PR #68125: fix: Control UI runtime content filtering (#68094)

• Repository: openclaw/openclaw
• Author: SARAMALI15792
• State: open | merged: False
• Link: https://github.com/openclaw/openclaw/pull/68125

Description (problem / solution / changelog)

Fix: Control UI Runtime Content Filtering

Problem Statement

When a new session starts (/new or /reset), the startup context prelude (daily memory notes, persona instructions) is concatenated directly into the user message body. Similarly, system events (heartbeat wakes, cron triggers) are prepended to user messages.

The Control UI renders all user turns in full with no filtering, causing runtime-injected content to appear as visible chat bubbles — as if the user typed them.

Impact: Confusing UX — users see internal runtime metadata displayed as their own messages.

Solution

Two-layer defense-in-depth filtering at the display boundary:

Layer 1: Gateway Stripping (Primary)

stripRuntimeInjectedContent() in src/gateway/server-methods/chat.ts:

• Detects and strips startup context blocks ([Startup context loaded by runtime])
• Detects and strips system event lines (System: [timestamp] / System (untrusted): [timestamp])
• Removes messages entirely if empty after stripping
• Handles all three message formats: content array, content string, text field
• Preserves non-runtime text blocks in multi-block messages (e.g., image + text)
• Preserves legitimate whitespace when no runtime content is present
• Anchors startup context detection to leading position only (prevents false positives)

Layer 2: UI Filtering (Defense-in-Depth)

shouldHideHistoryMessage() in ui/src/ui/controllers/chat.ts:

• Detects both startup context prefix and system event patterns
• Filters runtime-injected messages client-side if gateway stripping fails
• Matches existing NO_REPLY filtering pattern

Architecture

User Message Flow:
  Session Start → Startup Context Injected → Stored in Transcript
                                                    ↓
  chat.history Request → stripRuntimeInjectedContent() → Stripped Message
                                                    ↓
  Control UI → shouldHideHistoryMessage() → Clean Display

Key Design Decisions:

• No changes to message storage or prompt pipeline (model still receives full content)
• Content-pattern detection at display boundary only
• Backward compatible (old sessions without injected content pass through unchanged)
• Reference equality preserved when no changes needed (avoids unnecessary re-renders)

Changes

Test Coverage

Gateway Tests (17 tests)

• ✅ Returns messages unchanged when no runtime content present
• ✅ Removes user message that is entirely startup context prelude
• ✅ Removes user message that is startup context + bare reset prompt
• ✅ Strips system event lines from beginning, keeps user text
• ✅ Strips untrusted system event lines
• ✅ Removes message entirely when only system event lines
• ✅ Handles string content field (not array)
• ✅ Handles text field (not content)
• ✅ Does not modify assistant messages
• ✅ Does not modify user messages without runtime patterns
• ✅ Returns same reference when nothing changes (performance)
• ✅ Strips combined startup context + system events
• ✅ Preserves non-runtime text blocks in content arrays
• ✅ Preserves leading/trailing whitespace when no runtime content present
• ✅ Does not strip startup context prefix when it appears mid-message
• ✅ Strips startup context prefix only when at leading position
• ✅ Does not strip startup context prefix with leading whitespace

UI Tests (7 tests)

• ✅ Filters user messages that are entirely startup context prelude
• ✅ Keeps user messages that do not start with startup context prefix
• ✅ Filters user messages that are only system event lines
• ✅ Filters user messages with untrusted system event lines
• ✅ Keeps user messages that start with system event but have user text
• ✅ Existing NO_REPLY filtering still works
• ✅ Existing synthetic tool result filtering still works

Verification

# Gateway tests
npx vitest run src/gateway/server-methods/chat.runtime-inject-strip.test.ts
# Result: 17/17 PASS

# UI tests
npx vitest run ui/src/ui/controllers/chat.test.ts
# Result: 76/76 PASS

Edge Cases Addressed

Outcomes

Before:

• Startup context prelude visible as user message bubble
• System event blocks visible as user text
• Multi-text-block messages could be corrupted during stripping
• Legitimate whitespace could be trimmed from normal messages

After:

• Runtime-injected content invisible in Control UI
• Clean chat history showing only user-typed messages
• Multi-block messages preserved correctly
• Whitespace integrity maintained
• Model still receives full injected content (no functional changes to prompt pipeline)
• Defense-in-depth: gateway strips, UI filters as backup

Fixes

Closes #68094

Changed files

• src/gateway/server-methods/chat.runtime-inject-strip.test.ts (added, +259/-0)
• src/gateway/server-methods/chat.ts (modified, +170/-1)
• ui/src/ui/controllers/chat.test.ts (modified, +126/-0)
• ui/src/ui/controllers/chat.ts (modified, +41/-1)