openclaw - 💡(How to fix) Fix Lossless/pi-ai `completeSimple` fails with `Connection error` for OpenClaw `claude-cli` Sonnet transport [1 participants]

Summary

Lossless Claw / pi-ai completeSimple fails when using OpenClaw's synthetic claude-cli/claude-sonnet-4-6 provider path, even though raw Claude CLI Sonnet works and OpenClaw runtime auth can read Claude CLI OAuth credentials from Keychain.

This looks like a transport integration issue between Lossless Claw's simple-completion summarizer path and OpenClaw's claude-cli provider (api = anthropic-messages, baseUrl = cli://claude), not a model/auth outage.

Environment

• macOS host: Darwin arm64
• OpenClaw: 2026.4.25 (aa36ee6)
• Lossless Claw: 0.9.2
• Claude Code CLI: 2.1.119
• Production LCM remains healthy on GPT-5.5:
  • summaryModel = openai-codex/gpt-5.5
  • summaryProvider = null / absent
  • proactiveThresholdCompactionMode = deferred
  • summaryTimeoutMs = 180000
• Synthetic Claude provider config:
  • models.providers.claude-cli.api = anthropic-messages
  • models.providers.claude-cli.baseUrl = cli://claude
  • models.providers.claude-cli.models = []

Controls that pass

Raw Sonnet CLI succeeds:

claude --print --model sonnet 'Reply with exactly SONNET_RAW_OK'

SONNET_RAW_OK

OpenClaw runtime auth lookup for the synthetic Sonnet model also succeeds and reads Claude CLI OAuth credentials from Keychain:

{
  "auth": {
    "mode": "oauth",
    "hasApiKey": true,
    "hasRequest": false,
    "requestKeys": [],
    "error": null
  }
}

Production LCM on GPT-5.5 is healthy: 176 production summaries in the last 24h, 0 production fallbacks, fallback rate 0%, degraded=false.

Repro harness

This is a standalone non-production harness. It does not change OpenClaw config, restart the gateway, or write to lcm.db.

import { completeSimple } from '/Users/nonzero/.openclaw/extensions/lossless-claw/node_modules/@mariozechner/pi-ai/dist/stream.js';
import { getRuntimeAuthForModel } from '/opt/homebrew/lib/node_modules/openclaw/dist/runtime-model-auth.runtime.js';
import fs from 'node:fs';

const cfg = JSON.parse(fs.readFileSync('/Users/nonzero/.openclaw/openclaw.json', 'utf8'));
const lookupModel = {
  id: 'claude-cli/claude-sonnet-4-6',
  provider: 'claude-cli',
  model: 'claude-sonnet-4-6',
  api: 'anthropic-messages',
  contextWindow: 200000,
};
const runtimeAuth = await getRuntimeAuthForModel({
  model: lookupModel,
  cfg,
  workspaceDir: '/Users/nonzero/.openclaw/workspace',
});

const model = {
  id: 'claude-cli/claude-sonnet-4-6',
  name: 'claude-cli/claude-sonnet-4-6',
  provider: 'claude-cli',
  model: 'claude-sonnet-4-6',
  api: 'anthropic-messages',
  baseUrl: 'cli://claude',
  contextWindow: 200000,
  maxTokens: 4096,
};

const result = await completeSimple(
  { ...model, apiKey: runtimeAuth?.apiKey },
  { messages: [{ role: 'user', content: 'Reply with exactly LCM_SONNET_OK', timestamp: Date.now() }] },
  { maxTokens: 32, temperature: 0, apiKey: runtimeAuth?.apiKey }
);
console.log(JSON.stringify(result, null, 2));

Actual result

{
  "role": "assistant",
  "content": [],
  "api": "anthropic-messages",
  "provider": "claude-cli",
  "model": "claude-cli/claude-sonnet-4-6",
  "usage": {
    "input": 0,
    "output": 0,
    "cacheRead": 0,
    "cacheWrite": 0,
    "totalTokens": 0
  },
  "stopReason": "error",
  "errorMessage": "Connection error."
}

Expected result

Either:

1. completeSimple should route through OpenClaw's claude-cli / cli://claude transport and return content, or
2. it should fail with a specific transport/provider error explaining why synthetic claude-cli cannot be used from this Lossless/pi-ai path.

Impact

This does not currently affect production because LCM is running on GPT-5.5 and is healthy. It does block safely using Sonnet as a Lossless Claw summarization model, because the controlled LCM symptom is the same: empty content, zero usage, finish/stopReason=error, Connection error, then fallback/truncation summary.

Safety note

No production config changes were made during this smoke test:

• no summaryModel switch to Sonnet
• no gateway restart
• no lcm.db mutation
• no plugin allow/config mutation

Production remains on openai-codex/gpt-5.5.