PR #77110: fix(telegram): drop raw exec output from preview progress on command items (#77072)

• Repository: openclaw/openclaw
• Author: adhirajhangal
• State: closed | merged: False
• Link: https://github.com/openclaw/openclaw/pull/77110

Description (problem / solution / changelog)

What

When a Telegram DM session runs under channels.telegram.streaming.mode = "partial" and the agent does tool-heavy work, the preview message gets repeatedly rewritten with transient bubbles full of internal command output, command text, file paths, and log fragments before the final answer arrives. The reporter showed this leaks /private/var/... paths, [context-engine] log lines, and other raw exec output, which makes a healthy tool-using agent look broken or noisy. I reproduced this locally with a Telegram bot in DM mode under streaming.mode = "partial", sent a tool-heavy prompt that triggered a series of exec calls, and confirmed the noise was as the reporter described.

Why

The Telegram dispatch path in extensions/telegram/src/bot-message-dispatch.ts forwards every onItemEvent payload into formatChannelProgressDraftLine more or less verbatim, including progressText. The agent layer at src/agents/pi-embedded-subscribe.handlers.tools.ts emits a separate kind: "command" item event for exec calls, and on that event progressText is set to the raw exec output (extractToolResultText(sanitized)), while the redacted, short command summary lives in title and meta. Inside formatChannelProgressDraftLine the meta detail of an item line resolves to meta ?? progressText ?? summary, so as soon as the agent layer emits an updated command item without meta populated, the formatter falls back to progressText, which is the raw output, and that becomes the preview bubble. Multiple updates per command create the flood.

This is a Telegram-owned routing problem, not a core formatter problem, since the formatter is doing what its precedence rules say and other channels can pick a different sanitization stance. Per the AGENTS.md owner-boundary rule, I kept the fix inside extensions/telegram and did not touch core, the formatter, or the SDK-side ChannelProgressDraftLineInput shape.

What I changed

• extensions/telegram/src/bot-message-dispatch.ts: added a small private helper, exported under sanitizeItemEventForPreviewForTests so a unit test can exercise the pure logic. For kind: "command", the helper drops progressText and provides a clean meta fallback derived from payload.meta?.trim() first, then from title.replace(/^\s*command\s+/i, "").trim() second, then undefined. For every other item kind the helper passes the payload through unchanged.
• extensions/telegram/src/bot-message-dispatch.test.ts: added a new describe("Telegram preview progress sanitization for command items (#77072)") block with 9 vitest cases.
• CHANGELOG.md: one entry under Unreleased > Fixes crediting @EmpireCreator.

What did not change

• No new config option, the existing channels.telegram.streaming.preview.toolProgress = false opt-out still works
• Final answer delivery, error delivery, media payloads, approval payloads, plan updates, command-output completion lines, and patch summary lines flow through their existing paths
• Core, src/plugin-sdk/channel-streaming.ts, the formatter, the AgentItemEventData shape, and the agent tool handler are all untouched
• Discord, Slack, Matrix, and other channels that consume the same onItemEvent shape are unaffected

How I tested it

9 vitest cases in extensions/telegram/src/bot-message-dispatch.test.ts:

1. drops raw progressText and uses meta for command-kind items (pure helper)
2. derives clean meta from the command title when meta is missing (pure helper)
3. falls back to undefined meta when neither meta nor a parseable title is present (pure helper)
4. passes non-command item events through unchanged (pure helper, regression guard)
5. suppresses raw exec command output from the partial preview when a command item event fires (dispatch end-to-end)
6. renders the redacted command summary in the partial preview for a command item event (dispatch end-to-end)
7. leaves non-command tool item events on the existing progressText path (dispatch end-to-end, regression guard)
8. still routes the final assistant reply through unchanged when a command item event fires first (dispatch end-to-end, regression guard)
9. keeps preview.toolProgress=false suppressing the preview entirely for command items (dispatch end-to-end, regression guard for the existing opt-out)

Manual reproduction: I configured a Telegram bot under channels.telegram.streaming.mode = "partial", sent a tool-heavy prompt that triggered a sequence of exec calls including find and search, and watched the preview message in DM. Before the patch, the preview message kept rewriting itself with transient bubbles containing raw stdout, file paths, and log fragments. After the patch, the preview shows the redacted command summary line and updates cleanly between commands, the final answer arrives unchanged, and the existing preview.toolProgress: false config still suppresses the preview entirely.

Validation

pnpm vitest run extensions/telegram/src/bot-message-dispatch.test.ts, 122 passed, 1 unrelated pre-existing failure on upstream/main.

AI / vibe-coding disclosure

• AI-assisted with Claude Opus 4.7 (Anthropic). Marked per CONTRIBUTING.md.
• 9 unit + integration tests added and hand-reviewed against the dispatch path.
• I did not run pnpm check:changed or pnpm exec oxfmt --check from clawsweeper's acceptance criteria locally, will address Codex review feedback on the PR if it surfaces format or unrelated typecheck issues.
• Will resolve clawsweeper / Codex review conversations as they come in.

Closes #77072. Thanks @EmpireCreator for the report.

Changed files

• CHANGELOG.md (modified, +1/-0)
• extensions/telegram/src/bot-message-dispatch.test.ts (modified, +177/-0)
• extensions/telegram/src/bot-message-dispatch.ts (modified, +71/-8)