WEBDIENSTE/openswarm
1
0
Fork 0
mirror of https://github.com/openswarm-ai/openswarm.git synced 2026-09-13 21:27:41 +02:00
Code Issues Packages Projects Releases Wiki Activity

[eric] router: always-on history pruning clears old tool-result bodies on the wire, newest kept verbatim

Browse Source
This commit is contained in:
ciregenz
2026-08-31 15:54:13 -07:00
parent b9a63b3b80
commit 72fa1b4ecc
3 changed files with 337 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
backend/apps/agents/9router_gpt5_patch.js
+32 -3
View File
@@ -98,7 +98,32 @@ const _http = require('http');
} catch (_) {}
})();
const TARGET_HOSTS = new Set(['api.openai.com']);
const TARGET_HOSTS = new Set(['api.openai.com', 'api.anthropic.com']);
// ENG-418: continuous history pruning rides the same interceptor. Loaded lazily and fail-open so
// a missing or broken module costs the prune, never the request.
let _prune = null;
function historyPrune(bodyStr) {
try {
if (_prune === null) {
_prune = require(require('path').join(__dirname, '9router_history_prune.js'));
try { process.stderr.write('[history-prune] installed\n'); } catch (_) {}
}
return _prune.maybePrune(bodyStr);
} catch (_) {
_prune = { maybePrune: (b) => b };
try { process.stderr.write('[history-prune] FAILED to load; requests pass through unpruned\n'); } catch (_) {}
return bodyStr;
}
}
function transformBody(bodyStr, host) {
// Order matters only in that both must see valid JSON; each is a no-op off its own shape.
let out = bodyStr;
if (host === 'api.openai.com') out = maybeRewriteBody(out);
out = historyPrune(out);
return out;
}
const DEBUG = process.env.OPENSWARM_DEBUG_GPT5_PATCH === '1';
function _log(msg) {
@@ -169,6 +194,7 @@ function patchHttpRequest(orig) {
if (!TARGET_HOSTS.has(host)) {
return orig.apply(this, args);
}
const targetHost = host;
let req;
try { req = orig.apply(this, args); } catch (e) { throw e; }
@@ -217,7 +243,7 @@ function patchHttpRequest(orig) {
let bodyStr = '';
if (isStringMode === true) bodyStr = chunks.join('');
else if (isStringMode === false) bodyStr = Buffer.concat(chunks).toString('utf8');
const rewritten = maybeRewriteBody(bodyStr);
const rewritten = transformBody(bodyStr, targetHost);
if (rewritten !== bodyStr) {
const newBuf = Buffer.from(rewritten, 'utf8');
try {
@@ -270,7 +296,7 @@ if (typeof globalThis.fetch === 'function' && !globalThis.fetch.__openswarm_gpt5
try { host = new URL(url).hostname.toLowerCase(); } catch (_) { return origFetch.call(this, input, init); }
if (!TARGET_HOSTS.has(host)) return origFetch.call(this, input, init);
if (init && typeof init.body === 'string') {
const rewritten = maybeRewriteBody(init.body);
const rewritten = transformBody(init.body, host);
if (rewritten !== init.body) {
const newInit = Object.assign({}, init, { body: rewritten });
const newLen = String(Buffer.byteLength(rewritten, 'utf8'));
@@ -300,3 +326,6 @@ if (typeof globalThis.fetch === 'function' && !globalThis.fetch.__openswarm_gpt5
_log('fetch install failed: ' + (e && e.message ? e.message : String(e)));
}
}
// Exported for the wiring test only; --require ignores exports.
module.exports = { transformBody };
backend/apps/agents/9router_history_prune.js
+128
View File
@@ -0,0 +1,128 @@
// Always-on history pruning at the one wire we own (ENG-418 item 2, the hermes lift).
//
// The CLI resends the whole transcript every request and its own KEEP-RECENT microcompact is
// feature-gated off with no override, so old tool results pile up until the autocompact cliff
// (ENG-385: 100-292 tool-call chats dying at 80-160K). This module runs inside 9router via the
// same `node --require` patch that fixes GPT-5 max_tokens, and clears the BODIES of old
// tool_result blocks on the wire only: the CLI's transcript on disk stays complete, so nothing
// is destroyed, and a fresh session or a resume still has every byte.
//
// Rules, each one earned:
// - Newest KEEP_RECENT big results stay verbatim (the CLI's own microcompact keeps 5).
// - Small results are never touched: `git status` says "nothing to commit" in 200 bytes and that
// whole content IS the answer (the OmniRoute lesson: ratio bought by deleting answers is a sin).
// - Below ENGAGE_BYTES the body passes through BYTE-IDENTICAL, so short sessions keep their
// prompt-cache prefix and this module is provably a no-op where context is not scarce.
// - The stub is PASSIVE and carries no path and no instruction: an imperative inside tool output
// reads as an injection attempt and manufactures a security warning (drilled 2026-08-27).
// - Exact duplicates of a newer big result collapse even inside the keep window (hermes's rule).
// - Only tool_result CONTENT is replaced; ids, roles, ordering, assistant blocks and thinking
// signatures are untouched, so the API contract (every tool_use answered) cannot break.
// - Any parse trouble returns the body unchanged: fail-open, availability first.
'use strict';
const KEEP_RECENT = 5;
const MIN_STUB_BYTES = 2000;
const ENGAGE_BYTES = 300000;
const OFF = String(process.env.OSW_HISTORY_PRUNE || '').toLowerCase() === 'off';
const DEBUG = process.env.OSW_HISTORY_PRUNE_DEBUG === '1';
function log(msg) {
try { process.stderr.write('[history-prune] ' + msg + '\n'); } catch (_) {}
}
function blockText(block) {
// tool_result content is a string or a list of blocks; concatenate the text parts.
const c = block.content;
if (typeof c === 'string') return c;
if (Array.isArray(c)) {
let out = '';
for (const p of c) {
if (p && p.type === 'text' && typeof p.text === 'string') out += p.text;
}
return out;
}
return '';
}
function blockHasImage(block) {
const c = block.content;
if (!Array.isArray(c)) return false;
for (const p of c) if (p && p.type === 'image') return true;
return false;
}
function stubFor(block, reason) {
const n = blockText(block).length;
const img = blockHasImage(block);
let text;
if (reason === 'duplicate') {
text = '[Duplicate tool output: same content as a more recent result.]';
} else if (img && n < MIN_STUB_BYTES) {
text = '[Old screenshot cleared by OpenSwarm to keep this long chat inside the model\'s context window.]';
} else {
text = '[Old tool output cleared by OpenSwarm to keep this long chat inside the model\'s context window: '
+ n + ' characters from an earlier step' + (img ? ', plus a screenshot' : '') + '.]';
}
return [{ type: 'text', text }];
}
// bodyStr -> { body: string, stats } ; body === input means untouched.
function pruneBody(bodyStr) {
const none = { body: bodyStr, stats: null };
if (OFF) return none;
if (typeof bodyStr !== 'string' || bodyStr.length < ENGAGE_BYTES) return none;
let data;
try { data = JSON.parse(bodyStr); } catch (_) { return none; }
if (!data || !Array.isArray(data.messages)) return none;
// Collect every prunable tool_result in transcript order.
const candidates = [];
for (const msg of data.messages) {
if (!msg || msg.role !== 'user' || !Array.isArray(msg.content)) continue;
for (const block of msg.content) {
if (!block || block.type !== 'tool_result') continue;
const size = blockText(block).length;
if (size >= MIN_STUB_BYTES || blockHasImage(block)) candidates.push(block);
}
}
if (candidates.length <= KEEP_RECENT) return none;
const keepFrom = candidates.length - KEEP_RECENT;
const seenNewestFirst = new Set();
// Walk newest-first so a duplicate always collapses toward its most recent copy.
for (let i = candidates.length - 1; i >= 0; i--) {
const b = candidates[i];
const text = blockText(b);
if (i >= keepFrom) {
if (text) seenNewestFirst.add(text);
continue;
}
b.content = stubFor(b, seenNewestFirst.has(text) ? 'duplicate' : 'old');
}
let out;
try { out = JSON.stringify(data); } catch (_) { return none; }
const stats = {
stubbed: keepFrom,
kept: KEEP_RECENT,
savedBytes: bodyStr.length - out.length,
};
return { body: out, stats };
}
function maybePrune(bodyStr) {
try {
const { body, stats } = pruneBody(bodyStr);
if (stats && DEBUG) {
log('engaged: stubbed=' + stats.stubbed + ' kept=' + stats.kept + ' saved=' + stats.savedBytes + 'B');
}
return body;
} catch (_) {
return bodyStr;
}
}
module.exports = { pruneBody, maybePrune, KEEP_RECENT, MIN_STUB_BYTES, ENGAGE_BYTES };