Repository navigation
Conversation
makeDirectModeHandler forwarded upstream results without calling applyOutputSanitisation or spotlightForwarded, so redact, block, strip_control_chars and spotlight_untrusted never applied to /mcp/all. Call both from the direct handler the same way the call_tool_* handlers do: sanitise the raw result before the activity classification and truncation, and spotlight after truncation, refreshing the recorded response text. Fixes smart-mcp-proxy#1683
|
Codecov Report✅ All modified and coverable lines are covered by tests. 📢 Thoughts on this report? Let us know! |
Direct mode (/mcp/all, server__tool) now runs redact/block/strip/spotlight (smart-mcp-proxy#1683), so drop it from the not-covered list.
Dumbris
enabled auto-merge (squash)
October 20, 2026 04:46
Member
|
Thank you, @james-crowley! Great find, and a careful fix. The direct handler now runs sanitisation in exactly the same order as On the docs question: the page lives in this repo ( One note for later, not a blocker: on a |
Dumbris
added a commit
that referenced
this pull request
Oct 20, 2026
… block Direct routing (#1685) returned applyOutputSanitisation's block result before emitting the call's completion. It now emits the same error tool_call (error_class proxy_policy, block message only) as the call_tool_* paths. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Dumbris
added a commit
that referenced
this pull request
Oct 20, 2026
…ks a response (#1687) ## Summary When `output_sanitisation` blocks a response (`response_action: block`), the call_tool_* handlers returned before recording the call's completion. The upstream call **did** run (side effects may have happened), yet only a `policy_decision` was persisted: no `tool_call` activity row (started events are not persisted), an SSE `started` event with no completion, and no tool-call history entry. `handleCallToolVariant`, the legacy `handleCallTool` and direct routing mode (`makeDirectModeHandler`, added by #1685) now close the call on a block: - **Tool-call history:** the attempt is stored with `Error` set to the block message and `Response` cleared, so the blocked payload is never stored. - **Activity:** a `tool_call` completion is emitted with status `error`, `error_class: proxy_policy` and `fault_domain: proxy`. It carries the fixed block message, never the response. On call_tool_* paths the `internal_tool_call` record is emitted as well. - The new helper `recordOutputBlocked` follows the convention of the code_execution sub-call block path, which already closed its record this way. ## Counting / audit check - **Spec 107 audit:** the block's `policy_decision` already writes `tool_call outcome:blocked reason:output_sanitisation` (`isPostDispatchBlockKey`). The completion's audit write is the existing deduped no-op (`toolCallWritten`), so the audit log still has one line per attempt. - **Telemetry:** nothing is double-counted. The availability block counter comes only from `policy_decision`, the upstream-attempt bucket was counted at dispatch, and activation is not recorded. - **Activity/Usage call totals (`storage.CountsAsCall`):** the `policy_decision` (blocked) and the new `tool_call` (error) both count as failed calls. One blocked attempt therefore counts as two failed calls in those totals. The code_execution sanitisation block and the generation-changed refusal pair already counted this way. Separating them needs the post-dispatch block key on the `policy_decision` record, so that is left for a follow-up. ## Direct mode After #1685 merged, origin/main was merged into this branch and the direct-mode block now emits the same error `tool_call` completion. Direct mode keeps no tool-call history. ## Tests - `TestCallToolVariant_SanitisationBlockRecordsToolCallCompletion` and `TestLegacyCallTool_SanitisationBlockRecordsToolCallCompletion` drive a real in-process upstream that returns an AWS key with block mode on. Each asserts one persisted `tool_call` with status `error` and `error_class` `proxy_policy`, the `policy_decision` beside it, and the secret absent from every activity record. `TestDirectMode_SanitisationBlockRecordsToolCallCompletion` asserts the same for direct mode. All three failed before the fix (no `tool_call` record). - `TestRecordOutputBlocked_HistoryCarriesBlockMessageNotResponse` asserts the history entry carries the block message and no response. The direct-routing fixture's incarnation guard drops all history writes, so this is tested on the helper. - `go test -race ./internal/server/` with the unit-tests.yml `-skip` regex: pass. - `golangci-lint` v2 (`.github/.golangci.yml`), bare and `--build-tags server`, over `./internal/server/...`: no issues in changed files. Six existing staticcheck findings remain in untouched test files. 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description
Fixes #1683.
output_sanitisation(redact, block, strip control characters, spotlight) never ran for direct routing mode (/mcp/all,server__toolnames). The direct handler (makeDirectModeHandlerinmcp_routing.go) forwarded the upstream result without callingapplyOutputSanitisationorspotlightForwarded, which thecall_tool_*handlers andcode_executionalready call.This adds both calls to the direct handler, mirroring
handleCallToolVariant:applyOutputSanitisationruns on the raw result before the activity classification and truncation. A non-nil result is a block and is returned as-is.spotlightForwardedruns after truncation, andresponseTextis refreshed withforwardedTextso the recorded activity matches what the agent receives.With no
output_sanitisationconfigured both calls are no-ops, so default behavior is byte-for-byte unchanged. Error results (isError: true) and non-text blocks keep the existing exemptions insideapplyOutputSanitisation.I did not touch the docs page that lists direct mode under "Scope and limits"; that is in the separate docs site and I wasn't sure of the process there. Happy to add that edit here if you want it.
Testing
Verification:
internal/server/direct_output_sanitisation_test.godrives the handler mcp-go registered (directRoutingFixture) against a real in-process upstream: redact, block, spotlight, strip, and the opt-out default. The first four fail onmainand pass with this change. I added a smallokBodyoverride tocountingUpstreaminscope_fixture_test.goso an upstream can return a chosen payload.go test ./internal/server/passes with the e2e tests skipped (-skip 'TestE2E|TestBinary|TestMCPProtocol|TestCodeExecClientModeE2E'), because those need a built binary andnpxin my environment. CI will cover them.output_sanitisation: {spotlight_untrusted, response_action: redact, strip_control_chars}:/mcpcall_tool_readreturned[REDACTED:cloud_credentials]wrapped in«untrusted:...», while/mcp/allreturned the raw key. This patch is meant to make the second match the first.