Files
AgentHub/server/lib/event-payload.js
T
nntrivi2001 4aae61297f feat: analytics parity — tool cost, date ranges, budgets, query explorer, replay, snapshots (v2.4.0)
Analytics parity phases 2–7 (docs/plans/analytics-parity/), built on
138e706 and ported onto v2.3.1:
- Cost by AI tool (?tool= on cost/analytics/stats/workflows, by_tool,
  Codex fast/priority and flex tiers, agenthub_cost_usd_total{tool,model})
- Date ranges (from/to) on cost/analytics/workflows; Analytics range select
  with Today / This week (Monday) / This month (1st) / 7d / 30d / 90d / All /
  Custom; workflow filters carried into workflow-stats.js (read worker)
- Budgets and cost alerts as alert rule types (alert_rules rebuilt without
  its rule_type CHECK), GET /api/budgets/status, rule preview, spend pace,
  status-bar budget chip; rules stay editable in place
- Query Explorer (/query, allowlisted sort, CSV/JSON export) and scheduled
  reports (report_definitions/report_runs, Settings → Reports)
- Session Replay tab and immutable snapshots (session_snapshots, /snapshot/:token)
- Remote Codex sync, push ingest with a hashed token, privacy redaction and
  ignored folders (Settings → Data), Cursor sessions (no cost)
- Read-only MCP server, statusline, Grafana/Prometheus files

Also: resizable sidebar, pet run/wave animations, the workspace home keeps a
half-typed prompt, and the pre-commit hook runs only the related tests for a
small commit.
2026-10-06 17:12:28 +07:00

174 lines
6.6 KiB
JavaScript

/**
* @file event-payload.js
* @description Trims a hook payload before it is stored as `events.data`. A
* PostToolUse payload carries whole files — Edit's `originalFile`, Read's file
* `content`/`base64`, Write's `content` — and long command output, none of which
* any analysis reads: the server reads only `tool_input.skill`, the event views
* only preview. The full text stays in the session's transcript (and AgentHub's
* snapshot of it under `<data dir>/transcripts/`), which the Conversation tab
* reads. Kept: every field outside `tool_input`/`tool_response`, every number and
* flag, `structuredPatch` (the +/- line counts), and each long string's head and
* tail. Dropped fields are named in `_trimmed`, so a view can say so. When
* the user turned on secret redaction (`server/lib/privacy.js`), the payload is
* redacted first, before anything is cut.
* @author Nguyễn Ngọc Trí Vĩ <vinnt@smartgift.vn>
*/
const { redactEventJson } = require("./privacy");
// Strings inside tool_input / tool_response longer than this keep head + tail.
const STR_CAP = 8192;
const HEAD = 6000;
const TAIL = 2000;
// A payload this short cannot hold anything worth trimming, so it is not parsed.
const SKIP_UNDER = STR_CAP;
// Per tool, the tool_response fields that mirror a whole file or repeat
// tool_input. Other tools' fields (Grep's `content` is its matches) only get capped.
const EDIT_DROP = ["originalFile", "oldString", "newString"];
const DROP_RESPONSE = {
Edit: EDIT_DROP,
MultiEdit: EDIT_DROP,
NotebookEdit: EDIT_DROP,
Write: [...EDIT_DROP, "content"],
};
const DROP_RESPONSE_FILE = { Read: ["content", "base64"] };
function capString(s, hit) {
if (s.length <= STR_CAP) return s;
hit.n++;
return `${s.slice(0, HEAD)}\n…[${s.length - HEAD - TAIL} chars trimmed]…\n${s.slice(-TAIL)}`;
}
// Caps every string at any depth; `structuredPatch` is left whole.
function capDeep(v, key, hit) {
if (typeof v === "string") return capString(v, hit);
if (key === "structuredPatch" || v === null || typeof v !== "object") return v;
if (Array.isArray(v)) return v.map((x) => capDeep(x, undefined, hit));
const out = {};
for (const [k, x] of Object.entries(v)) out[k] = capDeep(x, k, hit);
return out;
}
/**
* @param {object} data A parsed hook payload.
* @param {{n: number}} [hit] Counts every drop and cap made, so a caller can tell
* "nothing to trim" apart.
* @returns {object} A trimmed copy (the input is not mutated).
*/
function trimHookPayload(data, hit = { n: 0 }) {
if (!data || typeof data !== "object" || Array.isArray(data)) return data;
const out = { ...data };
const dropped = new Set(Array.isArray(data._trimmed) ? data._trimmed : []);
const res = out.tool_response;
if (res && typeof res === "object" && !Array.isArray(res)) {
const r = { ...res };
for (const k of DROP_RESPONSE[data.tool_name] || []) {
if (typeof r[k] === "string" && r[k].length > 0) {
delete r[k];
hit.n++;
dropped.add(`tool_response.${k}`);
}
}
const fileDrop = DROP_RESPONSE_FILE[data.tool_name];
if (fileDrop && r.file && typeof r.file === "object") {
const f = { ...r.file };
for (const k of fileDrop) {
if (typeof f[k] === "string" && f[k].length > 0) {
delete f[k];
hit.n++;
dropped.add(`tool_response.file.${k}`);
}
}
r.file = f;
}
out.tool_response = r;
}
if ("tool_response" in out) out.tool_response = capDeep(out.tool_response, undefined, hit);
if ("tool_input" in out) out.tool_input = capDeep(out.tool_input, undefined, hit);
if (dropped.size > 0) out._trimmed = [...dropped];
return out;
}
/**
* The SQL-side entry point (`trim_event_data(?)`, registered in `server/db.js`):
* a JSON string in, a JSON string out. Short, null or unparseable input is
* returned unchanged, and so is a payload with nothing to trim — re-serializing
* it would round integers past 2^53. Trimming is idempotent.
* @param {string|null} json
* @returns {string|null}
*/
function trimEventData(json) {
// Opt-in secret redaction runs first, on the whole text: trimmed first, a
// private key cut in two would no longer match its pattern.
json = redactEventJson(json);
if (typeof json !== "string" || json.length <= SKIP_UNDER) return json;
let parsed;
try {
parsed = JSON.parse(json);
} catch {
return json;
}
if (!parsed || typeof parsed !== "object" || Array.isArray(parsed)) return json;
const hit = { n: 0 };
const out = trimHookPayload(parsed, hit);
// ponytail: a payload that IS trimmed is re-serialized whole, so a huge integer
// elsewhere in it still rounds; splice the JSON text instead if one ever matters.
return hit.n ? JSON.stringify(out) : json;
}
/**
* Trims the rows stored before trimming existed, a few rows per write
* transaction so a hook never waits long for the lock: ids are picked first and
* each row's payload is read alone, and a transaction ends after `chunk` rows or
* once the payloads it read pass `maxBytes` — a Read of an image can hold MBs of
* base64, and 200 of those in one go stalled the event loop. Idempotent: a crash
* just resumes from the start on the next run.
* @param {import("better-sqlite3").Database} db
* @param {{chunk?: number, maxBytes?: number}} [opts]
* @returns {Promise<{rows: number, bytesSaved: number}>}
*/
function sweepEventPayloads(db, { chunk = 200, maxBytes = 8 * 1024 * 1024 } = {}) {
const pick = db
.prepare("SELECT id FROM events WHERE id > ? AND length(data) > ? ORDER BY id LIMIT ?")
.pluck();
const read = db.prepare("SELECT data FROM events WHERE id = ?").pluck();
const update = db.prepare("UPDATE events SET data = ? WHERE id = ?");
const out = { rows: 0, bytesSaved: 0 };
// Returns the last id it handled; the next pick starts after it.
const trimChunk = db.transaction((ids) => {
let bytes = 0;
let last = ids[0];
for (const id of ids) {
last = id;
const data = read.get(id);
if (typeof data !== "string") continue;
bytes += data.length;
const t = trimEventData(data);
if (t !== data) {
update.run(t, id);
out.rows++;
out.bytesSaved += data.length - t.length;
}
if (bytes >= maxBytes) break;
}
return last;
});
let lastId = 0;
return new Promise((resolve, reject) => {
const step = () => {
try {
const ids = pick.all(lastId, SKIP_UNDER, chunk);
if (ids.length === 0) return resolve(out);
lastId = trimChunk(ids);
setImmediate(step);
} catch (err) {
reject(err);
}
};
step();
});
}
module.exports = { trimHookPayload, trimEventData, sweepEventPayloads, STR_CAP };