fix(terminal): continuation fragments of wrapped references are clickable

[ghackett]

Stacked on #174. A live testing pass there found that clicks on wrapped file references did nothing — and the failing half was the continuation fragment, the one holding the file name, i.e. exactly where you'd naturally click. Verified against a live headless terminal: with PR 174 as-is, only head-fragment clicks ever worked in Claude Code's real output format.

The two causes

• Slashless continuations were never candidates. FILE_PATTERN deliberately requires a slash, and a wrap point usually falls inside the basename, so the continuation row shows o.py:7) — no VTE match under the pointer, and the click handler returned before the stitcher was consulted. Now, when nothing under the pointer matches, the raw whitespace-delimited token at the clicked cell stands in as the stitch candidate (token_at_column, unit-tested). The stitcher's geometry gates plus the existence check keep Ctrl+clicks on ordinary prose inert — verified live for row-edge prose tokens and empty areas.

• Leading junk killed the join. The emitter wraps whatever surrounds the path along with it, so fragments arrive glued to prefixes — ⏺ Read(/a/b/c is Claude Code's own tool-call format, and the row-above token it contributes is Read(/a/b/fo, not /a/b/fo. The anchored match could only shed trailing punctuation, so every up-stitch through that format failed even when the continuation did contain a slash. Joins are now matched with a search over the joined string, with span guards requiring the hit to overlap the clicked fragment itself — one mechanism that sheds junk whether it sits in a contributed token or in a token-derived candidate, and that still refuses to resolve text the click never touched.

Verification

• 5 new unit tests: up-stitch through a Read(-glued head, a slashless token candidate keeping its :7 suffix, a glued-prefix token candidate, the span guard (a junk-leading candidate must not resolve the row above's path), and token_at_column edge cases. 64 pass total.
• Live headless run (real VTE, 40 rows of scrollback, pixel scrolling, Claude-style ⏺ Read(...)/⏺ Update(...) wrapped output): all 8 click scenarios pass — both fragments of a wrapped reference resolve with the line suffix, a head token glued to Update( resolves, plain unwrapped references still resolve, and prose tokens at row edges plus empty areas stay inert. Before the fix the same probe failed 4 of 8.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AXEJjBkzLLfVtVk6HhzFEn

[ghackett]

Follow-up from live testing: wrapped clicks were still dead in the app even with this branch. Ground truth from running the real claude CLI inside a headless VTE found the root cause — the CLI's repaint-style renderer leaves VTE's internal ring a full page away from what the adjustment describes (adjustment claimed rows 0–40 while the content sat at ring rows 42–80), so every adjustment-derived get_text_range read returned an empty row and the stitcher never saw any neighbour text. Feeding text into a terminal keeps ring and adjustment aligned, which is why the earlier probes never caught it.

Fixed in d5d86bd by reading row texts from the visible-screen snapshot (get_text_format) indexed by screen row straight from the click's y — no ring coordinates involved. A click map over every cell of a real Ink-rendered wrapped path (64 columns, both the bare and Read(...):42 forms) now resolves on all four rows, while box borders and prose stay inert.

[ghackett]

Second live-testing follow-up: continuation fragments containing a slash sometimes still misfired — clicking the head half opened the parent directory. A wrap-position sweep (all 143 viable splits of a long path, plain and Read(...):42 forms, clicking every cell of both rows) isolated the shape: the wrap falling right after a character FILE_PATTERN refuses to end on (…/collins/ ⏎ linkpatterns.py, or …/collins/. ⏎ hidden/…). The candidate arrives without the shed /, the end-of-row gate never opens, and direct resolution then "succeeds" on the existing directory prefix — exactly the hazard stitch-before-direct was meant to prevent.

Fixed in 99d70f9: the gate also opens when the row ends with the candidate plus a run of exactly the shed characters (:.,;!?/), and that run seeds the downward join. Sweep now passes 143/143; 27 unit tests and the 8-scenario feed probe pass clean.

[ghackett]

@claude review

[claude]

Claude finished @ghackett's task in 4m 44s —— View job

---

Review

• Gather context (read diff, files, CLAUDE.md)
• Review collins/linkpatterns.py changes
• Review collins/terminal.py changes
• Review new tests in tests/test_linkpatterns.py
• Post review feedback

I read the full diff against origin/feat/wrapped-file-references and traced the logic by hand for every new test case (the two comment threads describe two rounds of real bugs already caught by live testing — good sign that this got exercised hard before review).

Correctness — no bugs found. I hand-traced resolve_wrapped_reference and token_at_column against all 8 new/changed test cases in tests/test_linkpatterns.py, including the trickier ones:

• test_stitch_must_reach_into_the_candidate — confirmed the span guards (collins/linkpatterns.py:199-202) correctly reject the Read(/tmp/.../collins("x case: the only regex match found by finditer ends exactly at len(up), so m.end() <= len(up) correctly skips it, and the loop falls through to up="" / down="", which is skipped by the "bare candidate already failed" guard — net None, as asserted.
• test_stitch_downward_across_directory_boundary / ..._hidden_dir_boundary — the _SHED_CHARS trailing-run logic (collins/linkpatterns.py:151-162) correctly recovers the shed / and /. and seeds the downward join.
• test_prose_hanging_punctuation_does_not_false_join — confirmed the shed-chars gate can open on ordinary trailing punctuation (not just real wraps), but the join still fails existence resolution downstream ("collins/foo.py,and" doesn't exist), so the design is safe: the gate is permissive, the existence check is the real backstop.
• test_stitch_upward_sheds_prefix_glued_to_head — confirmed the Read( prefix is shed for free: _PATH_PRE's lookbehind can't start a match inside Read(...)'s word-chars, so finditer's only viable match starts right after the ( boundary.

The terminal.py half — reading rows from get_text_format (visible-screen snapshot) instead of grid get_text_range_format, and folding the pixel-scroll fractional offset into y before the row = (y + frac) // ch division — matches what commits d5d86bd/99d70f9 describe fixing, and looks consistent with the rest of the click-handling code (the ±1 row slop, _STITCH_ROWS_UP/DOWN limits).

Minor, non-blocking nits:

• collins/terminal.py:192 and :201 — _reference_roots(terminal) is now called twice per click in on_pressed (once in the new no-match branch, once in the kind == "file" branch). Could hoist above the if not uri: check to avoid the duplicate ancestor-widget lookup. Purely cosmetic, not worth blocking on.
• collins/linkpatterns.py:150 / :171 — when trail == "" (candidate sits exactly at the row's end with no shed chars) and rows_below is empty or contributes an empty chain, downs ends up with a duplicate "" entry, so that combination gets tried twice. Harmless (same input, same result), just a trivial bit of redundant work.

Verification note: I wasn't able to execute the test suite directly in this environment (Bash tool calls require interactive approval that isn't available here), but get_ci_status shows the CI workflow for this PR already completed successfully (run 30829874417), consistent with the 64 unit tests and live headless-VTE probe described in the PR body and the two follow-up fix commits.

No changes requested — this looks solid and ready as-is.
· branch