PR #74753: fix(agents): strip replay metadata placeholder from user-facing text

• Repository: openclaw/openclaw
• Author: Marvae
• State: closed | merged: False
• Link: https://github.com/openclaw/openclaw/pull/74753

Description (problem / solution / changelog)

Summary

• Problem: Internal replay placeholder [assistant copied inbound metadata omitted] is leaking into outbound Telegram messages after 2026.4.27
• Why it matters: Users see internal scaffolding text instead of (or in addition to) the actual assistant response
• What changed: Extended the existing replay placeholder stripping to also filter this metadata placeholder
• What did NOT change: Inline mentions of the placeholder (e.g. user asking "what does this mean?") are preserved

Regression: Introduced in ee8f41f56e (2026-04-26) which added the placeholder in replay-history.ts but did not add corresponding filtering in sanitize-user-facing-text.ts.

Change Type (select all)

• Bug fix
• Feature
• Refactor required for the fix
• Docs
• Security hardening
• Chore/infra

Scope (select all touched areas)

• Gateway / orchestration
• Skills / tool execution
• Auth / tokens
• Memory / storage
• Integrations
• API / contracts
• UI / DX
• CI/CD / infra

Linked Issue/PR

• Closes #74745
• Related to ee8f41f56e (commit that introduced the placeholder)
• This PR fixes a bug or regression

Root Cause (if applicable)

• Root cause: normalizeAssistantReplayContent in replay-history.ts synthesizes [assistant copied inbound metadata omitted] when stripped assistant replay text becomes empty, but sanitizeUserFacingText only stripped [tool calls omitted] placeholder lines
• Missing detection / guardrail: The new placeholder was not added to the outbound sanitizer
• Contributing context: If LLM echoes the replay history or certain edge cases occur, the placeholder leaks to user-facing delivery

Regression Test Plan (if applicable)

• Coverage level:
  • Unit test
• Target test file: src/agents/pi-embedded-helpers.sanitizeuserfacingtext.test.ts
• Scenario the test should lock in: Standalone [assistant copied inbound metadata omitted] lines are stripped, inline mentions are preserved
• Why this is the smallest reliable guardrail: Tests the central sanitizer that all outbound paths use
• Existing test that already covers this: None for this specific placeholder

User-visible / Behavior Changes

• Internal placeholder text no longer appears in delivered messages
• Users receive actual assistant responses instead of scaffolding text

Diagram (if applicable)

Before:
[replay-history synthesizes placeholder] -> [sanitizer misses it] -> [user sees "[assistant copied inbound metadata omitted]"]

After:
[replay-history synthesizes placeholder] -> [sanitizer strips it] -> [user sees actual response or empty string]

Security Impact (required)

• New permissions/capabilities? No
• Secrets/tokens handling changed? No
• New/changed network calls? No
• Command/tool execution surface changed? No
• Data access scope changed? No

Repro + Verification

Environment

• OS: macOS Darwin 25.3.0
• Runtime/container: Node.js v22.22.0
• Model/provider: Any
• Integration/channel: Telegram (reported), affects all channels

Steps

1. Trigger a session where replay-history synthesizes the metadata placeholder
2. Observe outbound message delivery

Expected

• Placeholder is stripped; user sees actual response

Actual

• Works as expected after fix

Evidence

• Failing test/log before + passing after

pnpm test src/agents/pi-embedded-helpers.sanitizeuserfacingtext.test.ts
# All tests pass including new regression tests

New tests verify:

expect(sanitizeUserFacingText("[assistant copied inbound metadata omitted]")).toBe("");
expect(sanitizeUserFacingText("Hello\n\n[assistant copied inbound metadata omitted]\nWorld\n")).toBe("Hello\n\nWorld\n");
// Inline mentions preserved
expect(sanitizeUserFacingText("What does [assistant copied inbound metadata omitted] mean?"))
  .toBe("What does [assistant copied inbound metadata omitted] mean?");

Human Verification (required)

• Verified scenarios: Unit tests pass, code review, regex logic inspection
• Edge cases checked: Standalone line, multiple lines, inline mention preservation
• What you did NOT verify: E2E session replay (unit tests cover the sanitizer path)

Review Conversations

• I replied to or resolved every bot review conversation I addressed in this PR.
• I left unresolved only the conversations that still need reviewer or maintainer judgment.

Compatibility / Migration

• Backward compatible? Yes
• Config/env changes? No
• Migration needed? No

Risks and Mitigations

• Risk: Regex could over-match and strip legitimate user content
  • Mitigation: Only matches exact standalone placeholder lines (with optional whitespace); inline mentions are preserved by design and tested

Changed files

• src/agents/pi-embedded-helpers.sanitizeuserfacingtext.test.ts (modified, +16/-0)
• src/agents/pi-embedded-helpers/sanitize-user-facing-text.ts (modified, +5/-4)
• src/auto-reply/reply/reply-utils.test.ts (modified, +16/-0)

PR #74785: fix(pi-embedded): strip [assistant copied inbound metadata omitted] placeholder from delivery

• Repository: openclaw/openclaw
• Author: hclsys
• State: closed | merged: False
• Link: https://github.com/openclaw/openclaw/pull/74785

Description (problem / solution / changelog)

Problem

The internal replay placeholder [assistant copied inbound metadata omitted] is leaking into outbound Telegram (and other channel) messages after the 2026.4.27 update, appearing as the entire message body instead of the actual assistant reply.

Fixes #74745.

Root cause

sanitize-user-facing-text.ts already strips [tool calls omitted] from user-facing output (landed in d30b8dccfd, fixing #74573). The same sentinel pattern OMITTED_INBOUND_METADATA_TEXT = "[assistant copied inbound metadata omitted]" synthesized by normalizeAssistantReplayContent in replay-history.ts was not covered by the stripping regex.

Solution

Extend TOOL_CALLS_OMITTED_PLACEHOLDER_LINE_RE to cover both placeholder variants using an alternation group:

-const TOOL_CALLS_OMITTED_PLACEHOLDER_LINE_RE = /^[ \t]*\[tool calls omitted\][ \t]*$/i;
+const TOOL_CALLS_OMITTED_PLACEHOLDER_LINE_RE =
+  /^[ \t]*\[(?:tool calls omitted|assistant copied inbound metadata omitted)\][ \t]*$/i;

Same standalone-line logic applies: only strip when the placeholder is the entire line content, so ordinary user discussion of the placeholder text is preserved.

Tests

Added 4 test assertions mirroring the existing [tool calls omitted] coverage — standalone strip, whitespace strip, mid-text strip, and inline-mention preservation. All 90/90 tests pass.

Changed files

• CHANGELOG.md (modified, +1/-0)
• src/agents/pi-embedded-helpers.sanitizeuserfacingtext.test.ts (modified, +19/-0)
• src/agents/pi-embedded-helpers/sanitize-user-facing-text.ts (modified, +4/-3)

PR #74828: fix: regression/data-loss edge case in the new metadata-only assistant drop path

• Repository: openclaw/openclaw
• Author: clawsweeper[bot]
• State: closed | merged: True
• Link: https://github.com/openclaw/openclaw/pull/74828

Description (problem / solution / changelog)

Summary

Found one regression/data-loss edge case in the new metadata-only assistant drop path.

What ClawSweeper Is Fixing

• Medium: Dropping metadata-only assistant turns can erase adjacent string user prompts during Anthropic replay validation (regression)
  • File: src/agents/pi-embedded-runner/replay-history.ts:321
  • Evidence: The new path returns null when an assistant block array becomes empty after stripInboundMetadata, so a history like user("First") -> assistant(metadata-only) -> user("Second") becomes two adjacent user turns. The next runtime boundary calls sanitizeSessionHistory and then validateReplayTurns in src/agents/pi-embedded-runner/run/attempt.ts:1918 and src/agents/pi-embedded-runner/run/attempt.ts:1933. For Anthropic replay, mergeConsecutiveUserTurns only preserves array content and drops string content via Array.isArray(...) ? ... : [] in src/agents/pi-embedded-helpers/turns.ts:352. A focused probe reproduced this exact loss:
  • Impact: Common persisted user messages use string content. After this commit, a metadata-only assistant replay turn between two string user turns can make Anthropic-class replay send an empty user message instead of the actual prior/latest prompts, causing bad replies or provider rejection. The new regression test covers only array-shaped user content at src/agents/pi-embedded-runner.sanitize-session-history.test.ts:1143, so it misses this production-shaped case.
  • Suggested fix: Preserve string user content when merging consecutive user turns, e.g. normalize strings into { type: "text", text } blocks before concatenation, and add a regression test using string user messages around the dropped metadata-only assistant turn.
  • Confidence: high

Expected Repair Surface

• src/agents/pi-embedded-runner/replay-history.ts
• src/agents/pi-embedded-runner/replay-history.test.ts
• src/agents/pi-embedded-runner.sanitize-session-history.test.ts

Source And Review Context

• ClawSweeper report: https://github.com/openclaw/clawsweeper/blob/main/records/openclaw-openclaw/commits/d80a8eb3adb6b26fb345fdc7b658dbd39f085e35.md

• Commit under review: https://github.com/openclaw/openclaw/commit/d80a8eb3adb6b26fb345fdc7b658dbd39f085e35

• Latest main at intake: 0603c2327d3e09540b4488596aa3062c6acfc550

• Original commit author: Peter Steinberger

• GitHub author: @steipete

• Highest severity: medium

• Review confidence: high

• Diff: bb449092626193dae04e91f9124caaf821caec1f..d80a8eb3adb6b26fb345fdc7b658dbd39f085e35

• Changed files: CHANGELOG.md, src/agents/pi-embedded-runner/replay-history.ts, src/agents/pi-embedded-runner/replay-history.test.ts, src/agents/pi-embedded-runner.sanitize-session-history.test.ts

• Code read: changed files in full, src/auto-reply/reply/strip-inbound-meta.ts, src/agents/pi-embedded-helpers/turns.ts, src/agents/pi-embedded-runner/run/attempt.ts, relevant validator tests

• Issue context: https://github.com/openclaw/openclaw/issues/74745

Expected validation

• pnpm check:changed

ClawSweeper already ran:

• pnpm docs:list
• pnpm install
• OPENCLAW_VITEST_MAX_WORKERS=1 pnpm test src/agents/pi-embedded-runner/replay-history.test.ts src/agents/pi-embedded-runner.sanitize-session-history.test.ts passed: 67 tests
• node --import tsx -e ...normalizeAssistantReplayContent + validateAnthropicTurns... reproduced the string-user content loss
• git diff --check bb449092626193dae04e91f9124caaf821caec1f d80a8eb3adb6b26fb345fdc7b658dbd39f085e35 passed
• One earlier test attempt failed because I accidentally passed Jest’s --runInBand; it was rerun successfully with the repo’s Vitest-compatible worker setting.

Known review limits:

• Did not run a live Telegram or Anthropic provider send; the finding is based on source trace plus a focused local replay-normalization probe.

ClawSweeper Guardrails

• Re-check the finding against latest main before changing code.
• Keep the patch to the narrowest behavior change and matching regression coverage.
• Do not merge automatically; this PR stays for maintainer review.

ClawSweeper 🐠 replacement reef notes:

• Cluster: clawsweeper-commit-openclaw-openclaw-d80a8eb3adb6
• Source PRs: none
• Credit: Detected by ClawSweeper commit review for d80a8eb3adb6b26fb345fdc7b658dbd39f085e35.; Original commit author: Peter Steinberger.
• Validation: pnpm check:changed

fish notes: model gpt-5.5, reasoning medium; reviewed against 99a6a311a9aa.

Changed files

• src/agents/pi-embedded-helpers.validate-turns.test.ts (modified, +20/-0)
• src/agents/pi-embedded-helpers/turns.ts (modified, +16/-2)
• src/agents/pi-embedded-runner.sanitize-session-history.test.ts (modified, +2/-10)