Skip to content

Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure

TypeScript

fix(guardrails): quieter pre_call apply-log; distinguish no_decision (review)

Andy Pavia83b19c4b54fa
diff --git a/mcp-host/package-lock.json b/mcp-host/package-lock.json
index 448e7c438..a14759f56 100644
--- a/mcp-host/package-lock.json
+++ b/mcp-host/package-lock.json
@@ -1,12 +1,12 @@
{
"name": "clerum-mcp-host",
- "version": "0.1.261",
+ "version": "0.1.262",
"lockfileVersion": 3,
"requires": true,
"packages": {
"": {
"name": "clerum-mcp-host",
- "version": "0.1.261",
+ "version": "0.1.262",
"license": "SEE LICENSE IN LICENSE",
"dependencies": {
"@anthropic-ai/sdk": "^0.32.0",
diff --git a/mcp-host/package.json b/mcp-host/package.json
index da2032a86..e997973b4 100644
--- a/mcp-host/package.json
+++ b/mcp-host/package.json
@@ -1,6 +1,6 @@
{
"name": "clerum-mcp-host",
- "version": "0.1.261",
+ "version": "0.1.262",
"description": "MCP Host - reads Host CRD and provides LLM access via OpenAI/Claude",
"main": "dist/main.js",
"engines": {
diff --git a/mcp-host/src/core/adapters/hookedLlmPort.ts b/mcp-host/src/core/adapters/hookedLlmPort.ts
index fbf889a68..daae663b4 100644
--- a/mcp-host/src/core/adapters/hookedLlmPort.ts
+++ b/mcp-host/src/core/adapters/hookedLlmPort.ts
@@ -163,17 +163,23 @@ export class HookedLlmPort implements LlmPort {
// and a reject blocks the call outright (no tokens spent).
const preContribs: Array<Contributor<ToolCompletionRequest, ToolCompletionResponse>> = []
for (const hook of this.hooks.preCall) {
- const before = approxRequestChars(req)
+ const preReq = req
const c = await hook.preCall(req)
if (c) {
preContribs.push(c)
if (c.rewrite) req = c.rewrite // already system-role-stripped + capability-gated (N4/F4)
}
- // Diagnostic (content-free, sizes only per §8.4): make the dial AND whether
- // the rewrite was actually applied observable, so a may_rewrite hook's
- // compaction can be confirmed end-to-end against the hook's own logs — the
- // `before→after` here should match the hook's reported before/after chars.
+ // Diagnostic (content-free, sizes only per §8.4): log only the meaningful,
+ // low-frequency outcomes — a rewrite that was APPLIED (with before→after
+ // chars, which should match the hook's own reported before/after), a deny,
+ // or a non-standard contribution (e.g. a reject downgraded to no_decision
+ // when the hook lacks may_deny). The common no-op — a hook that returned
+ // `continue` with nothing to do — is deliberately NOT logged: it fires on
+ // every tool-loop iteration, and the hook's own logs already record the
+ // dial. Sizes are measured only when a rewrite lands, so the no-op path
+ // stays free.
if (c?.rewrite) {
+ const before = approxRequestChars(preReq)
const after = approxRequestChars(req)
console.log(
`[HookedLlmPort] pre_call ${hook.id}: applied rewrite ${before}→${after} chars (Δ${before - after})`
@@ -182,8 +188,8 @@ export class HookedLlmPort implements LlmPort {
console.log(
`[HookedLlmPort] pre_call ${hook.id}: deny (reason=${logSafeReason(c.reasonCode)})`
)
- } else {
- console.log(`[HookedLlmPort] pre_call ${hook.id}: no rewrite (continue), ${before} chars`)
+ } else if (c) {
+ console.log(`[HookedLlmPort] pre_call ${hook.id}: ${c.decision} (no rewrite)`)
}
}
if (aggregateDecision(preContribs) === 'deny') return this.denyRefusal(preContribs, 'pre_call')