[Bug]: 2026.5.19 Discord shows failed file reads in chat

Existing issue check:

• Related open issue: https://github.com/openclaw/openclaw/issues/44905
• Related open issue: https://github.com/openclaw/openclaw/issues/78561
• Related open issue for the same class on Matrix: https://github.com/openclaw/openclaw/issues/85000
• I did not find an exact open Discord issue for visible print lines ... failed messages from failed file reads.

Bug type

Behavior bug

Beta release blocker

No

Summary

Discord can show internal failed file-read messages such as print lines 1-220 from memory/2026-05-22.md (agent) failed or print lines 1-220 from ~/.openclaw/workspace/skills/gemini/SKILL.md (agent) failed as normal app output.

Steps to reproduce

1. Run OpenClaw 2026.5.19 with Discord enabled.
2. Start an agent turn from a Discord-delivered workflow where the agent tries to read a missing workspace file.
3. In this install, the observed workflow was a scheduled Discord channel turn.
4. Observe the Discord channel after the turn finishes.

Expected behavior

Discord should show the user-facing reply or job output only. Failed internal file-read attempts should stay in logs or diagnostics and should not appear as app messages in the channel.

Actual behavior

Discord showed this internal failed tool message:

⚠️ 🛠️ `print lines 1-180 from memory/2026-05-20.md (agent)` failed

The same bug is still present after upgrading to 2026.5.20:

⚠️ 🛠️ `print lines 1-220 from memory/2026-05-22.md (agent)` failed

The same visible-failure shape was also observed for a skill instruction file:

⚠️ 🛠️ `print lines 1-220 from ~/.openclaw/workspace/skills/gemini/SKILL.md (agent)` failed

OpenClaw version

2026.5.19 (a185ca2)

Operating system

macOS 26.5 (25F71)

Install method

Source checkout with pnpm build. Gateway runs from the built source checkout.

Model

openai/gpt-5.5

Provider / routing chain

Discord -> OpenClaw Gateway -> Codex runtime -> openai/gpt-5.5

Additional provider/model setup details

Relevant local config:

agents.defaults.model.primary = openai/gpt-5.5
agents.defaults.models.openai/gpt-5.5.agentRuntime.id = codex
channels.discord.streaming.mode = off
messages.groupChat.visibleReplies = message_tool

Logs, screenshots, and evidence

OpenClaw version and source state at the time of the primary observed failure:

$ openclaw --version
OpenClaw 2026.5.19 (a185ca2)

$ node --version
v22.22.3

$ git -C openclaw-src describe --tags --exact-match
v2026.5.19

$ git -C openclaw-src rev-parse --short=12 HEAD
a185ca283a74

The primary 2026.5.19 failure came from an agent attempt to read a missing memory file:

sed -n '1,220p' memory/2026-05-20.md
sed: memory/2026-05-20.md: No such file or directory

The relevant file was missing on disk:

$ ls -l ~/.openclaw/workspace/memory/2026-05-20.md
ls: ~/.openclaw/workspace/memory/2026-05-20.md: No such file or directory

A later 2026.5.20 run showed the bug is still present. Local launchd job log shows the user-facing scheduled summary followed by the failed internal read message:

**Scheduled summary: 3 item(s) need attention**

Summary body omitted here because it contains site-specific operational details.

[normal user-facing summary content]

⚠️ 🛠️ `print lines 1-220 from memory/2026-05-22.md (agent)` failed
delivered scheduled summary

Gateway log for the same event:

<timestamp> ⚠️ 🛠️ `print lines 1-220 from memory/2026-05-22.md (agent)` failed

Impact and severity

Affected: Discord channels on this install.

Severity: Medium annoyance. The main work can still complete, but Discord shows internal failed tool messages that look like app errors.

Frequency: Seen on this install across 2026.5.19 and still present in 2026.5.20, with different memory files.

Consequence: Users see noisy implementation details in chat, and routine automation or skill-backed turns can look broken even when the useful output was delivered.

Additional information

This looks related to the broader Discord internal-output leak tracked in #44905, but the visible symptom here is narrower and easy to search for: failed file reads are rendered as print lines ... failed app messages.

Local workaround for the scheduled job was to stop using an agent turn for a simple summary and send the formatted summary directly. That avoids this one job path, but it does not address Discord showing failed file-read messages from other agent turns.