Strip OpenClaw channel metadata from chat turn user text

[Jurij89]

Summary

• Strip leading OpenClaw Telegram metadata blocks from persisted user-side chat-turn text in both W4a computeDelta and W4b message-hook persistence.
• Keep runtime model input behavior unchanged: the Telegram/OpenClaw channel wrapper is not modified, and stripping happens only inside adapter chat-turn persistence.
• Preserve assistant-side stripRecalledMemory behavior while adding user-side coverage for Telegram metadata pollution from Telegram chat turns persist OpenClaw channel metadata blocks alongside user text, polluting chat-turn memory and recall #380.

Related

• Fixes Telegram chat turns persist OpenClaw channel metadata blocks alongside user text, polluting chat-turn memory and recall #380
• Related to Fix Telegram chat turn persistence #362
• Related to Fix OpenClaw adapter stateDir isolation and rename workspace scratch dir #364

Files changed

File	What
packages/adapter-openclaw/src/ChatTurnWriter.ts	Adds Telegram-gated stripChannelMetadata and applies it before W4a user pairing and W4b inbound queueing so persisted user text and content-derived identity use stripped text for Telegram runtime wrappers.
packages/adapter-openclaw/test/ChatTurnWriter.test.ts	Adds T380 regressions for Telegram metadata stripping, preservation of real user text, W4b paths, stripped-text turnId hashing, and top-pasted metadata-shaped content safety.

Metadata stripping behavior

• Runs only when trusted OpenClaw channel context identifies the source as Telegram.
• Removes contiguous leading fenced JSON metadata blocks labelled Conversation info and optional Sender when they use the (untrusted metadata) heading form.
• Removes separator blank lines after stripped blocks, while preserving indentation on the first real user line.
• Preserves pure user text, multi-line text, metadata-shaped content in the middle, leading metadata-shaped content outside Telegram, repeated metadata labels after the Telegram wrapper, and broader labels such as Channel context after the wrapper.
• Does not strip Channel context or Message context without a concrete trusted wrapper source; those labels remain extensibility points rather than active stripping rules.
• Does not change the OpenClaw Telegram channel plugin or the raw runtime context seen by the agent before persistence.

Watermark/hash decision and migration note

• Stripping happens before W4a pending-user storage and before W4b inbound queueing for Telegram, so deterministicTurnId, contentHash, cross-path stamps, and typed W4b markers use the stripped persisted user text going forward.
• Numeric W4a watermarks and W4b counts are unchanged because they are pair-index/count based, not text-hash based.
• Existing persisted turns are not rewritten. If reset/compaction or state loss leaves only pre-fix metadata-hashed durable markers, those legacy markers may not suppress a stripped replay; PR Fix Telegram chat turn persistence #362's cold-start clamp still bounds W4a state-loss replay to the latest pair.
• Known ambiguity: without changing the Telegram plugin to add an out-of-band sentinel, a Conversation-only runtime wrapper followed immediately by user text that starts with a first Sender metadata-shaped block is indistinguishable from the current Conversation info + Sender runtime wrapper shape. The implementation chooses the issue's required wrapper stripping while bounding repeated/broader-label loss cases in tests.

Privacy notes

• Persisted userText no longer carries inline Telegram chat_id, message_id, sender_id, sender name, username, or timestamp from the channel metadata wrapper.
• Routing/session metadata outside userText remains unchanged and is out of scope for Telegram chat turns persist OpenClaw channel metadata blocks alongside user text, polluting chat-turn memory and recall #380.

Known risks or follow-ups

• No migration is included for already-persisted metadata-polluted turns; Telegram chat turns persist OpenClaw channel metadata blocks alongside user text, polluting chat-turn memory and recall #380 marked historical rewrite out of scope.
• The full adapter package test still has two unrelated baseline failures outside this diff, listed below.

Test plan

• pnpm install - completed; emitted non-blocking dkg-mcp bin warnings for an unbuilt MCP package and pnpm ignored-build-script notices.
• pnpm --filter @origintrail-official/dkg-adapter-openclaw exec vitest run test/ChatTurnWriter.test.ts -t T380 - passed, 16 tests.
• pnpm --filter @origintrail-official/dkg-adapter-openclaw exec vitest run test/ChatTurnWriter.test.ts - passed, 187 tests.
• pnpm --filter @origintrail-official/dkg-core run build - passed.
• pnpm --filter @origintrail-official/dkg-adapter-openclaw run build - passed.
• pnpm --filter @origintrail-official/dkg-adapter-openclaw exec vitest run test/ChatTurnWriter.test.ts test/ChatTurnWriter.crashRecovery.test.ts test/ChatTurnWriter.sanitize.test.ts - passed, 198 tests.
• git diff --check - passed with Windows LF-to-CRLF checkout warnings only.
• pnpm --filter @origintrail-official/dkg-adapter-openclaw test - ran 984 tests: 981 passed, 1 skipped, 2 failed in unrelated baseline tests: test/adapter-openclaw-extra.test.ts > plugin id equals package name expects @origintrail-official/dkg-adapter-openclaw but receives adapter-openclaw; test/setup.test.ts > writeDkgConfig > mirrors only autoUpdate.enabled from network default and preserves existing pins expects autoUpdate.repo to remain present.

[github-actions]

🔴 Bug: This heuristic strips any Telegram message that starts with one of these fenced blocks, even if the user authored it intentionally. A user debugging Telegram metadata and pasting a single Conversation info ... block as the first part of their message will have that content silently removed from persistence. Please gate stripping on a trusted adapter signal (or validate the parsed block against envelope metadata) instead of matching only on user-writable text.