Observer: \\\\s (4-backslash) → \\s (2-backslash) in template literal regex
- Per known-issues rule: regex literals need 2-backslash for \s, \d etc.
- Old: literal \\s matched, new: whitespace class matched correctly
http-bridge: /^[\s\\\/]*[\w_.-]+\s*[>»$#]\s*$/ → /^.*[>»$#]\s*$/
- Old pattern failed on ellipsis '…' prefix (U+2026 not in charset)
- Terminal prompt '…\\gravity_control >' was passing through as cmd
- New pattern: any text ending with prompt marker is skipped
- extractContextFromNearby: PROMPT_ONLY_RE skips code elements containing only terminal prompts (e.g. '\\gravity_control >')
- Multi-codeEl traversal: tries all code elements at each depth, picks longest non-prompt text
- http-bridge enrichment: validates extracted command is not just a prompt fragment before enriching
- Fixes: cmd='\\gravity_control >' (prompt-only) no longer sent to Discord as the command text
- extractContextFromNearby: span/div/p text fallback when pre/code not found (depth 0-8)
- collectSiblingButtons: extended to 5 parent levels, stop only when both action+reject found
- http-bridge: command enrichment — swap generic button text with actual command from description
- Fixes: cmd='Always run' now shows actual command text, Cancel button detection improved
Root cause: regex literals inside template literal (\...\) need exactly 2
backslashes (\\s, \\d) to produce \s, \d in the output. The v9 code had 4
backslashes (\\\\s) which produced literal \\s in the browser (matching
backslash+s instead of whitespace class). This broke:
- 'Running N commands' skip pattern (never matched)
- NOISE_CODE_RE (declare/import line filters)
- cleanLines() pre-strip word boundary \\b and split/join \\n
- cleanButtonText() whitespace strip \\s
The fix restores all regex-literal patterns to exactly 2 backslashes.
new RegExp() string patterns (like NOISE_RE) already used 4 backslashes
correctly (string→regex has one extra escaping layer).
Verified: node eval of generated code confirms /^Running\s*\d+\s*commands?$/i
correctly matches 'Running2 commands' and 'Running 4 commands'.
#task-619 #task-620
- http-bridge.ts: add req.setEncoding('utf8') to all POST handlers
to fix Korean text corruption in pending/chat/dump payloads
- observer-script.ts: add inline pre-strip in cleanLines() for
Material icon names concatenated by textContent without newlines
- observer-script.ts: apply cleanLines() to codeText extraction
- known-issues: document UTF-8 encoding and noise filter issues
- html-patcher: relocate inline script from after </html> to before </body>
- html-patcher: clean up duplicate </html> tags from previous bad insertions
- observer-script: add immediate BEACON fetch to /ping on script load
- http-bridge: add diagnostic request logging for non-polling endpoints
- devlog 003: crash recovery session notes
- Replace CSS class-based scanning with [data-testid='conversation-view'] + [data-step-index] traversal
- New extractCleanStepText(): clone-and-strip buttons/SVG/icons before text extraction
- New extractStepContext(): step-container-aware context with header + code block
- NOISE_RE: block Material icon names, button labels, UI artifacts
- Auto DOM structure dump on first conversation-view detection
- Enhanced deep-inspect with step element + button inventory
- known-issues: document AG Native SDK API incompatibility
* Restore step.content.parts traversal missing in prior bugfix
* Catch wide exceptions in bot.py chat_snapshot_scanner and move broken files to .json.failed to prevent loop aborts blocking the pending queue
- removed PATS in observer-script.ts
- step-probe.ts now handles 100% of pending detection with full RPC payload context
- DOM observer restricted to trigger-click polling only
* Removed writeChatSnapshot calls in step-probe.ts to prevent duplicate '⚡ 자동 실행됨' notifications since bot.py already broadcasts '🤖 자동 승인됨'.
* Docs: Update devlog and known-issues with Discord Bot cache deletion bugs and multi-workspace LS connection conflicts.
* Fixes Vikunja #593
