You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[v2][feature] Per-tool-call audit log — refile of #1554 (still applies to v2, measured) #2566
#1554 asked for every tool call to be recorded in an audit log. It was closed during the pre-v2 backlog clearing with "if it still applies to v2, please open a new issue" — it still does. Measured on self-host 2.0.0-beta.37 (2026-11-14, live instance):
The gap, measured on v2 self-host
No endpoint exposes ad-hoc tool-call history. The product API (103 routes) covers scheduled-runs and workflow-runs, but an agent run that calls a tool through /mcp leaves nothing queryable afterwards. The two questions from Record every tool call in an audit log #1554 remain unanswerable: what did this agent touch, and what did my policies actually stop?
Traces carry the timing but not the semantics. A tools/call produces spans like http.server POST → auth.authenticate → McpServer.@effect/mcp/2025-03-26/tools/call with duration and status, but every span arrives with attributes: {} — no tool name, no connection/account identity, no outcome. Pointing OTEL_EXPORTER_OTLP_ENDPOINT at a collector just moves the same empty spans elsewhere.
The ask (either or both)
Persist per-tool-call rows — tool path, app, connection/account, caller, outcome (ok | fail | blocked), duration — with a query API. Record every tool call in an audit log #1554's design (a row written at the execute chokepoint, which every plugin kind and host passes through) still looks right, and @midego1's PR branch from that thread may be a starting point.
At minimum, populate span attributes (tool, app, connection, outcome) so operators can ship their own audit pipeline via OTel without waiting for a first-class log.
Context
We run Executor self-hosted as the integration mounting point for another product, and had to build an MCP proxy in front of/mcp specifically to produce this audit trail — posting here as a signal of real demand. Happy to help test a implementation.
Related: #2548 (per-app logs tab) — complementary UI lane, not a substitute for a queryable per-call ledger.
Activity