Perf/hot path overhaul - #114

Merged

berges99 merged 28 commits into

mainfrom

perf/hot-path-overhaul

Aug 5, 2026

Commits on Aug 4, 2026

perf(core): cut framework overhead 2-4x across agent and workflow hot paths

berges99
and
cursoragent
committed
9d299cb
perf(core): fast paths for non-streaming handlers, approvals-off, and single-tool dispatch

berges99
and
cursoragent
committed
889852f
test(core): pin exact event-stream shapes for tools, agents, workflows

berges99
and
cursoragent
committed
bbe5917
perf(core): split __call__ into raw _stream() + collector wrapper at the API boundary

berges99
and
cursoragent
committed
ddab93c
refactor(core): decompose Runnable._stream into named lifecycle methods

berges99
and
cursoragent
committed
dc6a52f
bench: long conversations / long messages vs LangGraph

berges99
and
cursoragent
committed
6d37bce
perf(core): cache Message dumps + cut serialization overhead on long histories

berges99
and
cursoragent
committed
8520ce1
perf(core): sequential fast path for linear workflows + shared _run_step

berges99
and
cursoragent
committed
76db757
feat(core): warn once when a sync handler blocks the event loop

berges99
and
cursoragent
committed
4db236b
fix(core): restore caller's run context after nested top-level runnable

berges99
and
cursoragent
committed
afbe96a
bench: refresh all cross-framework results after the perf branch

berges99
and
cursoragent
committed
c922cb8
bench: clear traces per latency iteration in long-conversation harness

berges99
and
cursoragent
committed
9372eaa
fix(core): re-raise pending cancellation in agent tool fast paths

berges99
and
cursoragent
committed
5923a4e
fix(server): don't pre-load voice ONNX models for non-voice deployments

berges99
and
cursoragent
committed
b8e9948
perf(core): lazy provider-SDK imports — 'from timbal import Agent' 1.28s -> 0.39s

berges99
and
cursoragent
committed
de7d449
perf(core): per-provider SDK isolation — anthropic-only runs never import openai

berges99
and
cursoragent
committed
085695b

Navigation Menu

Uh oh!

Perf/hot path overhaul - #114

Perf/hot path overhaul#114
berges99 merged 28 commits into
mainfrom
perf/hot-path-overhaul

Commits on Aug 4, 2026

perf(core): cut framework overhead 2-4x across agent and workflow hot paths

perf(core): fast paths for non-streaming handlers, approvals-off, and single-tool dispatch

test(core): pin exact event-stream shapes for tools, agents, workflows

perf(core): split call into raw _stream() + collector wrapper at the API boundary

refactor(core): decompose Runnable._stream into named lifecycle methods

bench: long conversations / long messages vs LangGraph

perf(core): cache Message dumps + cut serialization overhead on long histories

perf(core): sequential fast path for linear workflows + shared _run_step

feat(core): warn once when a sync handler blocks the event loop

fix(core): restore caller's run context after nested top-level runnable

bench: refresh all cross-framework results after the perf branch

bench: clear traces per latency iteration in long-conversation harness

fix(core): re-raise pending cancellation in agent tool fast paths

fix(server): don't pre-load voice ONNX models for non-voice deployments

perf(core): lazy provider-SDK imports — 'from timbal import Agent' 1.28s -> 0.39s

perf(core): per-provider SDK isolation — anthropic-only runs never import openai

Commits on Aug 5, 2026

test: bump files-integration anthropic model to claude-sonnet-5

perf(codegen): lazy transformer registration — CLI --help 381ms -> 60ms

fix(core): sync generators yielding None no longer truncate the stream

fix(core): rename command-path loop variable shadowing the agent iteration counter

refactor(codegen,evals): extract shared helpers to kill copy-paste clones — net -187 lines

refactor(core): split llm_router into core/llm package — one module per concern, zero perf cost

Merge origin/main into perf/hot-path-overhaul

refactor(tests): mirror package layout — move tracing/llm/collector tests out of tests/core

feat(core): enable Anthropic automatic prompt caching by default — cache reads cost 0.1x

refactor(tests): move parallelism and files-integration tests to state/ and types/

fix(core): let server-side tools coexist with client tools and survive provider switches

fix(types): drop assistant turns left empty after cross-provider replay