github-mirror/n8n
1
0
Fork 0
mirror of https://github.com/n8n-io/n8n.git synced 2026-10-17 13:39:44 +02:00
Code Issues Actions 14 Packages Projects Releases Wiki Activity Graph

docs(core): Trim narration comments in agent session-log utils (no-changelog)

Browse Source
This commit is contained in:
Rob 2026-09-29 15:07:16 +01:00
parent e0648c3e62
commit 504bbfc0e9
2 changed files with 17 additions and 35 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

49
packages/cli/src/modules/mcp/tools/agents/agent-session-log.utils.ts
View File

@ -16,11 +16,7 @@ export const DEFAULT_TIMELINE_EVENT_CHAR_BUDGET = 2_000;
/** Character cap for free-text fields on execution summaries (userMessage, error). */
export const SUMMARY_FIELD_CHAR_BUDGET = 1_000;
/**
* A timeline event shaped for the MCP response. Values over the character
* budget are replaced by a truncated string preview, so the shape is looser
* than {@link TimelineEvent}.
*/
/** Looser than {@link TimelineEvent}: over-budget values become string previews. */
export type ShapedTimelineEvent = Record<string, unknown>;
type ToolCallEvent = Extract<TimelineEvent, { type: 'tool-call' }>;
@ -38,15 +34,11 @@ function isSoftFailure(event: ToolCallEvent): boolean {
}
/**
* True when a closed tool-call event counts as a failure. Covers hard
* failures (`success: false`) and soft failures (a workflow tool or sub-agent
* delegation that reported an error in its output). Declined approvals are
* not failures.
*
* Mirrors the failure rule in `computeExecutionFailureSummary`
* (`@/modules/agents/utils/execution-failure-summary`). A drift-pinning test
* in `agent-session-log.utils.test.ts` compares this filter's count against
* that summary, so the two rules cannot diverge silently.
* True for hard failures (`success: false`) and soft failures (a workflow
* tool or sub-agent delegation that reported an error in its output);
* declined approvals are not failures. Mirrors the rule in the agents
* module's `computeExecutionFailureSummary`; a drift-pinning test compares
* the two so they cannot diverge silently.
*/
export function isFailedToolCallEvent(event: TimelineEvent): event is ToolCallEvent {
return (
@ -66,13 +58,10 @@ function capAndScrubText(value: string, budget: number): string {
}
/**
* Applies the per-event character budget to every payload-bearing field of a
* timeline. Returns the shaped events and whether anything was cut, so the
* tool can tell the client to raise `truncate` when it needs full payloads.
*
* Tool payloads, suspensions, and HITL responses are secret-scrubbed at
* record time; model-generated `text` and `reasoning` content is not, so this
* boundary scrubs it before it leaves over MCP.
* Caps every payload-bearing field of a timeline at `budget` chars and says
* whether anything was cut. Also scrubs secrets: tool payloads are scrubbed
* at record time, but model-generated text and reasoning content is not, so
* it must not leave over MCP unscrubbed.
*/
export function truncateTimeline(
events: TimelineEvent[],
@ -87,9 +76,7 @@ export function truncateTimeline(
return scrubbed.slice(0, budget) + truncationMarker(scrubbed.length - budget);
};
// Caps a value of unknown shape at `budget` characters of its JSON form. A
// value over budget becomes its truncated JSON string, so the client always
// sees a preview instead of an oversized payload.
// An over-budget value becomes a truncated preview of its JSON form.
const capValue = (value: unknown): unknown => {
if (value === undefined || value === null) return value;
if (typeof value === 'string') return capText(value);
@ -140,10 +127,9 @@ export function truncateTimeline(
}
/**
* Compact per-execution summary shared by get_agent_session and
* get_agent_execution. Free-text fields are scrubbed and capped so a long
* prompt or provider error cannot blow the response budget. The `timeline`
* column is deliberately never copied here.
* Free-text fields are scrubbed and capped so a long prompt or provider
* error cannot blow the response budget; `timeline` is deliberately never
* copied here.
*/
export function toExecutionSummary(execution: AgentExecution) {
return {
@ -170,10 +156,9 @@ export function toExecutionSummary(execution: AgentExecution) {
}
/**
* True when an execution's timeline lives in a blob store but could not be
* read (missing, corrupted, or an unconfigured storage location). The detail
* read degrades to a null timeline in that case, and the tool must report it
* instead of presenting a clean empty run.
* True when a blob-stored timeline could not be read (missing, corrupted, or
* an unconfigured location) — the read degrades to null, which must not be
* presented as a clean empty run.
*/
export function isTimelineUnavailable(execution: AgentExecution): boolean {
return execution.storedAt !== 'db' && execution.timeline === null;

3
packages/cli/src/modules/mcp/tools/agents/agent-tools.service.ts
View File

@ -1143,9 +1143,6 @@ export class McpAgentToolsService {
const summary = toExecutionSummary(execution);
if (!input.includeTimeline) return { ok: true, execution: summary };
// A blob-stored timeline can be unreadable (missing, corrupted, or
// an unconfigured location); the detail read degrades to null. Say
// so instead of presenting a clean empty run.
if (isTimelineUnavailable(execution)) {
return {
ok: true,