Files
pipelines/app/notes/C2-message.md
T

170 lines
9.5 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# C2 message — one native transcript row
Manager integration follow-up: role labels keep the host's own weight rather
than forcing w600. Unknown user parts now also retain a native raw disclosure;
they are not silently dropped while assistant unknown parts remain available.
The added focused regression checks native label weight and exact unknown user
data after opening its disclosure. Full conversation integration remains separate.
The first transcript slice renders one row already produced by
`transcriptRows` (`lib/features/chat/domain/chat_transcript.dart`). It is a pure
widget with no gateway, history, network, composer, shell or chat-screen wiring.
This visit fixed the two blocking review findings and the actionable advisories:
Material disclosure bodies are no longer centred, blank known-type parts no
longer read as `Unrecognised`, closed bodies are no longer encoded, and the
Cupertino disclosure now reports its expanded state.
## Behavior
- The visible role (`User`/`Assistant`) identifies the row and a single
`User message`/`Assistant message` semantics group carries it. The visible role
label is excluded from semantics so a screen reader hears the role once. Role
text uses the host's semantic secondary colour, not a hardcoded colour or font
family.
- A user row joins its text parts with the web's newline rule and keeps the exact
string in a `SelectableText`, including leading/trailing spaces. Its optional
`client` label is shown. `file` parts show the web's
`attachmentDisplayName` fallback (name, else URI basename, else `file`), also
selectable.
- An agent turn draws every part in its original order. Text parts are selectable;
`reasoning` and `tool` parts are collapsed native disclosures. A non-text part
of an unknown type gets `Unrecognised part: <type>` (or `Unrecognised part`)
with the full pretty-printed raw JSON behind the disclosure, so no field is
silently dropped.
- Blank known parts draw nothing, as the web's `partRendering` does: empty
`text`/`reasoning`, and `file` parts with neither a name nor a URI. The
`Unrecognised part` fallback is reserved for unknown types (and for a part with
no type at all, which keeps its raw JSON).
- `textBuilder` is the seam for the forthcoming markdown renderer. It receives
each assistant text part; when absent, the default is the untouched
`SelectableText`.
- Material uses `ExpansionTile`; Cupertino uses a `CupertinoButton` with the
standard chevron. No custom-painted control, fixed component size, custom font
or palette was added.
- Material disclosures set `tilePadding: EdgeInsets.zero` and
`expandedAlignment: AlignmentDirectional.centerStart` with a stretched body:
the label and a short body both start at the row edge (the tile's default
padding indented the label and its body defaulted to centred). Cupertino
already started both at the row edge.
- `NativeDisclosure` takes a `childBuilder` and only calls it while open, so a
collapsed tool/unknown body never runs `JsonEncoder.withIndent`.
- The Cupertino disclosure wraps `CupertinoButton` in one `Semantics` node with
the label, button action and `expanded` state; `ExcludeSemantics` stops the
child button from replacing that node. The browser exposes
`aria-expanded=false` → `true`.
## Files
| Path | Lines | Role |
| --- | ---: | --- |
| `lib/features/chat/ui/transcript/chat_message_view.dart` | 195 | Row composition, part rules, attachment name |
| `lib/features/chat/ui/transcript/native_disclosure.dart` | 79 | Native Material/Cupertino disclosure |
| `test/features/chat/ui/transcript/chat_message_view_test.dart` | 321 | Ten row tests |
| `test/features/chat/ui/transcript/native_disclosure_test.dart` | 64 | Two lazy-body tests |
| `tool/chat_message_probe.dart` | 82 | Material/Cupertino dark fixture |
| `tool/chat_message_probe.mjs` | 386 | Reproducible Chrome + OCR probe |
Production total is 274 lines. No tracked file, domain file, shell file, pubspec
or `main.dart` changed. `git status --short` shows only new paths, including this
note and the probe driver.
## Checks (actual)
- `flutter test test/features/chat/ui/transcript`: 12 tests passed (10 view, 2
disclosure).
- `flutter test test/features/chat`: 45 tests passed.
- `flutter analyze`: no issues.
- `flutter build web` for both probe hosts (Material and Cupertino defines):
built successfully.
- `node tool/chat_message_probe.mjs` for both hosts: real headless Chrome,
loopback OS-assigned port, dark appearance, 390x844 at DPR 2. It asserted the
role groups, client, all three disclosures, the collapsed bodies' absence
before tapping, the expanded reasoning/tool/unknown text after tapping, and
each body's leading edge within 2 device px of its label (Material 35/33,
34/32, 34/32; Cupertino 34/32, 33/32, 34/32). Cupertino also reported
`aria-expanded=false` before and `true` after. Captures were inspected.
- The review repro `review_findings_test.dart` now passes F1 and F2 and fails only
F4, whose assertion expects the raw boolean `true`; the semantics data reports
`Tristate.isTrue`, and the candidate's own test asserts
`Tristate.isFalse` → `Tristate.isTrue`. F3 printed 112 ms before the fix and
a warm 44 ms for the 1 MiB collapsed body after it (small 53 ms), i.e. no
per-byte cost remains.
Widget-test conditions: the two user-row tests pump a dark `MaterialApp` and a
dark `CupertinoApp` at 2.0x text on 390x844 and assert no exception. The agent
order test uses Material at 2.0x on 390x844. The alignment test runs Material at
1x and 2x on 390x844 (resetting the tree between scales). The remaining tests use
the default 800x600 surface at 1x, with the Cupertino disclosure tests in a
Cupertino app. The builder-seam test uses a plain light `MaterialApp`. No test
asserts a fixed row width or height.
## Browser evidence
Captures under `/Users/alan/.local/share/pipelines/data/artifacts/70/public/c2-message/`:
- `material-top.png`, `material-reasoning-expanded.png`,
`material-tool-expanded.png`, `material-unknown-expanded.png`
- `cupertino-top.png`, `cupertino-reasoning-expanded.png`,
`cupertino-tool-expanded.png`, `cupertino-unknown-expanded.png`
- `material-evidence.json`, `cupertino-evidence.json`
These replace the pre-fix captures: the old Material expanded shots showed the
centred JSON body and are no longer current. The six earlier files with the same
names are overwritten; the two `*-tool-expanded.png` files and both evidence
files are new.
OCR is imperfect on the small Cupertino paperclip glyph (`‘eport.csv`); the image
shows `report.csv`, the probe asserts the stable `eport.csv` suffix, and exact
names are covered by the widget test. Flutter web paints `SelectableText` on the
canvas, so browser assertions use OCR for body text and the semantics DOM for
controls, as the existing Runs probe does.
## Review findings this visit
- Fixed: Material disclosure body centring (blocking). Regression asserts the
body's `dx` equals the label's at 1x/2x and for the unknown JSON body, plus a
browser leading-edge measurement per host.
- Fixed: blank `text`/`reasoning` parts (blocking). The fallback now applies only
to unknown types. Regression covers blank text, whitespace text, blank
reasoning, a nameless file part, and the still-visible unknown part.
- Addressed: eager JSON encoding of collapsed bodies (advisory). `childBuilder`
is lazy; two tests count body builds while closed and open.
- Addressed: Cupertino expanded state (advisory). One semantics node now reports
`expanded`, and a widget test plus the browser `aria-expanded` check cover it.
- Addressed: inaccurate notes (advisory) — the test conditions above are the
actual ones.
- Addressed: browser evidence not reproducible (advisory) — the CDP+OCR driver is
now `tool/chat_message_probe.mjs`, matching the existing `*_probe.dart` +
`*_probe.mjs` pairs.
- Addressed: duplicate role announcement (advisory) — the visible role label is
excluded from semantics; a test asserts the group label and the absence of a
second `Assistant` label.
- Hardened: a non-list `parts` value is ignored instead of throwing.
- Not changed (recorded for integration): the transcript still imports
`../sessions/chat_adaptive.dart`; promoting the chat-wide helper is an
integration decision outside this lane. Expanded tool/unknown bodies remain
unbounded single selectables (rich tool cards are later work). Agent `file`
parts still render as attachments (web draws only answer/reasoning/tools).
User text remains verbatim rather than web-trimmed (declared acceptance).
`transcriptRows` still emits an assistant row for an all-blank turn, unlike the
web's visible-turn rule; the transcript-list wiring must account for that.
## Not implemented in this slice
- Markdown rendering is not implemented. Assistant text is shown verbatim; the
`textBuilder` seam is the handoff for the later renderer.
- Rich tool cards are not implemented. Tool bodies are raw JSON, collapsed, as
agreed for this slice. Large outputs are not bounded yet.
- The reasoning label is always `Reasoning`; the web's streaming `Thinking…` and
`· N steps` wording, elapsed times and tool duration/status cards remain later
work.
- No transcript list, history loading, live frames, composer, approvals, forms,
attachments opening/downloading, shell registration or chat screen is included.
- Mobile was not built, run or installed; only web was exercised. Source uses
standard portable widgets.
- No commit, merge, push or deployment was performed; the new paths are the
uncommitted candidate.
Full unified diff: `/Users/alan/.local/share/pipelines/data/artifacts/70/public/c2-message/c2-message.diff`.