Merge branch 'main' into flutter/chat-messages

This commit is contained in:
2026-10-21 20:44:02 +01:00
41 changed files with 3618 additions and 222 deletions
+44
View File
@@ -0,0 +1,44 @@
# C2 — Chat thread controller
Manager integration follow-up: exposes subscribed/subscriptionError separately
from history errors, so readable stored history cannot hide a refused live watch.
Resync marks history loading while retaining the current transcript. A focused
regression covers refused subscribe + successful history, then successful resync.
The controller does not claim to fold form events; those are a separate UI slice.
`lib/features/chat/conversation/chat_thread_controller.dart` owns one open
conversation without any UI. It is constructed with the shared `ApiGateway` and
one `sessionId`, reuses `ChatRepository` for the wire and
`domain/chat_transcript.dart` for the fold, and exposes `messages`, `running`,
`loaded`, `loading`, `hasMore`, `cursor` and `error` through
`ChangeNotifier`.
Behaviour:
- `load()` calls `chat_subscribe` with the exact session body, then reads the
newest `chat_messages` page with `limit: 200` and an empty cursor. A refused
subscribe leaves history usable.
- `loadOlder()` is explicit only, requires `hasMore`, passes the server's opaque
`next_cursor` back verbatim, and prepends the older page by id.
- Only `chat_event` envelopes for this session are folded; replayed/live
text, reasoning, tool, step and user frames use the existing domain fold, so
unknown message and part fields survive.
- `resync()` settles the held turn, re-subscribes, then replaces the transcript,
cursor and `hasMore` with the newest authoritative page. Activity replayed
while re-attaching (for example a running step) survives the replacement. A
pushed `resync` event invokes the same path.
- History failures set `error` and keep the held transcript; they are never
reported as an empty success. Superseded reads and replies after disposal are
ignored.
- Disposal cancels only the local stream subscription. There is no
`chat_unsubscribe` request.
Checks (run from `app/`):
- `flutter test test/features/chat/conversation` — 10 tests passed.
- `flutter test test/features/chat` — 43 tests passed.
- `flutter analyze lib/features/chat test/features/chat` — no issues.
Not done: the controller is not yet wired to a screen, so there is no rendered
or browser change to capture; mobile builds and device verification were not
performed. No shared domain, shell, main, pubspec or server files changed.
+174
View File
@@ -0,0 +1,174 @@
# P1 wire — Pipelines destination on the shared session
## Implemented
- `lib/application/pipelines_destination.dart` registers the landed
`PipelinesScreen` on the existing `/pipelines` route through `AppSession`.
It reuses the session's single socket and gateway; no second connection,
fake rows or hidden request path was added.
- Before the first accepted connection the destination mounts no screen and
sends no `list_pipelines`, `list_runs` or `get_pipeline`. The full gate says
`Connecting to Pipelines…` or `Pipelines server unavailable`, with an
explicit native `Retry` and the existing missing-endpoint explanation.
- After the first connection the screen stays mounted through drops. A
status banner reports `Reconnecting to Pipelines…`; a reconnect advances
`syncToken`, and `PipelinesScreen.didUpdateWidget` calls the existing
`controller.load()` so the list and the selected detail refresh in place.
Selection (pipeline id and requested version) is preserved; there is no
remount and no return to the empty list.
- `PipelinesScreen` gained `embedded` and `syncToken`. Embedded mode skips its
own Material `Scaffold`/`AppBar` or Cupertino `CupertinoPageScaffold`/
`CupertinoNavigationBar`, because the shell already owns the single bar.
A native refresh action stays in the content (`IconButton`/`CupertinoButton`
with the same `Refresh` wording), using the real host selection.
- The connection gate and banner moved verbatim from `runs_destination.dart`
to `lib/application/connection_status.dart`, so Runs and Pipelines share one
implementation and one wording. Runs behavior is unchanged.
- Registration is a tiny entry in the existing `destinationBodies` map in
`lib/application/pipelines_app.dart`.
- No graph rendering, actions, templates, schedules, route expansion, native
control replacement, fixed component size/font/colour override, server,
web, legacy, Android or pubspec change was made. Graph still shows
`Graph rendering is pending (P2).`
## Files
Production:
- `lib/application/connection_status.dart` (new, moved shared chrome)
- `lib/application/pipelines_destination.dart` (new)
- `lib/application/pipelines_app.dart` (registration and import)
- `lib/application/runs_destination.dart` (uses the shared chrome)
- `lib/features/pipelines/screen.dart` (embedded framing, content refresh,
reconnect sync token)
Tests:
- `test/application/pipelines_destination_test.dart` (new)
- `test/application/pipelines_registration_test.dart` (new)
- `test/application/pipelines_replies.dart` (new serialized reply helper)
- `test/application/pipelines_app_test.dart` (route now renders the real list)
- `test/application/settings_registration_test.dart` (answers the mounted
Pipelines reads before visiting Settings)
- `test/features/pipelines/pipelines_screen_test.dart` (embedded framing and
content refresh)
Browser tooling (isolated, pure loopback fixtures):
- `tool/pipelines_wiring_probe_entry.dart` — exact `runPipelinesApp()`
composition
- `tool/pipelines_wiring_probe_server.mjs` — static bundle plus loopback
websocket fixture with drop/refuse/accept/state endpoints
- `tool/pipelines_wiring_probe.mjs` — real Chrome DevTools probe
## Checks actually run
| Check | Result |
| --- | --- |
| `flutter analyze` | `No issues found!` |
| `flutter test` | **595 passed** |
| `flutter test test/application` | **33 passed** |
| `flutter test test/features/pipelines` | **49 passed** |
| `flutter test test/features/pipelines/pipelines_screen_test.dart` | **6 passed** |
| `dart format --output=none --set-exit-if-changed lib test tool/pipelines_wiring_probe_entry.dart` | 273 files, 0 changed |
| `node --check` on both probe scripts | passed |
| `flutter build web` (default `lib/main.dart`) | `✓ Built build/web` |
| `flutter build web -t tool/pipelines_wiring_probe_entry.dart` | `✓ Built` |
| `node tool/pipelines_wiring_probe.mjs <build> <captures>` | exit 0, 25 checks, 9 captures, `probe.json` has no error |
Coverage added:
- missing native endpoint never opens a socket;
- connecting gate sends no list request; the screen mounts only after the
shared socket accepts;
- refused first connection is friendly, names no transport exception and
offers Retry;
- a drop retains the loaded rows and detail, then reconnect re-reads list and
the open detail without losing the selected row;
- the registered `/pipelines` route renders real rows, detail content hash and
source with exactly one Material or Cupertino bar, one socket;
- Cupertino host builds no `Material` widget;
- embedded framing adds no second bar and keeps a working content Refresh.
## Browser evidence (real Chrome, normal composition)
Directory: `/Users/alan/.local/share/pipelines/data/artifacts/69/public/p1-wire-captures`
- `01-gate-refused-material.png` — first connection refused: honest gate,
`Retry`, shell nav present, zero list requests recorded by the fixture.
- `02-list-material.png` — Material shell below its single bar with the real
`Release verification` / `Nightly documentation` rows.
- `03-detail-material.png` — selected pipeline Summary with
`sha256:fixture-3-v3`, versions and metadata.
- `04-source-material.png` — Source tab with the real fixture source text.
- `05-drop-detail-retained-material.png` — socket dropped: `Reconnecting to
Pipelines…` banner while the open detail stays visible at `/pipelines`.
- `06-reconnect-detail-refreshed-material.png` — after Retry: banner gone,
list and open detail re-read (fixture counters advanced for `list_pipelines`
and `get_pipeline`), same selection.
- `07-list-cupertino-dark.png` — macOS/dark host: Cupertino global bar and tab
bar, inset grouped rows, `Release verification` present.
- `08-detail-cupertino-dark.png` — Cupertino dark detail with content hash,
versions and native segmented tabs.
- `09-source-cupertino-dark.png` — Cupertino dark Source with source text.
- `probe.json` — fixture URL, 25 checks, captures and final counters.
The probe starts the fixture with websocket upgrades refused, asserts the gate
and that `listPipelines`, `listRuns` and `getPipeline` are all 0, then accepts
and drives the normal app. Hosts are selected through the browser platform
user agent (Linux for Material, macOS for Cupertino), which is what the normal
composition reads. OCR asserts the real row, content hash and source; the
captures were inspected to confirm one bar and Cupertino dark styling.
Reproduce:
```sh
flutter build web -t tool/pipelines_wiring_probe_entry.dart -o <build>
node tool/pipelines_wiring_probe.mjs <build> <capture-dir>
```
## Explicitly not done / unverified
- Graph rendering and graph nodes/edges remain P2; the existing pending notice
is unchanged.
- No action, template, schedule, route expansion or navigation redesign.
- No production server, real operator chat/run or authenticated preview was
touched; all browser data comes from the isolated loopback fixture.
- Android and iOS were not built, installed or run; mobile remains unverified.
- No push, deployment or preview-service change was made. The Land stage
committed the reviewed candidate on `flutter/p1-wire` and merged `main`
into it (see Landing below).
- No Material/Cupertino fixed sizes, palette or font overrides were added.
The only fixed breakpoint remains the existing `kPipelinesWideBreakpoint`.
## Diff
Tracked changes (`git diff --numstat`):
```
3 0 app/lib/application/pipelines_app.dart
4 184 app/lib/application/runs_destination.dart
57 16 app/lib/features/pipelines/screen.dart
26 5 app/test/application/pipelines_app_test.dart
3 0 app/test/application/settings_registration_test.dart
26 0 app/test/features/pipelines/pipelines_screen_test.dart
```
New files: `connection_status.dart` (195), `pipelines_destination.dart` (61),
three test files and three probe files. Net production change is ~120 added
lines after moving the shared chrome out of `runs_destination.dart`.
## Landing
- Committed the reviewed work as `8476cfd7 feat(app): wire the native
pipelines browser on the shared session` (15 paths, no unrelated files).
- Merged local `main` (`7be4456f`) into `flutter/p1-wire` with no conflicts;
the merge brought in the native run rename work only (`app/WORK.md`,
`lib/features/runs/**`, `notes/R2-title-wire.md`, its tests and probe).
- Post-merge checks on the merged candidate: `flutter analyze` clean,
`flutter test` **599 passed**, `dart format` 274 files 0 changed,
`node --check` on both probe scripts, `flutter build web` built, and the
Chrome wiring probe again reported **25 checks, 9 captures**, no error,
final `{subscribed:3, listPipelines:3, listRuns:3, getPipeline:3,
dropped:1, connections:1}`. The 595 in the table above is the pre-merge
branch count; the extra 4 tests are the merged rename coverage.
+23
View File
@@ -0,0 +1,23 @@
# Cancel from the actual run detail
The loaded detail's native action row reuses CancelRunAction alongside Rename.
No new confirmation or protocol implementation: Keep sends nothing, only Confirm
sends the exact run_id, and no canceled state appears before server acceptance.
The accepted run detail now folds lifecycle fields into the held list row as well
as the detail (the previous title-only update was insufficient for Cancel).
Detail-only payloads are not duplicated into the summary list; pending old detail
refreshes are invalidated after acceptance. Foreign run replies cannot replace
the open detail. Native controls still own sizing and appearance.
## Actual checks
- New normal-screen Material/Cupertino tests failed before registration, then
passed: Keep/no request, Confirm exact id, delayed accepted canceled state,
action disappears, Back shows canceled list row.
- 601 integrated tests pass; analyzer and normal web build pass.
- Normal-app real Chrome fixture probe passes23 captures, including Cupertino
confirm/Keep, Material confirm and accepted detail/list. The fixture records
exactly one cancel_run for its own synthetic run4. No real run was canceled.
- Fresh Cupertino-confirm and Material-accepted PNGs inspected in approved
temp/cancel-wired-browser. Existing history/reconnect/Settings checks pass too.
- Mobile and authenticated production mutations remain unverified.
+19
View File
@@ -0,0 +1,19 @@
# Recorded human decisions in run detail
RunDecisionNotice reuses pendingDecision and HumanDecisionAction. A standard
Material card or Cupertino section names the recorded checkpoint and offers
Decide. The existing native prompt owns option/feedback/confirmation/refusal
semantics; export/resource waits do not become invented human checkpoints.
Accepted server detail flows through the same lifecycle/list update as Cancel.
New actual-screen tests failed before wiring and now pass on both hosts: recorded
instructions, selection sends nothing, exact run/stage/option and trimmed feedback
only on confirmation, accepted result removes the card. Integrated614 tests pass;
analysis and normal web build are clean.
Real normal-app browser fixture passes28 captures, including actual Cupertino
waiting card and prompt, Material card/prompt/accepted result, Cancel/Keep and
the existing history/reconnect/Settings checks. The fixture records one exact
decide_run for its synthetic run5; opening/Cancel send nothing. Card and prompt
PNGs inspected in approved temp/decision-wired-browser. No production decision
was answered; mobile and live authenticated mutations remain unverified.
+140
View File
@@ -0,0 +1,140 @@
# S1-draft — pure saved-schedule editor draft and validation
Manager integration guard: invalid pipeline IDs and edited records without a
positive exact server identity/revision cannot build save payloads. This prevents
a malformed/historical edit from becoming a create or bypassing CAS. Three new
regressions failed against the original draft and pass after the fail-closed
guard. The original snapshot remains unchanged; no production save was tested.
## What changed
New files only; no existing feature, shell, transport, `main` or pubspec source
edited:
| File | Lines | Responsibility |
| --- | --- | --- |
| `lib/features/schedules/domain/editor/schedule_draft.dart` | 210 | mutable editor draft over `SavedSchedule`/`ScheduleRule`, web-mirrored validation, `save_schedule` record assembly |
| `test/features/schedules/domain/editor/draft_fixtures.dart` | 42 | shared server-record fixture and deep-copy helper |
| `test/features/schedules/domain/editor/schedule_draft_test.dart` | 134 | create/edit record assembly, CAS, unknown fields, workspace, source immutability |
| `test/features/schedules/domain/editor/schedule_draft_validation_test.dart` | 130 | timer/day/timezone/core field validation and invalid-field behavior |
The production file is one cohesive file (below the 220-line slice budget) and
keeps the existing domain pattern: `SavedSchedule` and `ScheduleRule` are
read-only views, `ScheduleDraft` is the separate editable state. Tests are split
by responsibility (record assembly vs. validation) so neither file exceeds the
working agreement's ~250-line aim.
## Review fix — workspace is now written into the record
Review visit 1 found that `build()` validated `workspace` but never wrote it, so
creates were refused by `save_schedule` and workspace edits were dropped. The
fix writes `record['workspace'] = workspace;` beside `name`/`version`/`enabled`,
untrimmed to match `workspace: draft.workspace` in `web/src/schedules.js`
(validation also runs on the untrimmed value). Direct probe against the real
classes:
```
create keys: enabled, id, input, name, notify_mode, pipeline_id, rule, version, workspace
create: {"id":0,...,"workspace":"/srv","rule":{"every":"24h","timezone":"UTC"},"notify_mode":"auto"}
edit workspace: /new (source /old)
edit revision: 13
```
New assertions cover the create record, an unchanged edit, a changed edit and
the untrimmed value.
## Behavior copied from `web/src/schedules.js`
`ScheduleDraft.build()` returns `ScheduleSaveResult(null, errors)` while any
field is invalid, so an invalid draft cannot yield a save payload. Errors are a
field-name → message map using the web editor's exact words:
- name `Enter a schedule name.` (stored trimmed, as the web record is)
- prompt `Enter a prompt.` (stored untrimmed)
- pinned version: digits only, safe integer, at least 1 — `Choose a positive
pinned version, not latest (0).`
- workspace must start with `/` and contain no NUL/CR/LF — `Enter an absolute
workspace path on the server.`
- interval is Go-style compound duration (`24h`, `1h30m`, `.5h`, `1.5m`) whose
matched text covers the whole value and totals at least one minute —
`Interval must be at least 1m (for example 24h or 1h30m).`
- clock times are comma-separated, trimmed, distinct `HH:mm` in 00:00–23:59 —
`Enter distinct 24-hour times, such as 09:00, 17:30.`
- weekday keys must be distinct values from `scheduleDays` —
`Choose valid, distinct weekdays.`
- notification mode is `auto`, `on` or `off` — `Choose a notification mode.`
- timezone is blank, `Local`, a known single-zone alias or a well-formed
`Area/Location` — `Enter a valid IANA timezone.`
Rule assembly mirrors the web editor: the prior rule is copied, `every`,
`times`, `days` and `timezone` are removed, then only the selected timer shape
is written. Empty weekdays mean every day (the key stays absent). A weekday set
that is unchanged keeps the server's original order (the web editor does this
so editing only a prompt does not reorder days). A blank timezone is omitted.
## Losslessness and CAS
- Edits start from `SavedSchedule.toJson()` (a deep copy): unknown top-level,
`input`, `rule` and run-option fields (`model`, `stage_models`, `write_paths`,
`allow_tools`, `title`, `output_format`, `next_due`, timestamps, future
fields) survive untouched.
- `input` keeps every unknown key and replaces only `prompt`.
- **Every build starts from a fresh copy**, so the source snapshot and the
draft are not mutated or aliased by a returned payload.
- The edit record keeps the exact server `revision`; it is never incremented,
invented or defaulted. An edit never rewrites `id`.
- A create has no source: it emits the web sentinel `id: 0`, no `revision`, no
unknown/run-option leftovers, `pipeline_id` from the caller, `enabled`,
`input: {prompt}`, `rule`, and `notify_mode`.
- `notify_mode` and `enabled` default the same way as the web draft when absent
from a malformed older record, but an existing non-default string/boolean is
preserved and validated rather than silently replaced.
## Checks actually run
```
flutter pub get
dart format lib/features/schedules/domain/editor test/features/schedules/domain/editor
flutter analyze # No issues found! (1.1s)
flutter test test/features/schedules/domain/editor # 9 passed
flutter test test/features/schedules # 56 passed
```
Focused coverage maps to the requested acceptance:
- create: sentinel `id: 0`, no revision, pipeline/version/prompt/workspace/rule;
- lossless edit: id 9/revision 13, unknown record/input/rule/run options,
workspace from the source, original map and `SavedSchedule` snapshot
unchanged, returned payload not aliased;
- changed workspace is written untrimmed, matching the web serializer;
- prompt-only edit keeps the server weekday order and clock rule;
- clock ↔ interval switching removes the other timer key and keeps unknown
rule fields;
- invalid timers (`59s`, `NaN`, `-1h`, `1.1x`, `Infinityh`, mixed text) and
valid parity cases (`.5h`, `1h30m`, `24h`);
- invalid clock times, duplicate weekdays, bad notification mode, invalid
timezones, bad name/prompt/version/workspace;
- invalid draft mutates nothing and has no record.
## Honest limits
- **Scope is the core editor slice.** Name, prompt, pinned version, workspace,
timer mode/every/times, weekdays, timezone, enabled and notification mode are
editable. Advanced JSON run options (`otherInput`, `stage models`, write
paths, model, title, output format) and tools are **preserved exactly but not
editable in this slice**; no control or parser for them was invented here.
- **Timezone is structural.** Dart ships no IANA database, so the check accepts
blank/`Local`, known single-zone aliases and a known area plus a well-formed
location (e.g. `Europe/London`, `Etc/GMT+5`). It rejects clearly malformed
values such as `Nowhere/Invalid` and `Europe/`, but cannot prove that every
location in an accepted area exists. The server remains the final validator.
This is documented at the function.
- **No UI, network, CRUD, run dispatch, polling or scheduling engine.** The
draft only assembles a record; the existing repository owns `save_schedule`,
and no server was contacted.
- **No browser capture applies.** This is a nonvisual pure domain item; there
is no rendered change, so no fixture or screenshot was produced.
- **Mobile unverified.** Only web-targeted Flutter tests and analysis ran. No
Android/iOS build, emulator, ADB or device was used; the code is pure Dart.
- **Not committed.** This stage forbids commit, push and deploy, so the
candidate remains in the `flutter/schedule-draft` worktree for the manager.