feat(rivetkit): trace actor to actor calls

[eersnington]

• Tracing HTTP action calls between actors, including routing, retries and wake time
• Connecting both actors' work in the same trace
• Testing that overlapping calls keep their traces and logs separate

[eersnington]

Stack for rivet-dev/actors

• #5734 docs(rivetkit): add actor tracing docs
• #5733 feat(rivetkit): forward otel sdk warnings to the pino logger
• #5732 feat(rivetkit): trace actor to actor calls 👈
• #5730 feat(rivetkit-core): trace queue sends and receives
• #5731 feat(rivetkit-core): trace raw http requests
• #5729 feat(rivetkit-core): trace scheduled invocations
• #5726 feat(rivetkit): send trace context from clients
• #5723 feat(rivetkit): pass trace context to js and logs
• #5727 feat(rivetkit): trace sqlite operations
• #5721 feat(rivetkit): trace actor invocations
• #5728 fix(pegboard-gateway): forward request ray id to actors
• #5724 test(rivetkit): let driver runtimes take extra environment
• #5722 fix(rivetkit): disable services in driver test runtimes
• main

Get stack: forklift get 5732
Push local edits: forklift submit
Merge when ready: forklift merge 5732

_{change zosqvyzx}

[claude]

Review: feat(rivetkit): trace actor to actor calls

Reviewed the diff across rivetkit-core, the NAPI bridge, and the TypeScript client. Overall this is a well-scoped, carefully-commented change that follows the repo's telemetry conventions closely (matches the existing SqliteOperationSpan pattern for span lifecycle/Drop semantics, reuses RivetError::extract for sanitized error.type, keeps rivetkit-core as the source of truth and the NAPI layer as pure bindings).

Correctness

• OutboundCallInvocation/OutboundCall lifecycle looks sound: finish() consumes/takes the span so a second call is a no-op, and Drop marks an un-finished call as actor.operation_abandoned, mirroring SqliteOperationSpan. The rivetkit-typescript/.../actor-handle.ts action() call site wraps every path (success, retry, destroy short-circuit) in try/finally-style call.finish()/call.finish(error), so I didn't find a leak in the JS call site itself.
• The header-precedence change in outbound-telemetry.ts (callSpan ?? active OTel context ?? invocation span) is consistent with the new parent_context()-based application-span parenting already used for SQL, and the two new driver tests (parents SQLite and outgoing calls under the actor's own application span, preserves vendor trace state across actor calls) exercise it well.
• Retrying via retryOnLifecycleBoundary intentionally reuses one callSpan across all attempts (per the comment), so multiple physical HTTP attempts collapse into one client-side span. That's a reasonable simplification, just noting it means individual retry attempts aren't separately visible in the call span.

Coverage gap worth a second look

• The new call span is only wired into ActorHandleRaw.action() (rivetkit-typescript/packages/rivetkit/src/client/actor-handle.ts:294-321). ActorConn's WebSocket-based action calls (actor-conn.ts:238) and raw ActorHandleRaw.fetch()/#fetchWithResolvedActor (actor-handle.ts:722) still only forward currentActorInvocation and never open a rivet.actor.call span. fetch() in particular is still an HTTP call between actors, so it's an inconsistent trace boundary with .action(). If this is intentionally out of scope (PR description says "Tracing HTTP action calls"), it'd be worth a one-line note/follow-up so it doesn't look like an oversight later.

Test coverage

• Good driver-level (NAPI, integration) coverage for isolation, application-span parenting, and vendor tracestate/invalid-version propagation. There's no Rust-level unit test under rivetkit-core/tests/ exercising start_outbound_call/OutboundCallInvocation (e.g. the dropped/never-finished → actor.operation_abandoned path, or parent_context() == None → None returned) in isolation from the full NAPI+driver stack. Not blocking, but a small core-level test would pin the Drop/finish semantics more cheaply than the e2e suite.

Minor nits

• Import order in client.ts/actor-handle.ts (StartCallSpan before CurrentActorInvocation) is alphabetically reversed vs the rest of the codebase's import style; likely harmless but may trip a formatter/lint pass.
• The new rivet.actor.call span omits the caller's own rivet.actor.id/rivet.actor.name fields that rivet.sqlite.operation spans include (only callee actor_name/action_name are attached). Presumably fine since it nests under the invocation span which already carries caller identity, but flagging in case that's an oversight rather than a deliberate choice.

No security concerns — actor_name/action_name come from the caller's own registry rather than untrusted input, and error text crossing the bridge goes through RivetError::extract/isCanonicalStructuredRivetError for sanitization rather than being trusted verbatim, consistent with the trust-boundary rules in CLAUDE.md.

🤖 Generated with Claude Code

[the-company-company]

🟠 2 medium-severity findings

_{Reviewed commit 384b32b.}

---

🟠 Medium · Keep outbound call tracing available on wasm

Every client created inside a wasm-hosted actor now receives beginOutboundCall, but this adapter always returns undefined. Its action calls therefore continue to propagate the invocation context directly and never emit the new client span, unlike the NAPI runtime.

Expose the core context operation through the wasm binding and forward it here (including its span and finish), rather than using this placeholder, so the portable CoreRuntime contract has the same tracing behavior on both runtimes.

Original location: "rivetkit-typescript/packages/rivetkit/src/registry/wasm-runtime.ts":550 (new side, not submitted inline).

[the-company-company]

🟠 2 medium-severity findings

_{Reviewed commit 7f1d947.}

[the-company-company]

🟠 2 medium-severity findings

_{Reviewed commit 41a28e0.}

[the-company-company]

🟠 2 medium-severity findings

_{Reviewed commit 945ecce.}

[the-company-company]

🟠 2 medium-severity findings

_{Reviewed commit 8395b1b.}

[the-company-company]

🟠 2 medium-severity findings

_{Reviewed commit d485500.}

[the-company-company]

🟠 2 medium-severity findings

_{Reviewed commit f9dcaf3.}

[the-company-company]

🟠 2 medium-severity findings

_{Reviewed commit ae9c32e.}

[the-company-company]

🟠 2 medium-severity findings

_{Reviewed commit 241f2c1.}

[the-company-company]

🟠 2 medium-severity findings

_{Reviewed commit 9d7ac37.}

[the-company-company]

🟠 Medium · Keep outbound call tracing available on wasm

Every actor client now receives beginOutboundCall, but this adapter always returns undefined. Actions issued by wasm-hosted actors therefore still propagate the invocation span directly and never emit the new client span, unlike the NAPI runtime.

Forward the core context operation through the wasm binding (including its span context and finish operation) rather than returning this placeholder so the portable CoreRuntime contract has matching tracing behavior.

[eersnington]

nope. out of scope

[the-company-company]

✅ No issues found

_{Reviewed commit aeeb217.}