Repository navigation
Releases
PR #29248 model reporting evidence
Evidence for pingdotgg#29248. These assets document the approved rework and patch-equivalent rebase. The release tag anchors rebased head 5695f47; finalization-public-proof.json identifies the tested, reviewed and rebased commits.
The three screenshots show isolated projection fixtures in the web client after reloads, not live Claude recovery. Runtime and regression coverage and untested surfaces are recorded in the JSON.
PR evidence: Unicode command-output overwrites
Verification artifacts for the focused Unicode cursor-overwrite correction in shared command-output folding. Replayed synthetic provider output; isolated development homes; no live user data.
The scenario uses the authenticated project API for setup and the existing free Replay provider. Copy unicode-output.e2e.ts into scripts/scenarios/ and place t3-unicode-replay.ndjson at /tmp/t3-unicode-replay.ndjson before running scenarios:prove. The setup API avoids the project CLI's one-second live-server probe on a busy machine.
Implementation written by GPT-6.1 Sol in Codex; independently reviewed by Claude Opus 5.5.
PR evidence: bounded unclosed TeX scans
Verification artifacts for the focused unclosed TeX tokenizer fix. Composer recording uses isolated development state and contains no pairing credentials. Operation counts compare upstream main d2c2595 with the fix; elapsed times are observational only.
Thread Find large-row fix evidence
Before/after captures and synthetic DOM-fixture proof for the focused thread Find argument-limit fix. Evidence only; no application release.
PR evidence: terminal link delimiter trimming
Terminal link delimiter trimming proof
The committed regression tests count full-string split scan work, with no wall-clock thresholds. Six URL/absolute-path cases fail on base and pass on head. All 76 parser tests pass on head.
The temporary scenario runs the real app with isolated state and a free replay-provider fixture turn. It waits for the mounted Ghostty surface, injects a wrapped URL followed by 128 unmatched closing parentheses, freezes further PTY writes to retain the diagnostic image, and dispatches a pointer-move event. It asserts the balanced URL, pointer cursor and a linear scan budget through the real terminal hover consumer. It is diagnostic evidence, not an end-to-end PTY timing or frame-rate benchmark. Base and head are expected to look identical: the defect is excess scan work.
The separate browser microbenchmark invoked the actual Vite-served extractTerminalLinks. At 5,000/10,000 extra closing parentheses its base median was 313.6/1219.3 ms versus 0.3/0.3 ms after (three samples each). Viewport limits constrain real hover inputs. These timings do not describe whole-app latency.
vp run check:changed upstream/main --concurrency 2 passed format, lint, both Knip phases, affected web/scripts typechecks, focused tests and the selected React compiler check. Packaged desktop and Windows were not run. Mobile uses a different terminal path.
Review-loop: one round, GPT-6.1 Sol on xhigh, APPROVE, no findings. Reviewer usage was not exposed.
The paired scenarios:prove run reproduced the expected scan-budget assertion failure on base (45,408 > 1,860). Its head setup exited 78 on the CLI ownership lock before running the test. Rerunning the identical scenario on the same head with --target web passed. proof-summary.json records this combined-run provenance; this is not a claim that one paired command exited successfully.
Capture limitation: after-terminal-hover.png caught a client resync after the synchronous diagnostic assertion, so it does not show the injected text. Do not use that image as visual proof. manual-client-after.png is a separate Browser-panel canvas capture on the fixed head, with shorter suffixes; it directly shows the balanced URL underline. The scenario assertion and scan transcript establish the long-suffix behavior.
PR evidence: TAR hard-link previews
PR evidence: bounded spreadsheet merge selection
Fixture: merges.xlsx, 15000 non-overlapping two-row merges in alternating A/B columns:
A1:A2, B2:B3, A3:A4, ..., B15000:B15001. Markers in A1 and A15001.
Select A15001:B15001. Its transitive merged selection must be A1:B15001.
In the workbook preview, Ctrl+Home, Ctrl+End, Left, Down lands in A15001; Shift+Right
extends to B15001. Copy must return 15001 TSV rows including both markers.
Helper reproduction (not an application benchmark), Node v26.12.1 on a loaded shared machine:
base: 15000 merges throws RangeError Maximum call stack size exceeded.
fix: same input returns {top:0,bottom:15000,left:0,right:1}.
Timing varies with machine load; no end-to-end latency claim is made.
Unit proof: focused coverMerges regression fails on base and passes on fix.
Additional scratch check: 10000 deterministic small-range comparisons match
original fixed-point expansion. No input mutation and no overlapping input
merges are needed to reproduce the defect.
Independent Browser panel observation on base 1530bec:
- Real workbook worker layout: 15000 merges; first A1:A2, last B15000:B15001.
- Native grid keydown path: Ctrl+Home, Ctrl+End, Left, Down, Shift+Right, Ctrl+C.
- Browser error: Uncaught RangeError: Maximum call stack size exceeded.
- Clipboard callback never received data.
- Captured manual-before.png after the failed paint/copy.
The preview is mounted in an isolated HTML harness using unmodified OfficeFilePreview,
the real archive preparation worker, and the real spreadsheet/WASM worker. The harness
only supplies the workbook, observes layout/errors, and captures the clipboard write.
Full chat app navigation was not used for this proof.
Final integrated result:
scenarios:prove failsOnBase=true (1530bec),
passesOnHead=true (8ccdf9467c27045ce36eae50c407dd9ee35c8d4d).
Independent Browser panel pass on head: 15000 merges, 15001 copied rows,
first marker Chain start, last marker Chain end, errors=[].
Before/after MP4 files are clips of separate scenario recordings, not a timing comparison.
The workbook also contains an unmerged B1 value to establish two used columns for keyboard navigation.
Reproduce in an isolated checkout:
- Put merge-proof.html and merges.xlsx in /tmp/t3-merge-proof/.
- Apply fixture-setup.patch and put spreadsheet-merge-chain.e2e.ts in scripts/scenarios/.
- Run vp run scenarios:prove scripts/scenarios/spreadsheet-merge-chain.e2e.ts --base 1530bec --head 8ccdf9467c27045ce36eae50c407dd9ee35c8d4d
- Revert the fixture patch afterwards. None of these proof fixtures is in the PR diff.
For a manual check, copy the HTML to apps/web/merge-proof.html and the workbook to
apps/web/public/merge-proof.xlsx, start isolated dev, and open /merge-proof.html.
Verified rebased head when evidence was recorded: 19f5288
apps/web/src/components/files/office/sheet-cells.ts: unchanged blob 458e3ba475e311f32f0315ef8fed5c6c460b2b8e
apps/web/src/components/files/office/sheet-cells.test.ts: unchanged blob 5ed46a3a7b65e80d0821a663f42fd543d1ccb02a
PR evidence: visibility-aware audio and server loading indicators
Loading indicator verification
Chromium 148.0.7778.0; standalone component fixture built from AudioPreview,
ThreadServersPanel, Spinner and index.css. Pending media metadata is held by a
network route; the environment, permission and stop-RPC hooks are stubbed to
keep the server-stop pending. Real component markup, controls and tooltip are
used. cn uses the same clsx/tailwind-merge libraries for this fixture.
No T3 dev server or live user data is used.
The recording compares upstream before (left) and patch after (right). It shows
normal motion, scroll clipping, reduced motion, a synthetic hidden-document
visibilitychange signal, then return to normal. Synthetic hidden-page proof
checks the application's document-visibility handling; it does not claim a
real background-tab GPU/CPU measurement.
Animation Web API currentTime is sampled twice, 700 ms apart. The JSON results
record both snapshots. Both original spinners advance in every case, including
clipped and reduced motion. After the patch, clipped animations are paused,
reduced-motion has animation-name: none, and the hidden-document signal pauses
both. Returning to normal motion resumes both. This proves animation scheduling
behavior only, not any CPU/GPU/battery saving.
The full application's integrated client pass is owned separately by the
primary agent and is not claimed by these component captures.
PR evidence: integrated performance fixes
Synthetic isolated-state primary runtime evidence for Markdown rendering and loading indicators; proof.txt states method and limits.
PR evidence: deferred automations panel
Synthetic isolated-state runtime proof for deferred thread automations loading. Screenshots, control-flow recording, and limits are documented in runtime-proof.txt.