feat: add browserless deterministic SVG export to record export - #171
feat: add browserless deterministic SVG export to record export#171ThomasK33 wants to merge 21 commits into
Conversation
aube audit --audit-level high (run by CI) reports 8 high advisories in transitive dev dependencies: brace-expansion (<5.0.9), js-yaml (<4.3.1), nanoid (<3.3.18), and postcss (<=8.5.17). Bump the existing overrides block to the fixed versions; each package resolves to a single version in the tree, so the bumps are mechanical.
- src/export/svg.ts: pure grid-frame -> SVG renderer (style runs, XML escaping, fixed 8.4x18 cell metrics, SMIL discrete animation) - src/replay/gridFrames.ts: offline libghostty-vt replay capturing de-duplicated styled grid frames with recorded hold durations - record export --format svg (+ --animate, --timing rejected for svg) - schema/messages format enum, recording filename .svg extension, review-bundle .svg classification, USAGE.md docs
…ites - unit tests for renderGridFramesToSvg determinism, style runs, XML escaping, cursor block, SMIL hold key times - unit tests for captureGridFrames coalescing, de-duplication, recorded holds, final mode, and native-unavailable export error - record export command tests for svg envelope, --animate/--timing validation - native-gated integration tests incl. byte-identical repeated exports - protocol/golden/artifact-path/review-bundle enum extensions
|
@codex review |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 3e37952d6e
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
The aube repository moved from endevco/aube to jdx/aube. GitHub attestation lookups via the old path now find nothing, so mise 2026.8.x fails 'mise install --locked' with 'Lockfile requires github-attestations provenance ... but verification was not performed', breaking every CI job at setup. Regenerate the aube lockfile entries with current mise (mise lock aube): URLs now point at jdx/aube and the unverifiable provenance pins are dropped. All artifact sha256 checksums are unchanged, so the pinned binaries are identical.
Cursor-forward output (e.g. ESC[10C) surfaces untouched columns as unstyled empty cells; merging them into a matching default-style text run re-anchored the run at the gap's first column and stretched the following glyphs across the gap via textLength. Zero-style empty cells now break the current run so the next glyph anchors at its true column, while styled wide-glyph spacers keep extending their run's cellCount/textLength. Added an invariant that emitted runs cover strictly advancing, non-overlapping columns within the row.
… feat/record-export-svg
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 4ed3a7bda7
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
- Preserve wide-glyph spans across styles: SnapshotCellSchema gains an optional width (>=2, leading cell only), the libghostty-vt backend propagates the native span (spacers stay width-free), and svg run grouping consumes a wide glyph's trailing spacers regardless of styling, so default-style CJK keeps its full textLength while zero-style gap padding still breaks runs. - Embed fonts in standalone SVGs: always inline the 21 KB JetBrains Mono latin subset as a woff2 data URI; inline the 2.5 MB Symbols Nerd Font only when a rendered frame contains a Private Use Area glyph. Deterministic (base64 of checked-in assets) with an SIL OFL 1.1 attribution comment. - Replace quadratic replay in captureGridFrames with an event cursor: each boundary feeds only the not-yet-applied validated event suffix, so every event is fed exactly once.
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: c215cdd7f4
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
- Timeline spans the full event log: leading non-visual gaps surface as a pre-visual (blank) frame held until the first visual boundary, and trailing non-visual events extend the final frame's hold to the last event timestamp (plus the final viewing hold). Model documented in captureGridFrames. - Honor hidden cursors when known: SvgGridFrame gains optional cursorVisible and the renderer skips the cursor block when false. The libghostty-vt native snapshot does not expose DECTCEM visibility (verified against the module's TerminalSnapshot shape), so captured frames leave it unset and keep drawing the cursor; documented. - Embed the Symbols Nerd Font whenever a rendered code point escapes the pinned JetBrains Mono latin subset (exact cmap coverage dumped offline and hardcoded), not just PUA glyphs, so symbols like U+26A1 no longer fall back to environment fonts. Over-embedding for glyphs covered by neither face (e.g. CJK) is accepted and documented.
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: b725d76037
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
A running-but-silent session (e.g. sleep) has a valid empty event log; svg export previously failed its >=1-event invariant with EXPORT_ERROR while asciicast already supports header-only exports for this state. captureGridFrames now synthesizes a single blank frame from the manifest's initial dimensions (cursor at 0,0, final viewing hold) without booting the native backend — backends reject snapshot() before the first replayed event. capturedAtSeq follows the empty-log convention (0, matching asciicast); frameCount 1, event counts 0, rendererBackend 'none'. Output stays deterministic.
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 0edf33d440
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Glyphs outside both bundled faces (notably CJK and most emoji) render via the viewer's monospace fallback, mirroring the reference ghostty-web renderer's own system-font fallback. Documented in docs/USAGE.md and at the embedding predicate: no CJK face is bundled by the pinned profiles and glyph-path conversion would require a font-shaping dependency. Layout stays deterministic under fallback fonts because text runs are grid-pinned via textLength.
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: c4f53a7288
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
A mixed-width run (e.g. A + CJK + B) merged into one <text> scales natural advances uniformly via textLength, so a fallback font whose wide-glyph advance ratio differs shifts interior glyphs off their columns even though the run endpoint stays correct. Runs are now homogeneous in cell width: each wide glyph forms its own single-glyph run spanning its trailing spacers, while width-1 glyphs keep merging by style. Within a homogeneous-width run a monospace fallback has equal natural advances, so uniform scaling lands every glyph exactly on its column. Gap and spacer semantics unchanged.
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d30294e6f7
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Still and animated SVG exports at the same seq both resolved to recording-<seq>-svg.svg: the second export overwrote the first file while the artifact manifest kept both entries, leaving one entry's sha256/bytes/animated metadata pointing at the other export's content. recordingFilename now accepts an optional variant label appended before the extension; record export passes 'animated' for animated SVG so it resolves to recording-<seq>-svg-animated.svg while still SVG and the existing asciicast/webm names stay byte-identical. The default-path seq invariant is unchanged.
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: c624435dda
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
With 6-decimal rounding, a 1 ms hold inside a timeline longer than ~16m40s serialized both frame boundaries to the same keyTime, giving the frame a zero-length interval and silently dropping a recorded transition. keyTimes now serialize with 12 decimals (millisecond resolution up to 10^12 ms timelines) with trailing zeros trimmed for canonical output, and renderVisibilityAnimation asserts serialized boundaries stay strictly increasing for positive holds. Also documents in gridFrames.ts that captured frames omit cursorVisible because the native snapshot exposes no DECTCEM field.
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 1ab6654e80
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
A background-styled empty width-1 cell between same-style glyphs (e.g. colored field padding) matched the run's style and merged, stretching the run across its column so the following glyph drifted off position. Every empty cell that is not a wide-glyph trailing spacer now breaks the current text run so the next glyph re-anchors at its true column. Backgrounds are grouped independently of text runs via groupRowIntoBackgroundRuns: styled empties still shade their columns, and adjacent same-bg cells merge into one continuous rect regardless of other styling differences.
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 2727b55720
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Same-seq SVG exports with different --profile values resolved to the same default path: the second export overwrote the first file while the artifact manifest appended a second entry with stale sha256/bytes/renderProfileHash pointing at replaced content. The SVG filename variant now carries the resolved profile name (sanitized by recordingFilename), so defaults become recording-<seq>-svg-<profile>.svg and recording-<seq>-svg-<profile>-animated.svg. asciicast/webm naming is unchanged; webm's pre-existing profile collision is out of scope.
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: b55a1c5486
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
A width-1 symbols-face glyph (e.g. U+E0A0) beside same-style latin text merged into one run; the two embedded faces have different natural advances, so uniform textLength scaling drifted interior glyphs off their columns. Each glyph is classified against the pinned latin subset coverage (latin vs symbols/fallback path) and runs now break when the classification changes, exactly like the width-boundary split. Glyphs covered by neither bundled face (CJK etc.) classify as fallback too — both resolve past the latin face, sharing advance behavior. Every run is now homogeneous in (style, width, face), so uniform scaling anchors each glyph to its column under monospace faces.
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 7e84625308
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Math.max(...frames.map(...)) and lines.push(...renderFrameElements()) argument-spread per-frame arrays; past V8's argument limit (~125k elements, reachable by long recordings or dense frames) they throw a RangeError. Canvas dimensions now come from a loop and frame elements are appended one-by-one. Array-literal spreads elsewhere are iteration-based and unaffected.
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: e5a7863859
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Animated capture keeps every distinct frame's dense styled grid in memory until rendering; within the supported 50 MB event-log limit a recording can produce enough distinct frames to exhaust the heap. captureGridFrames now enforces a retained-cell budget (default MAX_ANIMATED_GRID_CELLS = 20M cells, ~10k distinct 80x24 frames, adapting to terminal size; override-able via the maxGridCells option for tests). Exceeding it fails fast with EXPORT_ERROR suggesting a still SVG or WebM export — frames are never silently dropped because exported artifacts are evidence. Documented in docs/USAGE.md.
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f5e8e3c854
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
For an animated recording that resizes from a larger grid down, renderGridFramesToSvg sizes the canvas to the per-frame maxima of cols/rows, but captureGridFrames reported only the final frame's dimensions — record export then published width/height in CLI JSON and the artifact manifest smaller than the actual viewBox. capture.cols/rows now carry the same per-frame maxima the renderer sizes from (documented on GridFrameCapture). Still exports capture a single frame, so their dimensions are unchanged; the blank-synthesized empty-log capture is likewise unaffected.
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 29ac24b21c
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
The two-bucket face classification merged adjacent width-1 glyphs where one resolves to the embedded Symbols Nerd Font and the other to the viewer's system fallback (e.g. alpha + U+26A1): different real fonts with different natural advances under one aggregate textLength drifted interior glyphs. GlyphFace now has three buckets — latin, symbols, system — with symbols classified by the pinned SymbolsNerdFontMono-Regular.ttf cmap coverage (dumped offline from the checked-in asset's format-12 cmap: 10,410 code points in 28 ranges, hardcoded with regeneration notes). Run grouping already breaks on face change. The embedding predicate is narrowed to glyphs the symbols face actually covers: system-fallback glyphs (CJK etc.) gain nothing from the 2.5 MB payload and no longer trigger it.
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: e9cce6c948
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Adjacent system-fallback glyphs (covered by neither bundled face) can resolve to different system fonts per script (e.g. Greek + Hebrew), so even a same-bucket merged run scales unequal natural advances under one aggregate textLength and drifts interior glyphs. System-classified glyphs never merge and close their run immediately (wide system glyphs close via the existing spacer branch), so every system glyph is pinned individually via its own x + textLength and cannot drift regardless of which face the viewer picks. Latin and symbols runs keep merging: each is a single known monospace face with uniform advances. Guarantee documented: runs are homogeneous in (style, width, face) and system-fallback glyphs are always singleton runs.
|
@codex review |
|
Codex Review: Didn't find any major issues. What shall we delve into next? Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
1 participant
Problem
agent-tty's only visual artifacts are Playwright-produced PNGs and WebM videos: every visual export needs Chromium, and WebM additionally doesn't render inline on GitHub. There is no visual artifact that is cheap to produce in CI, diffable, or hash-stable.
Mechanism
record export --format svg [--animate]— the first renderer-grade artifact with zero Playwright/ffmpeg footprint:src/replay/gridFrames.ts) builds validated replay input from the session's event log and replays it through the libghostty-vt backend (in-process WASM, no browser), capturingsnapshot({ includeCells: true })at output/resize boundaries. Same-timestamp events are coalesced and consecutive identical frames are de-duplicated with merged hold durations. Still export captures only the final frame; if the optional@coder/libghostty-vt-nodepackage is unavailable the command fails with a clear error rather than falling back to a browser.src/export/svg.ts) groups consecutive same-style cells into<text>runs pinned to the grid withtextLength, draws per-cell backgrounds/underlines and a translucent block cursor, and uses fixed cell metrics (8.4×18 px at the pinned fontSize 14) so output is deterministic — byte-identical across repeated exports — with no timestamps, random IDs, or environment-dependent content. Animation uses ID-free SMIL discrete visibility with key times derived purely from recorded event-log timing (--timingis rejected for svg). Only the two locked profiles (reference-dark/reference-light) are supported; no new themes.RecordExportResultSchema,.svgviarecordingFilename, artifact kindrecordingwith metadata (frameCount,animated,profileName,renderProfileHash,rendererBackend, event counts), and.svgclassification in the review-bundle tooling. Docs updated indocs/USAGE.md.Idea credit: adopted from the tcut review (
docs/notes/tcut-review.md, §3.3) — reimplemented for agent-tty's replay/profile model, no code copied, no new dependencies.How to verify
Dogfood proof (isolated
AGENT_TTY_HOME): repeated still exports byte-identical (29bfb9253a6b…, 704 bytes; a larger colorful session:dc369920f3f5…, 11,226 bytes twice); animated export of an 11-frame session reportsdurationMs: 4601,rendererBackend: libghostty-vt. The rendered SVG shows correct bold/ANSI-color runs and the cursor block.Validation run:
npm run typecheck,npm run lint,npm run format, unit suites (svg,gridFrames,record-export, protocol/golden/review-bundle/artifact-storage), andtest/integration/record-export.test.tswith the native package available (all svg cases executed). New tests fail without the change; native-dependent integration cases skip cleanly when the optional package is missing.Note: the CI
auditstep currently fails on every fresh run due to new upstream advisories; the fix is in the separate overrides-bump chore PR (#168).Generated with
mux• Model:anthropic:claude-fable-5• Thinking:xhigh