[Bug] TUI crashes with URIError in cellMarker when rendering markdown tables

[vichaunter]

The TUI crashes with URIError: URI malformed when rendering a markdown table whose cell content contains lone surrogates or characters that encodeURIComponent cannot handle. The crash kills the entire session with no recovery.

URIError: URI malformed
    at encodeURIComponent (<anonymous>)
    at cellMarker (chunk-MMG2DX73.js:26349:56)
    at markTableCell (chunk-MMG2DX73.js:26376:10)
    at Markdown.renderTable (chunk-MMG2DX73.js:27155:35)
    at Markdown.renderToken (chunk-MMG2DX73.js:26790:33)
    at Markdown.renderBlock (chunk-MMG2DX73.js:26647:29)
    at Markdown.render (chunk-MMG2DX73.js:26616:27)
    at Container.render (chunk-MMG2DX73.js:18807:32)

Root cause: cellMarker() calls encodeURIComponent(content) unconditionally on table cell text:

function cellMarker(kind: CellMarker["kind"], row: number, column: number, segment: number, content?: string): string {
  const encodedContent = content === undefined ? "" : `:${encodeURIComponent(content)}`;
  return `${TABLE_MARKER_PREFIX}${kind}:${row}:${column}:${segment}${encodedContent}\x07`;
}

(Called from markTableCell().)

Per ECMA-262 §19.2.6.5, encodeURIComponent must throw URIError on lone surrogates (U+D800–U+DFFF). This is not Node.js-specific or platform-specific — it's mandated by the JS spec and applies to every conformant engine (V8, SpiderMonkey, JSC), every Node version, and every OS.

Lone surrogates can reach agent-produced strings through several paths:

• Tool output containing partially decoded binary data
• JSON.parse of strings with escaped lone surrogates (\uD800)
• TextDecoder with invalid UTF-8 sequences (non-fatal mode, the default)

Steps to reproduce

1. Start a session where an agent produces markdown tables from external data or tool output
2. When a table cell contains lone surrogates (U+D800–U+DFFF), the TUI crashes immediately
3. The crash kills the entire session — no recovery

Minimal reproduction:

encodeURIComponent('\uD800'); // URIError: URI malformed

Expected behavior

The TUI should gracefully handle any string content in markdown table cells without crashing. A try/catch around encodeURIComponent with a fallback that replaces lone surrogates (e.g. with U+FFFD) would prevent the crash without affecting normal table rendering.

Environment

• Prime Agent version: 0.9.1
• Installation method: Stable release installer
• Operating system: Linux / Ubuntu 24.04
• CPU architecture: x86_64 / amd64
• Shell and terminal: bash 5.2 in xterm-256color

Additional context

This crash has occurred multiple times across different sessions. It's not deterministic — it depends on the agent's output containing specific characters in a table cell. A local workaround is to patch cellMarker in the dist bundle to wrap encodeURIComponent in try/catch and replace lone surrogates with U+FFFD, but this gets overwritten on every prime-agent update.

[tttboy123]

I reproduced this against upstream main at 1c07eaa.

The public table-selection path still reaches the unconditional encodeURIComponent(content) call: passing an unpaired high or low UTF-16 surrogate through markTableCell throws URIError: URI malformed before the table can render.

I prepared a focused candidate change that replaces only unpaired surrogates with U+FFFD before encoding while preserving valid surrogate pairs. The regression test exercises the public markTableCell → extractTableCellSelectionRegions path with an unpaired high surrogate, an unpaired low surrogate, and a valid emoji pair.

Red/green evidence:

• Before the change, the focused test failed at selection-metadata.ts:38 with URIError: URI malformed.
• After the change, the focused test passed: 1 test, 0 failures.
• Repository npm run check passed, including formatting, linting, type checking, installer checks, and browser smoke checks.

Candidate branch: tttboy123:fix/2001-tui-lone-surrogates

Candidate commit: 271bc49

I have not opened an upstream PR. If maintainers want this fix, I can add the required TUI changelog fragment and open a focused PR after an explicit invitation, per CONTRIBUTING.md.

[tttboy123]

Follow-up: I expanded the candidate into a complete fork-local Draft PR for review:

tttboy123/prime-agent#1

The candidate now:

• isolates surrogate normalization in a named table-marker encoding boundary;
• replaces only unpaired high/low UTF-16 surrogates with U+FFFD;
• preserves valid surrogate pairs and the existing marker format/API;
• covers the public marker round trip;
• adds an end-to-end Markdown.render() → getSelectionRegions() regression; and
• includes the required TUI changelog fragment.

Both regressions were verified red against upstream main (1c07eaa) with URIError: URI malformed, then green on fork commit f458aec.

Fresh validation on the final commit:

• marker regression: 1 passed, 0 failed;
• end-to-end Markdown regression: 1 passed, 0 failed;
• npm run check: passed;
• clean working tree, with the remote branch at the same commit.

The Draft PR targets only the fork's main; no upstream PR has been opened. If maintainers want this fix, I can update the changelog attribution to the upstream PR number and submit the focused upstream PR after an explicit invitation.