with a DEBUG
log, and with sensitive=True the processor copied tool arguments and agent
messages to input.value / output.value regardless of TraceConfig,
FI_HIDE_INPUTS or FI_HIDE_OUTPUTS.

instrument() now takes config (a TraceConfig; TypeError otherwise) and
builds TraceConfig() from the environment when it is omitted. The
processor takes the same config. With sensitive=True, hide_inputs drops
gen_ai.input.messages, gen_ai.tool.call.arguments, input.value and
input.mime_type; hide_outputs drops gen_ai.output.messages,
gen_ai.tool.call.result, output.value and output.mime_type. Default
sensitive=False behaviour is unchanged (all content keys removed).

Tests: each flag via config and via env against a real Semantic Kernel
execute_tool span and an agent span, a no-hiding control, a user-built
processor reading FI_HIDE_OUTPUTS, and the wrong-type rejection. Test
fixtures and the contract subprocess env no longer inherit FI_HIDE_*.

approved: Nikhil 2026-10-03 blanket
Refs: TH-8241
…count

R4: fi_instrumentation's TracerProvider.add_span_processor shuts down and
empties the chain register() set up on its first call, which also drops
this package's processor while instrument() still reports instrumented.
The README now says to call instrument() after any add_span_processor of
your own, and to uninstrument()/instrument() again otherwise. A test pins
both the drop and the recovery on fi's TracerProvider (it passes on the
previous head too: the reset is existing fi behaviour, not changed here).

R3: the README warns against co-enabling a client-level instrumentor
(traceAI-openai etc.): Semantic Kernel's chat span already carries the
tokens, the client span repeats them, and the processor only touches
Semantic Kernel spans, so trace and session sums double. Use one or the
other.

approved: Nikhil 2026-10-03 blanket
Refs: TH-8241
R6: the 3.13 classifier and the opentelemetry >=1.29.0 floor were not
backed by a test run (README listed 3.10-3.12 and sdk 1.45.0 only).

The full suite now passes on Python 3.10, 3.11, 3.12 and 3.13 with
semantic-kernel 1.44.1 and 1.38.0, and with opentelemetry-api/-sdk/
-exporter-otlp-proto-http/-proto 1.29.0 on both SK versions, so the
3.13 classifier and the 1.29.0 floor stay as declared.

The one 1.29.0 failure was in the test-only snapshot processor of the
contract app: opentelemetry-sdk 1.29's base SpanProcessor.force_flush
returns None, which made provider.force_flush() report False. The
snapshot processor now returns True. README tested-versions table,
test command, pyproject comment and CHANGELOG updated.

approved: Nikhil 2026-10-03 blanket
Refs: TH-8241
@nik13
nik13 requested a review from NVJKKartik October 10, 2026 08:15
nik13 and others added 3 commits October 11, 2026 20:26
… spans

With sensitive=True, Semantic Kernel writes a model call's messages only as
INFO log records inside the chat span (model_diagnostics/decorators.py
_set_completion_input / _set_completion_response), so every LLM span in
Observe had tokens but no input or output.

instrument(sensitive=True) now puts a proxy in front of that module's
logger. Its info() appends each message record to the current chat span as
gen_ai.input.messages / gen_ai.output.messages, then calls the real logger
with the same arguments, so log handlers see what they saw before. The
processor lifts those to input.value / output.value as it does for agent
spans, and TraceConfig hide_inputs / hide_outputs drop them the same way.
No proxy with sensitive=False; uninstrument() restores the original logger.
It replaces a module attribute, like the tracer routing; no Semantic Kernel
function is wrapped.

Tested on semantic-kernel 1.38.0 and 1.45.0 (unary and streaming), and in a
live run on CLIProxy where the chat spans render as User / Assistant / Tool
messages in Observe.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
output.value on chat spans was the one-item gen_ai.output.messages list.
Observe's chat view renders an output object {role, content} as the
assistant reply and nothing for a list, so the completion was only visible
in the attributes table. Put a single completion on output.value as its
message object; more than one stays a list. gen_ai.output.messages is
unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
fi-instrumentation-otel 1.0.0's register() raises AttributeError
('HTTPSpanExporter' object has no attribute '_headers') with
opentelemetry-exporter-otlp-proto-http 1.45, so a fresh install that
resolves 1.0.0 crashes on startup. Use the floor the other TH-8103
packages already declare.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…unless sensitive

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
nik13 and others added 2 commits October 21, 2026 14:48
0.1.0 has never been published, so the privacy fix and the 1.45.0 test run belong in its entry, not under Unreleased.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@nik13
nik13 merged commit 26956e6 into dev Oct 26, 2026
4 checks passed
@nik13
nik13 deleted the feat/th-8241-semantic-kernel-python branch October 26, 2026 09:35
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant