4.0 KiB
Next session prompt (written 2026-10-22, late)
Continue the Flutter port in /Users/alan/repos/pipelines. Read app/AGENTS.md, app/WORK.md, app/PARITY.md and this note first. Main is at the commit that added this note; 1607 tests pass and flutter analyze is clean.
Landed this session: runs-pipeline-url (main's filter menu superseded the
lane's pipeline menu; URL routing kept), chat /stop + /help (+ the web's
Stopping the running turn… line), live New marks (active chat cleared on
leave), /new /close /rename, and chat create dialog parity (fixed New chat
title, Home (~) sends ~, harness picker only with >1 backend, preloaded
instructions, model note, Start in …).
Preview: build/web published (JS 1e4e1877); rollback web-before-chatcmds-1792704916.
Superseded WIP kept for reference only: flutter/chat-paste (9baed7ea, landed via chat-paste-2), flutter/chat-create-dialog (09d17b18, redone as flutter/chat-create-parity).
Open decision: Flutter's /rename sends the shared server rename (matching its own Rename dialog); the web's /rename is device-local. The usage line still says "on this device".
Also landed: chat row refusals dismiss with OK (not yet published to the preview).
Pipeline runs (Task v6 with --model openai/gpt-6-luna, one worktree each):
155 native boot uses the saved endpoint, 156 sent attachments on the
optimistic turn; both landed. The protocol-mismatch warning was already done
(8b9a298d). Runs 143-145 and 153 are stuck cancel-requested awaiting server
restart recovery; do not restart the server. Preview JS c0a551f9, rollback
web-before-runs155-156-1792707430.
Run 157 landed: a connected foreground return re-reads the visible Pipelines/Runs list (no text-field guard or scroll restore yet).
Runs 158-160 landed: foreground refresh skips dialogs/pages/text fields; a lost
launch survives a reload (the dialog now awaits the stored intent; key uses a
blank endpoint, not per server); list scroll is kept on re-read (tests only).
Main was briefly red at 3503060d (run_start_wiring_test lacked the
SharedPreferences mock); fixed by run-intent-test-fix.
Runs 161-162 landed: pending launch keyed per server; deep-linked chats marked read.
Remaining: native OS deep-link registration and the phone install need a
reachable endpoint (operator decision); /rename server-vs-device decision;
layout work (create dialog in the shell) after functionality.
Mobile: phone needs a reachable endpoint (operator decision), then device
verification and iOS portability audit.
Design round (2026-10-23, done directly; pipelines stopped)
All pipeline runs were cancelled: the server's agent cleanup leaves runs stuck
"running" (143-145, 153, 169, 175, 184) until a restart, which is the operator's
call. Work since then is direct, one defect per branch, each with a regression
test proven to fail without the fix and sh tool/check_full.sh before landing.
Landed: chat conversation bar leads (outside the scroller); chat rows put the state word on the title line and recency first in the meta; chat toolbar keeps New chat beside search; run actions share one compact style; run tabs and visit tabs line up with the 16 inset; run detail status kicker; coloured status glyphs on run rows and pipeline rows (lib/ui/status/run_status_icon.dart); decision banner; visit pane alignment and field-card padding; Settings URL field and Notifications section; schedules page follows the connection (a real bug: the list never rebuilt after its read); search empty message; template stage counts. Preview JS c6ed91c6, rollback build/web-before-design2-1792758894.
Not defects (checked against the web): narrow pipeline graph scale (the web fits the same way, MIN_SCALE 0.3); Cupertino inset-grouped 20 vs 16 (native inset).
Tools: tool/shoot_real_app.mjs (captures any route of the real app at 390/1280 against the loopback fixture), tool/check_full.sh (strict landing gate), tool/publish_preview.sh (rename swap with rollback). Project pipelines are written in app/pipelines/ but unpublished: the server runs without -allow-authoring.