Thanks for your interest in improving the F# SDK for ARCP. This document covers how to report issues, propose changes, and get a change merged.
By participating you agree to the Code of Conduct.
ARCP is two things in two places, and a change belongs to exactly one of them:
- The protocol — the wire format, message semantics, lease rules, error taxonomy, feature flags. These live in the specification repository. If your idea changes what goes on the wire or what a conformant runtime must do, it is a spec change — open it there, not here. This SDK implements the spec; it does not define it.
- This SDK — how the protocol is expressed idiomatically in F#: bugs, ergonomics, performance, missing-but-specified features, docs, tests. Those belong here.
When in doubt, open an issue here and we'll redirect if it's really a protocol question.
A change to this SDK must keep it a faithful client of ARCP v1.1 (draft). Concretely:
- Don't invent wire behavior. No envelope fields, event kinds, error codes, or feature flags that the spec doesn't define. If you need one, it's a spec proposal first.
- Negotiate honestly. Only advertise a feature flag in
session.helloonce the SDK actually implements it. The feature matrix in the README must match what the code negotiates — a row markedSupportedis a promise. - Respect the semantics. Sequence numbers stay gap-free and monotonic;
LEASE_EXPIREDandBUDGET_EXHAUSTEDstay non-retryable; the effective feature set is the intersection of client and runtime advertisements. Tests must not paper over a semantic the spec requires. - Stay layered. This SDK controls runtimes. It does not expose tools (that's MCP) or export telemetry (that's OpenTelemetry). PRs that blur those layers will be asked to move the logic out.
Open an issue with: the SDK version and F# version, the runtime you
connected to, a minimal reproduction (the smallest program that triggers it),
what you expected, and what happened. A failing test is the best possible bug
report. Wire-level traces (the envelopes exchanged) help enormously for protocol
behavior — redact any auth.token or provisioned-credential value first.
For anything beyond a small fix, open an issue describing the problem before writing code, so we can agree on the approach. Small, focused PRs review faster than large ones; if a change is big, say so early and we'll help break it down.
The SDK targets .NET 10 (net10.0). The required SDK version is pinned in
global.json (10.0.203, rollForward: latestFeature) — install a matching
.NET SDK from dot.net and the .NET CLI
will respect it automatically. NuGet is the package manager; all package
versions are pinned centrally in Directory.Packages.props. Clone, restore
local tools (Fantomas), then restore project dependencies:
git clone https://github.com/agentruntimecontrolprotocol/fsharp-sdk.git
cd fsharp-sdk
dotnet tool restore
dotnet restore ARCP.slnx
dotnet build ARCP.slnx --configuration ReleaseTwo layers must pass before a PR merges:
-
Unit tests — this SDK's own suite:
dotnet test tests/Arcp.UnitTests/Arcp.UnitTests.fsproj -
Conformance — the SDK's behavior against the reference runtime. New protocol-facing code (session negotiation, event sequencing, lease handling, error mapping) needs a test that exercises the real exchange, not a mock that assumes the answer. The integration suite under
tests/Arcp.IntegrationTestsruns the F# client against the in-processArcp.Runtimereference implementation; run it withdotnet test tests/Arcp.IntegrationTests/Arcp.IntegrationTests.fsproj. To point it at an out-of-process runtime, set theARCP_RUNTIME_URLenvironment variable to its WebSocket endpoint before invoking the suite.
CI runs both on every PR. A PR that changes which feature flags the SDK negotiates must also update the README feature matrix in the same change.
The full coverage report is regenerated with:
dotnet test ARCP.slnx --collect:"XPlat Code Coverage" \
--results-directory TestResults/review-coverage
reportgenerator \
-reports:"TestResults/review-coverage/*/coverage.cobertura.xml" \
-targetdir:"TestResults/coverage-report" \
-reporttypes:"TextSummary"Install the report tool once with
dotnet tool install -g dotnet-reportgenerator-globaltool. The summary lands
at TestResults/coverage-report/Summary.txt. The target is ≥ 80 % line
coverage; transport and async-state-machine paths drive most of the
remaining branch gaps and additions there are welcome.
Formatting is enforced by Fantomas
(restored as a local tool); the compiler runs with
TreatWarningsAsErrors=true plus an explicit WarningsAsErrors list
(FS0025;FS0026;FS0040;FS0064) so warnings are part of the build gate. Run
the same commands CI runs:
dotnet tool restore
dotnet fantomas --check src tests samples
dotnet build ARCP.slnx --configuration ReleaseMatch the surrounding code. Public API changes need doc comments and an entry in the changelog. Prefer clarity over cleverness in a library others build on.
-
Write focused commits with present-tense, imperative subjects (
add result_chunk reassembly, notadded/adds). -
Reference the issue a PR closes (
Closes #123). -
Keep the PR description honest about scope and any spec sections touched.
-
Rebase on the default branch and ensure CI is green before requesting review.
-
Sign off your commits to certify the Developer Certificate of Origin:
git commit -s -m "your message"
Releases are cut by maintainers. Pushing a v* tag (for example v1.1.0) to
the default branch triggers the publish GitHub Actions workflow, which packs
Arcp and Arcp.Cli and pushes the resulting .nupkg/.snupkg artifacts to
nuget.org. The SDK is versioned with
semantic versioning independently of the protocol version it speaks; a protocol
version bump is noted in the changelog when the negotiated ARCP version changes.
By contributing, you agree that your contributions are licensed under the project's Apache-2.0 license.