Summary

• Problem: Issue [Bug]: anthropic-vertex-provider adds cache_control to active-memory system block — triggers "Found 5" error when active-memory is enabled #91982 reports that anthropic-vertex requests are intermittently rejected with FailoverError: LLM request rejected: A maximum of 4 blocks with cache_control may be provided. Found 5. In the reporter's production deployment the rejection fires whenever active-memory recall injects context into the turn, forcing a model fallback on every hit; the only workaround is disabling active-memory.
• Root cause: the request passes through two cache-control writers. The shared anthropic-messages transport (src/llm/providers/anthropic.ts) builds the payload with full marker budgeting: it splits the system prompt at the cache boundary into a stable prefix (with cache_control) and an intentionally uncached dynamic suffix, marks the tool array, and gives the message pass only the remaining budget (4 - system - tools), so legitimate payloads can carry exactly four markers. The vertex plugin then ran applyAnthropicPayloadPolicyToParams again on that finished payload inside its onPayload hook (createAnthropicVertexOnPayload in extensions/anthropic-vertex/stream-runtime.ts). That helper is written for raw payloads (the agents-side anthropic-transport-stream.ts builds its system blocks with no markers and boundary text intact): any system text block with no boundary marker and no cache_control gets a marker added unconditionally. Post-transport, the dynamic suffix is exactly such a block, so the shim adds a fifth marker and Anthropic rejects the request. Turns that don't fill the budget stay at four and pass, which is why the failure tracks active-memory hits rather than every request.
• Provenance: the shim was added in 1a13c34f5b ("close cache boundary transport gaps"), when the then-external transport did not understand the OpenClaw cache boundary and the plugin-level policy pass was the only splitter. eef24d452f ("preserve provider prompt cache boundaries", shipped in v2026.6.2-beta.1) moved boundary splitting and marker budgeting natively into the shared transport, which turned the shim from a gap-closer into a double-application. The reporter's first failures on 2026.6.5 line up with that release.
• Fix: delete the shim and forward the caller's onPayload hook unchanged. The shared transport is the single owner of cache-control budgeting for this API family, which is already how the plain anthropic provider behaves (its wrapper applies the payload policy only for service_tier, without enableCacheControl).
• What changed: extensions/anthropic-vertex/stream-runtime.ts removes createAnthropicVertexOnPayload and its policy imports and passes options?.onPayload straight through (net −35 lines); the stale tests that encoded the pre-split payload shape are replaced with regression tests for the budgeted-payload invariant.
• What did NOT change (scope boundary): config surface unchanged (no schema, defaults, doctor migrations, or docs/reference/config). Plugin surface unchanged (no exported signature, manifest, api.ts, or SDK changes; createAnthropicVertexStreamFn keeps its signature). The shared transport, the agents-side anthropic transport (which builds raw payloads and correctly applies the policy once), and the anthropic plugin's service-tier wrapper are untouched.

Reproduction

1. Configure anthropic-vertex as the primary provider with prompt caching at its defaults, and enable active-memory for the agent.
2. Drive a turn whose payload legitimately uses the full marker budget — an agentic turn where the transport marks the stable system prefix, the tool array, the last user text block, and the trailing tool_result fallback (the reporter hits this whenever a memory recall injects context).
3. The system prompt contains the cache boundary, so the transport emits a cached stable prefix plus an uncached dynamic suffix.
4. Before this PR: the vertex onPayload shim re-runs the payload policy on the finished payload, adds cache_control to the dynamic suffix, and Vertex StreamRawPredict returns 400 A maximum of 4 blocks with cache_control may be provided. Found 5., triggering a model fallback.
5. After this PR: the payload reaches the wire exactly as the transport budgeted it (at most four markers, dynamic suffix uncached) and the request is accepted.

Real behavior proof

Behavior addressed (#91982): an anthropic-vertex payload that already carries the transport's full cache-control budget is no longer re-marked on its way out, so requests that previously breached Anthropic's four-marker cap with Found 5 now keep exactly the budgeted markers and the dynamic system suffix stays uncached.

Real environment tested (Linux x64, Node v22.22.3 — Vitest against the production vertex stream runtime): createAnthropicVertexStreamFn from the shipped plugin module, exercised through its injectable transport seam; the regression payload mirrors the shared transport's real output shape (split system prefix/suffix, marked tools, marked user text, marked trailing tool_result).

Exact steps or command run after this patch: pnpm test extensions/anthropic-vertex/stream-runtime.test.ts; pnpm tsgo:extensions && pnpm tsgo:extensions:test; node scripts/run-oxlint.mjs and format check on the changed files; .agents/skills/autoreview/scripts/autoreview --engine claude --thinking claude=max.

Evidence after fix (Vitest output for the touched test file):

 Test Files  1 passed (1)
      Tests  25 passed (25)

Observed result after fix: the transport options forward the caller's payload hook unchanged; a payload carrying the transport's four budgeted markers keeps exactly four after the hook path runs, with the dynamic suffix block still uncached; when the caller supplies no hook, the transport sets none.

What was not tested: a live Vertex AI StreamRawPredict round-trip against GCP (no Vertex credentials in this environment). The reporter's GCP Cloud Logging captures in #91982 document the live failure mode this removes.

Repro confirmation: both new tests fail on the unpatched tree — the marker count comes back as expected 5 to be 4, the same five-marker breach the reporter logged, and a transport-injected hook appears where none was supplied — and pass with the fix, so the tests cover the production change.

Risk / Mitigation

• Risk: vertex payloads could lose prompt caching if nothing re-applied markers. Mitigation: the shared anthropic-messages transport applies cache-control budgeting for every request on this API family before onPayload runs, and the vertex stream function routes all requests through it — the same single-owner arrangement the plain anthropic provider already ships with.
• Risk: a downstream onPayload consumer relied on the shim re-shaping its returned payload. Mitigation: every runtime supplier was audited (anthropic payload logging, command delivery logging, diagnostics byte accounting, and the before_provider_request plugin hook); all of them inspect or patch the finished payload and none construct boundary-marked system arrays, and the plain anthropic path likewise sends hook results without re-policing.

Change Type (select all)

• Bug fix

Scope (select all touched areas)

• Integrations

Linked Issue/PR

Fixes #91982