@openagents-org/agent-launcher 0.2.216 → 0.2.218
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/package.json +1 -1
- package/registry.json +3 -0
- package/src/adapters/codex.js +43 -9
- package/src/adapters/decision-log.js +2 -0
- package/src/adapters/gemini.js +5 -0
- package/src/adapters/hermes.js +13 -0
- package/src/adapters/opencode.js +54 -6
- package/src/adapters/utils.js +24 -1
- package/src/config.js +36 -8
- package/src/daemon.js +34 -3
- package/src/probe.js +12 -1
- package/src/utils.js +5 -2
package/package.json
CHANGED
package/registry.json
CHANGED
package/src/adapters/codex.js
CHANGED
|
@@ -25,7 +25,7 @@ const https = require('https');
|
|
|
25
25
|
|
|
26
26
|
const { whereBinary } = require('../paths');
|
|
27
27
|
const BaseAdapter = require('./base');
|
|
28
|
-
const { redactSecrets, formatAttachmentsForPrompt, decodeConsoleOutput, BENIGN_CLI_OUTPUT } = require('./utils');
|
|
28
|
+
const { redactSecrets, formatAttachmentsForPrompt, decodeConsoleOutput, BENIGN_CLI_OUTPUT, networkErrorText } = require('./utils');
|
|
29
29
|
const { buildOpenclawSystemPrompt } = require('./workspace-prompt');
|
|
30
30
|
|
|
31
31
|
const IS_WINDOWS = process.platform === 'win32';
|
|
@@ -39,7 +39,15 @@ const MAX_INLINE_ATTACHMENT_CHARS = 50000;
|
|
|
39
39
|
// then went quiet left the run hanging forever with nothing in the log.
|
|
40
40
|
// Overridable per instance (_directIdleTimeoutMs / _directTotalTimeoutMs).
|
|
41
41
|
const DIRECT_IDLE_TIMEOUT_MS = 120000;
|
|
42
|
+
// The overall deadline is checked, not enforced, while the model is still
|
|
43
|
+
// producing: a reasoning model thinking for minutes (22k streamed events in one
|
|
44
|
+
// report) was cut off at 300s and everything it had sent was thrown away. Each
|
|
45
|
+
// time it comes due with new text or reasoning since the last check it moves
|
|
46
|
+
// on by the same step, up to the hard ceiling below; a stream that only sends
|
|
47
|
+
// keep-alives, or nothing new, still ends there. What reply text there is at
|
|
48
|
+
// that point is posted, marked as cut off.
|
|
42
49
|
const DIRECT_TOTAL_TIMEOUT_MS = 300000;
|
|
50
|
+
const DIRECT_HARD_TIMEOUT_MS = 900000;
|
|
43
51
|
const TEXT_FILE_RE = /\.(md|markdown|txt|text|csv|tsv|json|jsonl|ya?ml|toml|ini|xml|html?|log|sql|sh|py|js|mjs|cjs|ts|tsx|jsx|java|go|rs|rb|php|c|h|cc|cpp|hpp|cs|swift|kt)$/i;
|
|
44
52
|
|
|
45
53
|
// How much of a failure's own words reach the channel.
|
|
@@ -983,13 +991,23 @@ class CodexAdapter extends BaseAdapter {
|
|
|
983
991
|
const mod = parsed.protocol === 'https:' ? https : http;
|
|
984
992
|
const idleMs = this._directIdleTimeoutMs || DIRECT_IDLE_TIMEOUT_MS;
|
|
985
993
|
const totalMs = this._directTotalTimeoutMs || DIRECT_TOTAL_TIMEOUT_MS;
|
|
994
|
+
const hardMs = Math.max(totalMs, this._directHardTimeoutMs || DIRECT_HARD_TIMEOUT_MS);
|
|
986
995
|
const started = Date.now();
|
|
987
996
|
// What arrived before things stopped, so a stalled relay can be told
|
|
988
997
|
// apart from one that answered in a shape we don't read.
|
|
989
998
|
const progress = { status: 0, bytes: 0, events: 0 };
|
|
999
|
+
let reply = null; // what the stream has carried so far, once it is one
|
|
990
1000
|
const seconds = () => Math.round((Date.now() - started) / 1000);
|
|
991
|
-
const
|
|
992
|
-
|
|
1001
|
+
const contentChars = () => (reply ? reply.text.length + reply.reasoning.length + reply.toolArgs.length : 0);
|
|
1002
|
+
// Text and reasoning counted apart: "still thinking" and "a relay
|
|
1003
|
+
// sending empty events" looked the same as bytes and events alone.
|
|
1004
|
+
const note = () => {
|
|
1005
|
+
const content = reply && contentChars()
|
|
1006
|
+
? `, ${reply.text.length} reply chars, ${reply.reasoning.length} reasoning chars`
|
|
1007
|
+
: '';
|
|
1008
|
+
return `HTTP ${progress.status || 'no response'}, ${progress.bytes} bytes, `
|
|
1009
|
+
+ `${progress.events} events${content}, ${seconds()}s`;
|
|
1010
|
+
};
|
|
993
1011
|
|
|
994
1012
|
let settled = false;
|
|
995
1013
|
let deadline = null;
|
|
@@ -1024,7 +1042,7 @@ class CodexAdapter extends BaseAdapter {
|
|
|
1024
1042
|
return;
|
|
1025
1043
|
}
|
|
1026
1044
|
|
|
1027
|
-
|
|
1045
|
+
reply = { text: '', reasoning: '', toolArgs: '', error: '', finishReason: '' };
|
|
1028
1046
|
let streamed = false;
|
|
1029
1047
|
let body = ''; // the raw response, kept until it turns out to be an SSE stream
|
|
1030
1048
|
let buffer = '';
|
|
@@ -1068,14 +1086,30 @@ class CodexAdapter extends BaseAdapter {
|
|
|
1068
1086
|
`The LLM API stopped responding, no data for ${Math.round(idleMs / 1000)}s (${note()})`));
|
|
1069
1087
|
req.destroy();
|
|
1070
1088
|
});
|
|
1071
|
-
|
|
1072
|
-
|
|
1073
|
-
|
|
1089
|
+
let charsAtCheck = 0;
|
|
1090
|
+
const onDeadline = () => {
|
|
1091
|
+
const elapsed = Date.now() - started;
|
|
1092
|
+
const chars = contentChars();
|
|
1093
|
+
if (chars > charsAtCheck && elapsed < hardMs) {
|
|
1094
|
+
charsAtCheck = chars;
|
|
1095
|
+
deadline = setTimeout(onDeadline, Math.min(totalMs, hardMs - elapsed));
|
|
1096
|
+
if (deadline.unref) deadline.unref();
|
|
1097
|
+
return;
|
|
1098
|
+
}
|
|
1099
|
+
const partial = reply ? CodexAdapter._completionText(reply) : '';
|
|
1100
|
+
if (partial) {
|
|
1101
|
+
this._log(`Direct API reply cut off at ${seconds()}s (${note()})`);
|
|
1102
|
+
finish(null, `${partial}\n\n_(Cut off: the model was still writing after ${seconds()}s.)_`);
|
|
1103
|
+
} else {
|
|
1104
|
+
const what = reply && reply.reasoning ? 'still thinking, no reply text yet' : 'no usable reply';
|
|
1105
|
+
finish(new Error(`The LLM API request went past ${seconds()}s with ${what} (${note()})`));
|
|
1106
|
+
}
|
|
1074
1107
|
req.destroy();
|
|
1075
|
-
}
|
|
1108
|
+
};
|
|
1109
|
+
deadline = setTimeout(onDeadline, totalMs);
|
|
1076
1110
|
if (deadline.unref) deadline.unref();
|
|
1077
1111
|
|
|
1078
|
-
req.on('error', (err) => finish(err));
|
|
1112
|
+
req.on('error', (err) => finish(err && err.message ? err : new Error(networkErrorText(err))));
|
|
1079
1113
|
req.write(payload);
|
|
1080
1114
|
req.end();
|
|
1081
1115
|
});
|
|
@@ -146,6 +146,8 @@ function isRecapEligible(msg, currentMessage) {
|
|
|
146
146
|
if (!msg) return false;
|
|
147
147
|
const mt = msg.messageType || 'chat';
|
|
148
148
|
if (mt === 'status' || mt === 'thinking' || mt === 'loading') return false;
|
|
149
|
+
// Control signals (a queue cancel) are not conversation for the agent.
|
|
150
|
+
if (mt === 'queue_cancel') return false;
|
|
149
151
|
const text = (msg.content || '').trim();
|
|
150
152
|
if (!text) return false;
|
|
151
153
|
if (text === currentMessage) return false;
|
package/src/adapters/gemini.js
CHANGED
|
@@ -290,6 +290,11 @@ class GeminiAdapter extends BaseAdapter {
|
|
|
290
290
|
|
|
291
291
|
let cmd;
|
|
292
292
|
const cleanEnv = { ...(this.agentEnv || process.env) };
|
|
293
|
+
// Without it a Gemini CLI with folder trust on (the default for a Google
|
|
294
|
+
// sign-in) stops with "Gemini CLI is not running in a trusted directory"
|
|
295
|
+
// in any working directory the user never trusted by hand. It is what the
|
|
296
|
+
// CLI's own --skip-trust sets; an explicit value of the user's stands.
|
|
297
|
+
if (cleanEnv.GEMINI_CLI_TRUST_WORKSPACE == null) cleanEnv.GEMINI_CLI_TRUST_WORKSPACE = 'true';
|
|
293
298
|
|
|
294
299
|
let _shouldRetry = false;
|
|
295
300
|
for (let attempt = 0; attempt < 2; attempt++) {
|
package/src/adapters/hermes.js
CHANGED
|
@@ -62,6 +62,14 @@ const TOOLSET_WARNING_RE = /^Warning: Unknown toolsets:/;
|
|
|
62
62
|
// (t("cli.resume.session_not_found")), so this only matches the English text;
|
|
63
63
|
// under another locale a lost session is not recognised as lost.
|
|
64
64
|
const SESSION_NOT_FOUND_RE = /session not found/i;
|
|
65
|
+
// What Hermes says when a resumed session was opened on credentials that are
|
|
66
|
+
// no longer in config.yaml: "This chat can't use the API key it was started
|
|
67
|
+
// with any more (model.api_key: removed from config.yaml since this session
|
|
68
|
+
// started), so nothing can run in it." A session pins its credentials, so a
|
|
69
|
+
// session begun on the device-wide `model:` key an earlier version wrote
|
|
70
|
+
// (handed back once each agent got a provider of its own) could never run
|
|
71
|
+
// again; only a new session can. English only, like SESSION_NOT_FOUND_RE.
|
|
72
|
+
const SESSION_CREDENTIALS_GONE_RE = /(?:can't|cannot) use the (?:API key|credentials?) it was started with|removed from config\.yaml since this session started/i;
|
|
65
73
|
const MAX_HISTORY_ENTRIES = 12;
|
|
66
74
|
const OPENAGENTS_RULES = [
|
|
67
75
|
'\n## OpenAgents-specific Rules',
|
|
@@ -826,6 +834,11 @@ class HermesAdapter extends BaseAdapter {
|
|
|
826
834
|
async _converse(channelName, content) {
|
|
827
835
|
const resumeId = this._channelSessions[channelName] || null;
|
|
828
836
|
const run = await this._attemptTurn(channelName, content, resumeId);
|
|
837
|
+
if (resumeId && !run.stopped && SESSION_CREDENTIALS_GONE_RE.test(`${run.text || ''}\n${run.detail || ''}`)) {
|
|
838
|
+
this._log(`Hermes session ${resumeId} was started on credentials no longer in config.yaml, starting a new one`);
|
|
839
|
+
this._forgetSession(channelName);
|
|
840
|
+
return this._attemptTurn(channelName, content, null);
|
|
841
|
+
}
|
|
829
842
|
const neverBegan = !run.stopped && run.exitCode !== 0 && !run.sessionId;
|
|
830
843
|
if (!resumeId || !neverBegan) return run;
|
|
831
844
|
|
package/src/adapters/opencode.js
CHANGED
|
@@ -43,6 +43,11 @@ const OPENCODE_PINNED_VERSION = '1.17.11';
|
|
|
43
43
|
// Supported floor — below this we block the run with `unsupported_version`.
|
|
44
44
|
// The 1.17.x line is the stream-json shape this adapter targets.
|
|
45
45
|
const OPENCODE_MIN_VERSION = '1.17.0';
|
|
46
|
+
// First major this adapter cannot drive. OpenCode 2 (`@opencode/cli`, its
|
|
47
|
+
// Homebrew formula and curl installer) is a new command line: `run` has no
|
|
48
|
+
// `--dir`, so every message died with "Unrecognized flag: --dir" while the
|
|
49
|
+
// preflight, seeing only "newer than tested", let it through.
|
|
50
|
+
const OPENCODE_UNSUPPORTED_MAJOR = 2;
|
|
46
51
|
// Highest version whose JSON behavior is treated as verified. Newer is allowed
|
|
47
52
|
// but flagged 'degraded' (the run still proceeds; any failure is classified).
|
|
48
53
|
const OPENCODE_TESTED_MAX_VERSION = '1.17.11';
|
|
@@ -70,7 +75,7 @@ const FAILURE_MESSAGES = {
|
|
|
70
75
|
cli_not_found: `OpenCode CLI not found. Install it (\`npm install -g opencode-ai@${OPENCODE_PINNED_VERSION}\`) and try again.`,
|
|
71
76
|
cli_not_executable: 'OpenCode CLI was found but could not be started. Reinstall it, then retry.',
|
|
72
77
|
install_incomplete: `OpenCode is installed, but its setup never finished: the opencode-ai package's postinstall script did not run (npm \`ignore-scripts\`, or a pnpm/bun install), so the \`opencode\` command is only a placeholder. Reinstall OpenCode from the launcher — or, for a copy you installed yourself, run \`npm install -g opencode-ai@${OPENCODE_PINNED_VERSION} --ignore-scripts=false\` — then retry.`,
|
|
73
|
-
unsupported_version: `This OpenCode CLI version is not supported
|
|
78
|
+
unsupported_version: `This OpenCode CLI version is not supported: OpenAgents runs the opencode-ai 1.x CLI (${OPENCODE_MIN_VERSION} or newer), and OpenCode 2.x changed its command line. Reinstall OpenCode from the launcher, or run \`npm install -g opencode-ai@${OPENCODE_PINNED_VERSION}\`, and retry.`,
|
|
74
79
|
model_missing: "No model is configured for OpenCode. Set a model (LLM_MODEL, e.g. `gpt-4o`) in this agent's configuration, then retry.",
|
|
75
80
|
credential_missing: 'OpenCode has no API key or sign-in configured. Add an API key (LLM_API_KEY) or run `opencode auth login`, then retry.',
|
|
76
81
|
auth_failed: "OpenCode's provider rejected the credentials (authentication failed). Check the API key or sign-in and retry.",
|
|
@@ -159,7 +164,10 @@ class OpenCodeAdapter extends BaseAdapter {
|
|
|
159
164
|
// opencode hangs waiting for interactive provider/model selection.
|
|
160
165
|
// Connected agents run unattended: let the CLI resolve permission prompts
|
|
161
166
|
// automatically instead of rejecting tools outside the working directory.
|
|
162
|
-
|
|
167
|
+
// --thinking adds the model's reasoning to the JSON stream as `reasoning`
|
|
168
|
+
// events, shown live as thinking (like Claude and Codex). Available since
|
|
169
|
+
// the minimum supported 1.17.0.
|
|
170
|
+
const args = ['run', '--format', 'json', '--thinking', '--dangerously-skip-permissions', '--dir', runCwd, '--model', model];
|
|
163
171
|
if (sessionId) args.push('--session', sessionId);
|
|
164
172
|
return args;
|
|
165
173
|
}
|
|
@@ -504,6 +512,7 @@ class OpenCodeAdapter extends BaseAdapter {
|
|
|
504
512
|
const eventType = String(event.type || '').toLowerCase();
|
|
505
513
|
if (['step_start', 'step-start', 'step_finish', 'step-finish', 'tool_use', 'tool'].includes(eventType)) return null;
|
|
506
514
|
if (OpenCodeAdapter._isToolEvent(event)) return null;
|
|
515
|
+
if (OpenCodeAdapter._isReasoningEvent(event)) return null;
|
|
507
516
|
if (eventType.includes('error')) return null;
|
|
508
517
|
|
|
509
518
|
const part = event.part;
|
|
@@ -574,6 +583,13 @@ class OpenCodeAdapter extends BaseAdapter {
|
|
|
574
583
|
}
|
|
575
584
|
|
|
576
585
|
async _handleStreamEvent(event, msgChannel, responseState = null) {
|
|
586
|
+
// Reasoning is a live thinking trace only: it never counts toward the reply.
|
|
587
|
+
const reasoning = OpenCodeAdapter._reasoningFromEvent(event);
|
|
588
|
+
if (reasoning !== null) {
|
|
589
|
+
if (reasoning.trim()) await this.sendThinking(msgChannel, reasoning.trim());
|
|
590
|
+
return;
|
|
591
|
+
}
|
|
592
|
+
|
|
577
593
|
const status = OpenCodeAdapter._toolStatusFromEvent(event);
|
|
578
594
|
if (status) {
|
|
579
595
|
// A tool call resets `finalText` (the "text since the last tool"), so the
|
|
@@ -1075,8 +1091,21 @@ class OpenCodeAdapter extends BaseAdapter {
|
|
|
1075
1091
|
* unusable cwd) and lets anything ambiguous through to a real run.
|
|
1076
1092
|
*/
|
|
1077
1093
|
_preflight() {
|
|
1078
|
-
|
|
1079
|
-
|
|
1094
|
+
// The path found at start-up is kept, but not past the file it names: a
|
|
1095
|
+
// reinstall or an antivirus quarantine took it away, and every message
|
|
1096
|
+
// after that failed "CLI not found" until the agent restarted, even with
|
|
1097
|
+
// a working copy installed by then.
|
|
1098
|
+
let binary = this._opencodeBinary;
|
|
1099
|
+
if (binary && !fs.existsSync(binary)) {
|
|
1100
|
+
const found = this._findOpencodeBinary();
|
|
1101
|
+
if (found !== binary) {
|
|
1102
|
+
this._log(`${binary} is gone; ${found ? `using ${found}` : 'no other copy found'}`);
|
|
1103
|
+
this._versionProbe = null;
|
|
1104
|
+
}
|
|
1105
|
+
binary = found;
|
|
1106
|
+
}
|
|
1107
|
+
binary = binary || this._findOpencodeBinary();
|
|
1108
|
+
this._opencodeBinary = binary || null;
|
|
1080
1109
|
if (!binary) return { ok: false, category: 'cli_not_found' };
|
|
1081
1110
|
|
|
1082
1111
|
// Before every other check: the placeholder a skipped postinstall leaves
|
|
@@ -1100,7 +1129,7 @@ class OpenCodeAdapter extends BaseAdapter {
|
|
|
1100
1129
|
ok: false,
|
|
1101
1130
|
category: 'unsupported_version',
|
|
1102
1131
|
detail: `detected ${probe.version || 'unknown'}`,
|
|
1103
|
-
diagnostic: `detected ${probe.version}, requires >= ${OPENCODE_MIN_VERSION}`,
|
|
1132
|
+
diagnostic: `detected ${probe.version}, requires >= ${OPENCODE_MIN_VERSION} and < ${OPENCODE_UNSUPPORTED_MAJOR}.0.0`,
|
|
1104
1133
|
};
|
|
1105
1134
|
}
|
|
1106
1135
|
// 'unknown' (unparseable) and 'degraded' (newer than tested) still run.
|
|
@@ -1262,6 +1291,24 @@ class OpenCodeAdapter extends BaseAdapter {
|
|
|
1262
1291
|
// ------------------------------------------------------------------
|
|
1263
1292
|
|
|
1264
1293
|
/** True when an event is a tool call, across opencode's version-varying shapes. */
|
|
1294
|
+
/** A `--thinking` reasoning event (`{type:"reasoning", part:{type:"reasoning", text}}`). */
|
|
1295
|
+
static _isReasoningEvent(event) {
|
|
1296
|
+
if (!event || typeof event !== 'object') return false;
|
|
1297
|
+
if (String(event.type || '').toLowerCase() === 'reasoning') return true;
|
|
1298
|
+
const part = event.part;
|
|
1299
|
+
return !!(part && typeof part === 'object' && String(part.type || '').toLowerCase() === 'reasoning');
|
|
1300
|
+
}
|
|
1301
|
+
|
|
1302
|
+
/** Reasoning text from a reasoning event ('' when empty), or null for any other event. */
|
|
1303
|
+
static _reasoningFromEvent(event) {
|
|
1304
|
+
if (!OpenCodeAdapter._isReasoningEvent(event)) return null;
|
|
1305
|
+
const part = (event.part && typeof event.part === 'object') ? event.part : event;
|
|
1306
|
+
const text = part.text || part.content || '';
|
|
1307
|
+
// Summaries of several reasoning steps arrive glued together
|
|
1308
|
+
// ("**Checking A****Testing B**"); give each heading its own paragraph.
|
|
1309
|
+
return typeof text === 'string' ? text.replace(/\*\*\*\*(?=\S)/g, '**\n\n**') : '';
|
|
1310
|
+
}
|
|
1311
|
+
|
|
1265
1312
|
static _isToolEvent(event) {
|
|
1266
1313
|
if (!event || typeof event !== 'object') return false;
|
|
1267
1314
|
const t = String(event.type || '').toLowerCase();
|
|
@@ -1383,11 +1430,12 @@ class OpenCodeAdapter extends BaseAdapter {
|
|
|
1383
1430
|
return 'empty_response';
|
|
1384
1431
|
}
|
|
1385
1432
|
|
|
1386
|
-
/** 'ok' | 'degraded' (newer than tested) | 'unsupported' (too old) | 'unknown'. */
|
|
1433
|
+
/** 'ok' | 'degraded' (newer than tested) | 'unsupported' (too old, or 2.x) | 'unknown'. */
|
|
1387
1434
|
static _classifyVersion(version) {
|
|
1388
1435
|
// Non-semver (or unreadable) → 'unknown': allow the run, never block on it.
|
|
1389
1436
|
if (!version || !/^\d+\.\d+/.test(String(version))) return 'unknown';
|
|
1390
1437
|
if (OpenCodeAdapter._cmpVer(version, OPENCODE_MIN_VERSION) < 0) return 'unsupported';
|
|
1438
|
+
if (Number(String(version).split('.')[0]) >= OPENCODE_UNSUPPORTED_MAJOR) return 'unsupported';
|
|
1391
1439
|
if (OPENCODE_TESTED_MAX_VERSION && OpenCodeAdapter._cmpVer(version, OPENCODE_TESTED_MAX_VERSION) > 0) return 'degraded';
|
|
1392
1440
|
return 'ok';
|
|
1393
1441
|
}
|
package/src/adapters/utils.js
CHANGED
|
@@ -247,8 +247,30 @@ function decodeConsoleOutput(buf, isWindows = process.platform === 'win32') {
|
|
|
247
247
|
* from stdin...", which filled the smoke test's timeout message. Node's own
|
|
248
248
|
* process warnings ("(node:1234) Warning: …", "(Use `node --trace-warnings
|
|
249
249
|
* ...`") come ahead of the real error from npm-shipped CLIs like CodeBuddy.
|
|
250
|
+
* Newer Claude Code reports an unknown model as a one-line signal instead,
|
|
251
|
+
* "[claude-code:unrecognized_model] {"model":…}", on every run against a relay
|
|
252
|
+
* model; read as the failure, it turned a timed-out smoke test into "wrong
|
|
253
|
+
* model" and skipped the direct API check that would have said why.
|
|
250
254
|
*/
|
|
251
|
-
const BENIGN_CLI_OUTPUT = /claude\.ai connectors are disabled|doesn't recognize the model|assumes a context window|auto-compact keeps this session|CLAUDE_CODE_DISABLE_UNKNOWN_MODEL_WINDOW_ENFORCEMENT|Model metadata for .* not found|Defaulting to fallback metadata|^\s*Reading (prompt|additional input) from stdin|^\s*\d{4}-\d{2}-\d{2}T[\d:.]+Z?\s+(TRACE|DEBUG|INFO|WARN)\s|^\s*\(node:\d+\) (\[\w+\] )?\w*Warning:|^\s*\(Use `node --trace-(warnings|deprecation) \.\.\.`/i;
|
|
255
|
+
const BENIGN_CLI_OUTPUT = /claude\.ai connectors are disabled|doesn't recognize the model|\[claude-code:unrecognized_model\]|assumes a context window|auto-compact keeps this session|CLAUDE_CODE_DISABLE_UNKNOWN_MODEL_WINDOW_ENFORCEMENT|Model metadata for .* not found|Defaulting to fallback metadata|^\s*Reading (prompt|additional input) from stdin|^\s*\d{4}-\d{2}-\d{2}T[\d:.]+Z?\s+(TRACE|DEBUG|INFO|WARN)\s|^\s*\(node:\d+\) (\[\w+\] )?\w*Warning:|^\s*\(Use `node --trace-(warnings|deprecation) \.\.\.`/i;
|
|
256
|
+
|
|
257
|
+
/**
|
|
258
|
+
* What a failed socket says went wrong. Node 20+ tries every address a host
|
|
259
|
+
* resolves to (happy eyeballs); when all of them fail it raises an
|
|
260
|
+
* AggregateError whose own message is EMPTY, with the reasons on `errors` and
|
|
261
|
+
* the code on `code`. Read as `e.message` alone that was "": the smoke test
|
|
262
|
+
* took it for success and crashed on the missing body ("Cannot read
|
|
263
|
+
* properties of undefined (reading 'slice')"), and a chat run failed with no
|
|
264
|
+
* reason at all.
|
|
265
|
+
*/
|
|
266
|
+
function networkErrorText(err) {
|
|
267
|
+
if (!err) return 'unknown network error';
|
|
268
|
+
if (err.message) return err.message;
|
|
269
|
+
const inner = Array.isArray(err.errors)
|
|
270
|
+
? [...new Set(err.errors.map((e) => e && e.message).filter(Boolean))].join('; ')
|
|
271
|
+
: '';
|
|
272
|
+
return inner || err.code || String(err) || 'unknown network error';
|
|
273
|
+
}
|
|
252
274
|
|
|
253
275
|
/** CLI output with the known-harmless lines taken out. */
|
|
254
276
|
function stripCliNoise(text) {
|
|
@@ -257,6 +279,7 @@ function stripCliNoise(text) {
|
|
|
257
279
|
|
|
258
280
|
module.exports = {
|
|
259
281
|
SESSION_DEFAULT_RE,
|
|
282
|
+
networkErrorText,
|
|
260
283
|
BENIGN_CLI_OUTPUT,
|
|
261
284
|
stripCliNoise,
|
|
262
285
|
decodeConsoleOutput,
|
package/src/config.js
CHANGED
|
@@ -3,6 +3,16 @@
|
|
|
3
3
|
const fs = require('fs');
|
|
4
4
|
const path = require('path');
|
|
5
5
|
|
|
6
|
+
// Windows refuses to replace or open a file another process holds open (the
|
|
7
|
+
// daemon and its `agn` children read daemon.yaml all the time). Such a refusal
|
|
8
|
+
// passes in milliseconds, so both sides wait a little and try again.
|
|
9
|
+
const BUSY_CODES = ['EPERM', 'EACCES', 'EBUSY'];
|
|
10
|
+
const BUSY_WAITS_MS = [10, 25, 50, 100, 200];
|
|
11
|
+
|
|
12
|
+
function pause(ms) {
|
|
13
|
+
try { Atomics.wait(new Int32Array(new SharedArrayBuffer(4)), 0, 0, ms); } catch {}
|
|
14
|
+
}
|
|
15
|
+
|
|
6
16
|
/**
|
|
7
17
|
* Manages ~/.openagents/daemon.yaml — the agent/network configuration file.
|
|
8
18
|
* Compatible with the Python SDK's daemon_config.py format.
|
|
@@ -19,14 +29,30 @@ class Config {
|
|
|
19
29
|
|
|
20
30
|
// -- Read --
|
|
21
31
|
|
|
32
|
+
/**
|
|
33
|
+
* The config as it is on disk. A read that hits the file while it is held
|
|
34
|
+
* (Windows) or while a writer that could not replace it is rewriting it in
|
|
35
|
+
* place (empty for that moment) is tried again: read as "no agents", it made
|
|
36
|
+
* a probe started right after a save report "No agent named '<name>' on
|
|
37
|
+
* this device", and that result stood for an hour.
|
|
38
|
+
*/
|
|
22
39
|
load() {
|
|
23
|
-
|
|
24
|
-
|
|
40
|
+
for (let attempt = 0; ; attempt++) {
|
|
41
|
+
const last = attempt >= BUSY_WAITS_MS.length;
|
|
42
|
+
try {
|
|
43
|
+
if (!fs.existsSync(this.configFile)) break;
|
|
25
44
|
const text = fs.readFileSync(this.configFile, 'utf-8');
|
|
26
|
-
|
|
45
|
+
// An empty file is tried again only briefly: a writer's in-place
|
|
46
|
+
// rewrite is over within that, and a file that really is empty must
|
|
47
|
+
// not slow down every read.
|
|
48
|
+
if (text.trim() || last || attempt >= 2) return parseYaml(text);
|
|
49
|
+
} catch (err) {
|
|
50
|
+
if (last || !BUSY_CODES.includes(err.code)) {
|
|
51
|
+
console.error('Failed to load config:', err.message);
|
|
52
|
+
break;
|
|
53
|
+
}
|
|
27
54
|
}
|
|
28
|
-
|
|
29
|
-
console.error('Failed to load config:', err.message);
|
|
55
|
+
pause(BUSY_WAITS_MS[attempt]);
|
|
30
56
|
}
|
|
31
57
|
return { version: 2, agents: [], networks: [] };
|
|
32
58
|
}
|
|
@@ -51,19 +77,21 @@ class Config {
|
|
|
51
77
|
* started right after a create read it empty and reported the new agent as
|
|
52
78
|
* "Unknown agent type '<name>'", and a read-modify-write in that window
|
|
53
79
|
* saves the empty list back. Windows can refuse the rename while a reader
|
|
54
|
-
* holds the file open, so that case retries
|
|
80
|
+
* holds the file open, so that case retries for a moment (the hold is a
|
|
81
|
+
* read, gone in milliseconds) and only then writes in place.
|
|
55
82
|
*/
|
|
56
83
|
save(config) {
|
|
57
84
|
fs.mkdirSync(this.configDir, { recursive: true });
|
|
58
85
|
const text = serializeYaml(config);
|
|
59
86
|
const tmp = `${this.configFile}.${process.pid}.tmp`;
|
|
60
87
|
fs.writeFileSync(tmp, text, 'utf-8');
|
|
61
|
-
for (let attempt = 0; attempt
|
|
88
|
+
for (let attempt = 0; attempt <= BUSY_WAITS_MS.length; attempt++) {
|
|
89
|
+
if (attempt > 0) pause(BUSY_WAITS_MS[attempt - 1]);
|
|
62
90
|
try {
|
|
63
91
|
fs.renameSync(tmp, this.configFile);
|
|
64
92
|
return;
|
|
65
93
|
} catch (err) {
|
|
66
|
-
if (!
|
|
94
|
+
if (!BUSY_CODES.includes(err.code)) {
|
|
67
95
|
try { fs.unlinkSync(tmp); } catch {}
|
|
68
96
|
throw err;
|
|
69
97
|
}
|
package/src/daemon.js
CHANGED
|
@@ -224,9 +224,19 @@ class Daemon {
|
|
|
224
224
|
try {
|
|
225
225
|
// By agent name, not type: the probe runs on this agent's own env, so a
|
|
226
226
|
// signed-in agent is tested on its sign-in, not the type's saved key.
|
|
227
|
-
const r = await this._runAgn(['probe', name, '--json']);
|
|
228
227
|
let parsed = null;
|
|
229
|
-
|
|
228
|
+
let r;
|
|
229
|
+
for (let attempt = 0; attempt < 2; attempt++) {
|
|
230
|
+
r = await this._runAgn(['probe', name, '--json']);
|
|
231
|
+
parsed = null;
|
|
232
|
+
try { parsed = JSON.parse(r.stdout.trim()); } catch {}
|
|
233
|
+
// The child found no agent of that name while this daemon has one: it
|
|
234
|
+
// read daemon.yaml mid-save (Windows). Ask once more instead of
|
|
235
|
+
// keeping "No agent named '<name>' on this device" for an hour.
|
|
236
|
+
const unseen = parsed && parsed.code === 'unknown_type' && this.config.getAgent(name);
|
|
237
|
+
if (!unseen || attempt > 0) break;
|
|
238
|
+
await new Promise((resolve) => setTimeout(resolve, 500));
|
|
239
|
+
}
|
|
230
240
|
if (!parsed || typeof parsed !== 'object') {
|
|
231
241
|
// The probe process itself died before it could answer — a fault on
|
|
232
242
|
// this device's OpenAgents install, not in the agent.
|
|
@@ -277,6 +287,24 @@ class Daemon {
|
|
|
277
287
|
})();
|
|
278
288
|
}
|
|
279
289
|
|
|
290
|
+
/**
|
|
291
|
+
* The connection test of an agent whose runtime was the wrong version ran
|
|
292
|
+
* before the repair (or at creation, while it was under way) and said
|
|
293
|
+
* "<version> is not supported"; nothing tested it again until the hourly
|
|
294
|
+
* sweep, so the device card kept that error for an agent that now worked.
|
|
295
|
+
*/
|
|
296
|
+
_reprobeAfterRepair(agentType) {
|
|
297
|
+
const due = this.config.getAgents().filter((a) => (a.type || 'openclaw') === agentType
|
|
298
|
+
&& this._probes[a.name] && this._probes[a.name].code === 'version_incompatible');
|
|
299
|
+
if (!due.length) return;
|
|
300
|
+
(async () => {
|
|
301
|
+
for (const a of due) {
|
|
302
|
+
try { await this._probeAgent(a.name); } catch {}
|
|
303
|
+
}
|
|
304
|
+
this._nodeHeartbeat();
|
|
305
|
+
})();
|
|
306
|
+
}
|
|
307
|
+
|
|
280
308
|
/** Report a runtime install that failed while creating `name`, as its probe result. */
|
|
281
309
|
_recordInstallFailure(name, type, error) {
|
|
282
310
|
let install = null;
|
|
@@ -1984,7 +2012,10 @@ async _runNodeCommand(n, cmd) {
|
|
|
1984
2012
|
|
|
1985
2013
|
if (typeof adapter.forgetRuntime === 'function') adapter.forgetRuntime();
|
|
1986
2014
|
const next = adapter.preflight();
|
|
1987
|
-
if (next && next.ok !== false)
|
|
2015
|
+
if (next && next.ok !== false) {
|
|
2016
|
+
this._log(`${name}: supported ${agentType} runtime installed`);
|
|
2017
|
+
this._reprobeAfterRepair(agentType);
|
|
2018
|
+
}
|
|
1988
2019
|
return next;
|
|
1989
2020
|
}
|
|
1990
2021
|
|
package/src/probe.js
CHANGED
|
@@ -549,6 +549,13 @@ async function probeAgentType(connector, type, opts = {}) {
|
|
|
549
549
|
const timeoutMs = opts.timeoutMs
|
|
550
550
|
|| (entry.probe && entry.probe.timeout_s ? entry.probe.timeout_s * 1000 : DEFAULT_TIMEOUT_MS);
|
|
551
551
|
const env = stripForCliLogin(type, { ...getEnhancedEnv(), ...agentEnv }, connector.registry, opts.agentEnv);
|
|
552
|
+
// What the CLI needs set to answer headless at all, declared per agent
|
|
553
|
+
// (Gemini refuses to run in a folder it does not trust, and the probe runs
|
|
554
|
+
// in the home directory). The agent's own value, if any, stands.
|
|
555
|
+
const probeEnv = (entry.probe && entry.probe.env) || {};
|
|
556
|
+
for (const [name, value] of Object.entries(probeEnv)) {
|
|
557
|
+
if (env[name] == null) env[name] = String(value);
|
|
558
|
+
}
|
|
552
559
|
|
|
553
560
|
const baseUrl = directBaseUrl(entry, agentEnv);
|
|
554
561
|
const protocol = directProtocol(type, agentEnv);
|
|
@@ -608,7 +615,11 @@ async function probeAgentType(connector, type, opts = {}) {
|
|
|
608
615
|
// same question directly: one round trip, and it says why.
|
|
609
616
|
const unexplained = code === CODE.TIMEOUT || code === CODE.CLI_ERROR || code === CODE.EMPTY_RESPONSE;
|
|
610
617
|
if (unexplained && canCallDirect) {
|
|
611
|
-
|
|
618
|
+
// `[1m]` is Claude Code's own 1M-context switch (`deepseek-v4-pro[1m]`),
|
|
619
|
+
// taken off before its requests go out; sent as part of the name, the
|
|
620
|
+
// relay would reject a model the agent actually runs on.
|
|
621
|
+
const wireModel = model ? String(model).replace(/\[1m\]$/i, '') : model;
|
|
622
|
+
const direct = await directApiVerdict(agentEnv, entry, wireModel, baseUrl, protocol);
|
|
612
623
|
if (direct) return done(direct);
|
|
613
624
|
}
|
|
614
625
|
if (unexplained && staticVerdict) return done(staticVerdict);
|
package/src/utils.js
CHANGED
|
@@ -22,6 +22,8 @@ const VENDOR_HOSTS = /(^|\.)(openai\.com|moonshot\.ai|deepseek\.com|anthropic\.c
|
|
|
22
22
|
const MODEL_REJECTED =
|
|
23
23
|
/model.{0,60}?(is not available|does not exist|not found|unknown model|invalid model|no such model|not supported)|model_not_found/is;
|
|
24
24
|
|
|
25
|
+
const { networkErrorText } = require('./adapters/utils');
|
|
26
|
+
|
|
25
27
|
/** One HTTP request. Resolves { status, data } or { error }; never rejects. */
|
|
26
28
|
function request(url, { method = 'POST', headers = {}, body = null, timeout = 15000 } = {}) {
|
|
27
29
|
const https = require('https');
|
|
@@ -36,7 +38,7 @@ function request(url, { method = 'POST', headers = {}, body = null, timeout = 15
|
|
|
36
38
|
res.on('data', (chunk) => { data += chunk; });
|
|
37
39
|
res.on('end', () => resolve({ status: res.statusCode, data }));
|
|
38
40
|
});
|
|
39
|
-
req.on('error', (e) => resolve({ error: e
|
|
41
|
+
req.on('error', (e) => resolve({ error: networkErrorText(e) }));
|
|
40
42
|
// Name the host and the deadline: a bare "Request timed out" on the
|
|
41
43
|
// device card did not say whether the provider or the network was slow.
|
|
42
44
|
req.on('timeout', () => { req.destroy(); resolve({ error: `Request to ${parsedUrl.host} timed out after ${Math.round(timeout / 1000)}s` }); });
|
|
@@ -94,7 +96,8 @@ const OUTPUT_LIMIT_REACHED = /max_tokens or model output limit was reached/i;
|
|
|
94
96
|
/** Read one completion response into the result shape. */
|
|
95
97
|
function completionResult(res, { isAnthropic, model }) {
|
|
96
98
|
if (res.error) return { success: false, error: res.error };
|
|
97
|
-
const
|
|
99
|
+
const status = res.status;
|
|
100
|
+
const data = String(res.data == null ? '' : res.data);
|
|
98
101
|
if (status >= 400 && OUTPUT_LIMIT_REACHED.test(data || '')) {
|
|
99
102
|
return { success: true, model: model || '?', response: '' };
|
|
100
103
|
}
|