Summary

This adds higher-signal gateway stall diagnostics for slow Telegram turns where the transport is healthy but the Gateway/embedded Codex run stops making useful progress.

Changes:

• Add a lightweight diagnostic phase ring and emit diagnostic.phase.completed for Gateway and CLI startup spans.
• Include the current phase, recent completed phase spans, and bounded active/waiting/queued work labels in diagnostic.liveness.warning events and logs.
• Preserve phase/work attribution in the diagnostic stability recorder and bundle parser.
• Mark stalled embedded runs with terminalProgressStale=true when the last bridge progress looked terminal, such as rawResponseItem/completed, but the Gateway still considers the embedded run active.
• Add OPENCLAW_TRACE_SYNC_IO=1 support in scripts/run-node.mjs to pass Node --trace-sync-io through the normal source runner for startup I/O investigations.
• Enable OPENCLAW_TRACE_SYNC_IO=1 by default for pnpm gateway:watch Gateway children, while preserving explicit overrides like OPENCLAW_TRACE_SYNC_IO=0.
• Document the new liveness/stall fields and sync-I/O trace behavior.
• Improve Slack socket-mode startup retry logs so missing SDK error payloads report unknown error instead of undefined.

Why

The Telegram path was not the bottleneck: transport probes were fast, but Gateway sessions could remain active after the Codex bridge had already reported terminal-looking progress. Existing liveness warnings only showed aggregate counts (active, waiting, queued) plus event-loop/CPU numbers, so diagnosing the next stall required manual tmux/log archaeology.

With this change, the next warning should say which phase was active, what recent startup/runtime spans completed, which session/work label is blocking, and whether the bridge emitted terminal progress without closing the embedded run. Gateway watch also records sync I/O traces by default so local dev stalls have file/module-loader evidence without needing to remember the env flag.

Operator impact

No behavior change to scheduling, Telegram delivery, recovery thresholds, or restart policy.

Manual one-off source-runner tracing remains available:

OPENCLAW_TRACE_SYNC_IO=1 pnpm openclaw gateway --force

Gateway watch enables the same trace mode by default for the watched child. To suppress Node sync I/O trace output in watch mode:

OPENCLAW_TRACE_SYNC_IO=0 pnpm gateway:watch

Verification

• pnpm test src/logging/diagnostic.test.ts
• pnpm test src/infra/run-node.test.ts
• pnpm test src/infra/watch-node.test.ts
• pnpm test src/infra/gateway-watch-tmux.test.ts
• pnpm tsgo:core
• pnpm tsgo:core:test
• pnpm format:docs:check
• pnpm docs:check-mdx
• pnpm tsgo:extensions:test
• pnpm tsgo:extensions
• pnpm test extensions/slack/src/monitor/provider.reconnect.test.ts
• git diff --check