fix(ui-tui): render agent URLs as visible, clickable text

[agentforce314]

Summary

URLs the agent prints are clickable again. The TUI markdown renderer replaced every visible URL with a remote-fetched page <title> (silently HTTP-GETting each URL the agent printed) or a slug-derived label, leaving the real URL only in OSC 8 metadata. Terminals without OSC 8 support (e.g. Apple Terminal) strip that metadata, so the URL was invisible, unclickable, and uncopyable — and in the default inline mode the TUI never captures the mouse (the in-process click opener is alt-screen-gated), so terminal-native URL detection over the visible text is the only affordance that works everywhere. There was no visible URL for it to detect.

Reproduced by driving the real TUI under a PTY: the agent's reply rendered as You can find the documentation at Guide. — no URL anywhere on screen.

Changes in ui-tui/src/components/markdown.tsx:

• Bare URLs and autolinks render verbatim (accent + underline, still OSC 8-wrapped).
• [label](url) renders as label (url) with the target visible; no duplication when the label already is the URL/address.
• The remote page-title fetch is gone from the renderer — no more network request per rendered URL, and labels no longer mutate into page titles mid-conversation.

After the fix, the same PTY drive renders visit https://example.com/docs/guide. with the URL as plain visible text, still OSC 8-wrapped (Cmd+click in VS Code/iTerm2, plain click in fullscreen mode), trailing punctuation outside the link.

Test Coverage

All changed code paths have direct unit coverage — three tests in ui-tui/src/__tests__/markdown.test.ts pin the new contract (bare URL verbatim; label (url); no URL duplication for self-labeled links and mailto autolinks), replacing the two tests that pinned the old label-hiding behavior. End-to-end verification via a live PTY drive of the built TUI (before/after captures of raw terminal bytes incl. OSC 8 sequences).

Pre-Landing Review

Specialist dispatch (testing, maintainability, performance, design — 68-line diff) + checklist pass + Claude adversarial subagent (Codex CLI not installed). 10 findings, 0 critical:

• [AUTO-FIXED] (performance + maintainability, multi-specialist, 9/10) externalLink.ts — the title-fetch machinery orphaned by this fix (~430 lines: browser-UA HTTP fetch, HTML title parsing, SSRF guards, module-level caches) was production-dead, kept green only by its own tests. Deleted; normalizeExternalUrl retained. This also makes the CHANGELOG claim "the stealth title fetch is gone" literally true.
• [AUTO-FIXED] (design, 7/10) labeled mailto links leaked the scheme into the visible suffix (Email me (mailto:…) instead of the bare address); the suffix now renders through defaultLinkLabel, matching how autolinked addresses display.
• [AUTO-FIXED] (testing, 5/10) trailing-punctuation trimming truncated wikipedia-style …/Foo_(bar) bare URLs' click targets (stripped the balancing paren → 404). Replaced with balanced-paren-aware trimming; prose punctuation (,.;:!? and unbalanced )) still stays outside the link.
• [AUTO-FIXED] (testing, 7/10) long-URL wrap test added — lines.join('') must reconstruct the URL exactly (no characters lost across hard wraps).
• [AUTO-FIXED] (testing, 6/10) OSC 8 regression test added — raw render output must carry the exact click target (incl. balanced paren) as hyperlink metadata.
• [AUTO-FIXED] (testing, 5/10) scheme-less-label dedup test added ([example.com/docs](https://example.com/docs) renders once).
• [AUTO-FIXED] (maintainability, 4/10) rationale comment deduplicated between test file and component.
• [DEFERRED] (design, 6/10) links inside markdown table cells still render label-only through stripInlineMarkup (no visible URL, no OSC 8) — the same invisibility this PR fixes for paragraphs persists in the table block context. Pre-existing; fixing it changes table widths, so it needs its own decision. Sits next to the existing cell-markdown TODO in markdown.tsx.
• [ACCEPTED] (design, 4/10) very long URLs hard-wrap and can visually dominate link-dense paragraphs — the explicit tradeoff for copyability/clickability; revisit only if real transcripts prove noisy.

Adversarial review (Claude subagent; Codex CLI not installed): the load-bearing worry — terminal-escape injection through now-verbatim model-controlled URL text — is structurally impossible: the renderer strips every C0 control (incl. ESC/BEL) from text-node values at the cell-clustering pass on both inline and fullscreen paths, and ANSI-bearing assistant messages divert to the <Ansi> sanitizer before reaching the markdown renderer at all. The click opener keeps its http(s)-only allowlist + no-shell spawn. Findings disposition:

• [FIXED] trimBareUrl was O(n²) on model-controlled input: it rescanned the whole string per trimmed character, so a printed URL followed by tens of thousands of ) would freeze the TUI synchronously in the unbounded non-streaming render path (reachable via prompt-injected tool output echoed into the transcript). Rewritten as a single O(n) pass (paren balance counted once, tracked while trimming), with a 50k-paren-flood regression test that blows the test timeout pre-fix.
• [HARDENED] Unicode bidi-control codepoints (U+202A–202E, U+2066–2069) survive the C0 filter and could visually reorder a displayed domain on software-bidi terminals (Trojan-Source-style spoofing; click target stays true). Now stripped from the visible link text only — OSC 8 target and opener input untouched — with a regression test. The prior behavior (rendering remote-fetched page titles) had equal-or-greater exposure.
• [HARDENED] the display sanitizer also drops C0/DEL controls locally. The vendored renderer already strips them at the cell-clustering layer, but the reviewer's point stands: the guarantee shouldn't depend on renderer internals or on paths that serialize text outside the cell buffer.
• [PRE-EXISTING, unchanged] OSC 8 forwards whatever scheme the model wrote (javascript:, file:) into the escape's URI field; terminals ignore non-standard schemes there and the in-process opener rejects non-http(s). This diff does not change what reaches <Link>.
• [NOTED] the SSRF-guard tests were deleted along with the title-fetch machinery they guarded; if link-title fetching is ever reintroduced, its guards + regression corpus must come back with it.
• [ACCEPTED] no length bound on verbatim URLs — deliberate (the wrap-integrity test pins full-URL survival).

Verdict: merge — the only production-breaking risk it identified was accidental inclusion of unrelated concurrent edits from the shared working tree, which this PR avoids by being built in an isolated worktree from origin/main with explicit-path staging.

Design Review

Design specialist (TUI-scoped) ran as part of the specialist dispatch — see Pre-Landing Review.

Eval Results

No prompt-related files changed — evals skipped.

Plan Completion

No plan file detected.

TODOS

No TODO items completed in this PR.

Documentation

Documentation is current — no docs reference the removed title-fetch behavior.

Test plan

• ui-tui vitest: 1411 passed / 4 skipped (after rebase onto feat(memory): port openclaude's /memory — TUI picker overlay, $EDITOR spawn, backend memory controls #693); the 7 failures (createGatewayEventHandler ×2, statusRule ×2, useConfigSync ×2, virtualHeights ×1) are byte-identical on a clean origin/main checkout — pre-existing, unrelated to this diff
• pytest tests/ (the CI merge gate): 8441 passed, 7 skipped, 0 failed (8m09s)
• tsc --noEmit clean; eslint clean on all changed files
• Live PTY verification of the built TUI (URL visible on screen + OSC 8-wrapped, trailing punctuation outside the link)

🤖 Generated with Claude Code