openclaw - 💡(How to fix) Fix Webchat persistence regression still firing on 2026.5.6 — PR #77839 fix appears to have not shipped [1 comments, 2 participants]

Summary

Issue #77518 was closed as completed when PR #77839 merged on 2026-05-05. After upgrading to 2026.5.6 (c97b9f7) on macOS today (2026-05-06), the cli session reset: provider=claude-cli reason=missing-transcript event is still firing on embedded webchat sessions at the same rate as before — 4 events in the first 90 minutes of the upgraded gateway.

Inspection of the 2026.5.6 dist suggests the PR's changes were not actually included in this release.

Environment

• OpenClaw 2026.5.6 (c97b9f7) (npm latest, installed via npm update -g openclaw 2026-05-06 14:38 EDT)
• Node 22.22.2 on macOS 26.4.1 (arm64)
• Gateway running as LaunchAgent, embedded webchat, claude-cli/claude-opus-4-7 model

Symptom

agent/cli-backend logs cli session reset: provider=claude-cli reason=missing-transcript, the gateway invalidates the prior CLI session and starts a fresh one. Visible to the user as missing prior turns / context loss on session resume.

Timeline post-upgrade on this host (gateway started 14:40 EDT):

2026-05-06T15:20:01 EDT — recent jsonl 1.19 MB, 236 asst / 131 user turns
2026-05-06T15:54:23 EDT
2026-05-06T16:25:47 EDT — recent jsonl 23 KB, 0 asst / 1 user turn (newly opened session)
2026-05-06T16:50:37 EDT — recent jsonl 88 KB, 19 asst / 11 user turns

That's 4 events in 1h30m on bare 2026.5.6, identical cadence to what was observed on 2026.5.4 before #77518 was closed.

Evidence the PR didn't ship in 2026.5.6 dist

Grepping /opt/homebrew/lib/node_modules/openclaw/dist/*.js on this 2026.5.6 install:

The most direct check — agent-command*.js:

// 2026.5.6 dist, agent-command*.js line 883:
if (result.meta.executionTrace?.runner === "cli") try {
    sessionEntry = await attemptExecutionRuntime.persistCliTurnTranscript({ ... });

PR #77839 was supposed to change this gate to runner === "cli" || runner === "embedded" and pass embeddedAssistantGapFill: runner === "embedded". Neither change is present in the 2026.5.6 dist. The runner-tag gate that drops embedded webchat turns from .jsonl is unchanged.

A local backport script that I had applied on 2026.5.4 (mirror of PR #77839) was wiped by the upgrade as expected, but had also been a no-op for this symptom — which makes sense given the gate is the issue, not the helpers.

Reproduction

1. Install 2026.5.6, run gateway under LaunchAgent / launchd on macOS.
2. Open an embedded webchat session, send a few prompts (no model override needed — see #77518 update on 2026-04-25 for context: this fires on clean sessions, not just post-override).
3. Watch ~/.openclaw/logs/gateway.log for cli session reset lines with reason=missing-transcript.

In our case the symptom appears Mac-specific: a separate Linux host running the same 2026.5.4/2026.5.6 dist with comparable webchat traffic shows zero missing-transcript events over a 7-day span. Whatever environment difference triggers the runner-tag mismatch on Mac, the gate in agent-command*.js line 883 is the underlying mechanism either way.

Ask

Either:

• Confirm whether PR #77839 was intended to ship in 2026.5.6 and if so investigate why the dist doesn't reflect the change, or
• Reopen #77518 and treat the closure as premature — the symptom is unchanged on the latest release dist.

Happy to attach the gateway log slice (/tmp/openclaw/openclaw-2026-05-06.log filtered to reason=missing-transcript) and diff the patched-vs-unpatched files if useful.

References: #77518, #77839