Files
pipelines/app/WORK.md
T

19 KiB
Raw Blame History

Flutter delivery queue

Behavioral parity with the existing web app, native controls, web-only testing. All work under app/. Existing applications remain running and unchanged.

First wave (independent worktrees, Task v6, Luna for every stage)

Item Scope Source Verification Landing Preview
F0 Scaffold and gateway contract 0e0ef54 analyze + 1 scaffold test pass main not served
F1 Socket transport, endpoint rules, reconnect run 29 + 1e4ecea 12 regressions pass, analyze/web build pass main cb1ea90, ancestry verified —
F2 Adaptive shell and standard platform controls run 31 + a561c45 6 widget tests; real Chrome navigation + inspected phone/desktop PNGs main 2f60a0e, ancestry verified served skeleton
F3 Run models, live event fold, actions run 32 + 1515ebe 11 regressions/analyze pass; 5 source files 105–226 lines main 37d0823, ancestry verified —
F4 Chat models, history/live fold, request facade run 33 / flutter/f4-chat pipeline passed; wire omissions carried into C1 pending C1 adapter fixes —
F5 Isolated authenticated web preview tooling run 30 + ae352bb 6 proxy fixtures incl auth/bytes/origin/idle sockets pass main 160ab45, ancestry verified http://127.0.0.1:8096/
P1 Pipeline list/detail/source/version browser run34 / flutter/p1-pipelines UIReview found whole-word tab-label scaling defect; visit3 export waits on duplicated inline screenshot data held for exact export recovery then review-directed fix pending
C1 Chat session browser and management 35 →36 →42 / flutter/f4-chat original export42 repaired; Review resumed pending review/landing pending
N1 Connection settings/preferences run37 / flutter/n1-settings original export repaired; Review resumed pending review/landing pending
R1 Run list/detail/live timeline controller43 + list48 + detail46 +35a8aab manager checked retained candidate and real browser captures merged main; ancestry verified built into8096 preview
D1 Recorded changes/diff viewer run39 / flutter/d1-diff-viewer visual-review audit now export-waiting; inline-payload correction staged held for exact export recovery pending
R2a Human-decision native confirm run47 / commit42b21b8 manager inspected source + reran 7 exact-confirm/refusal/native tests + analyze, pass main3e6f950, ancestry verified reusable component only; wiring/browser pending
R2b Native cancel-run confirmation run51 →52 / correction5449370 manager reran23 action tests + analyze, pass; known refusal now immediate and dismissible while resync pending merged main; ancestry verified reusable component only; wiring/browser pending
X1 Native search and result targets run49 / commit51140cd manager reran36 tests, checked native browser evidence + source; all pass merged main, ancestry verified standalone screen; shell wiring pending
I1 App-owned socket/session + live Runs shell registration run50 timeout → manager finish /35a8aab 68 focused tests; integrated375 tests; clean analyze/build; normal-entry browser probe passed18 captures incl deep links/history/reconnect merged main; ancestry verified 8096 bundle rebuilt, auth preserved; live authenticated smoke not verified
E1 Narrow authorized server export-bound fix run53 / source22801a9 / deployb4627e2 bounded regressions + full isolated Go suite pass; exact installed revision verified source merged; deployed minimal patch on original server revision installed; seven exact exports complete, Review resumed
R2c Native retry/restart confirmation run54 / commit178f22a manager reran17 focused tests + analyze, pass; exact modes/stable intent keys checked main804e531, ancestry verified reusable only; wiring/browser pending
R2d Native rename-run dialog run55 / commitf2643a0 manager reran11 focused tests + analyze, pass; clean scoped diff main3dde288, ancestry verified reusable only; screen/list update wiring pending
S1a Saved schedules repository/revision safety run56 / commitb963965 manager reran18 focused tests + analyze, pass; revision/unknown field preservation checked maina290858, ancestry verified —
T1a Templates/validate/create-only requests run57 / candidateef2f463 manager reran13 focused tests + analyze, pass; exact create-only/source preservation checked merged main, ancestry verified —
S1b Native read-only schedule list run58 / commitebfa7f7 manager reran29 tests/analyze and inspected2x phone PNG; browser-review advisory accepted merged main, ancestry verified standalone; navigation/editor pending
P2a Graph identities/recorded-visit model run59 / commit2c10507 manager reran41 focused tests + analyze, pass; scoped 4 pure domain files merged main, ancestry verified no renderer/layout yet
P2b Content-sized graph node layout run60 / commitc1f140b manager checked caller-supplied extents + reran21 tests/analyze, pass mainec9e889, ancestry verified geometry only, no renderer
A1a Artifact records/grouping/safe URL navigation run61 timeout → resume63 / commitff0f812 preserved780 source lines; manager reran60 tests before/after resume, pass; URL security checked merged main, ancestry verified data only
P2c Edge routes and measured label geometry run62 / commitb381290 manager reran20 tests, checked geometry entrypoint/measured inputs, pass merged main, ancestry verified no renderer yet

Next ready work

Current execution target (user correction)

Target approximately20 minutes END TO END per task, not per model stage. One independently testable behavior, focused regression, short handoff; concurrent small lanes, prompt integration. No multi-hour feature-sized runs or blind retry/review loops. When a slice overruns, inspect retained progress and split unfinished behavior or finish directly. This changes task selection/prompts, not pipeline definitions. Keep existing in-flight work intact.

Ready micro-slices after the necessary audit-fix service window:

  • Graph: native measured stage-node widget + selection callback ONLY.

  • Graph: edge/label painter over already-routed geometry ONLY (parallel node lane).

  • Artifacts: read-only visit-group list + selection callback ONLY (no viewer/share).

  • Schedules: editor input validation + unchanged revision payload tests ONLY (no UI). Keep each item separately reviewable; renderer composition, artifact preview and schedule save integration are later slices, not bundled into those tasks.

  • P1 pipelines list/detail/source/version; P2 graph and stage/visit selection.

  • R1 runs list/detail/timeline; R2 decisions/run/retry/cancel/title and attention.

  • C1 chat list/group/search/star/rename/close/create; C2 transcript/markdown/reasoning/tools; C3 composer/attachments/queue/stop; C4 forms/permissions/models/metrics/shells.

  • A1 artifact gallery/preview/share/download; D1 diff viewer.

  • S1 schedules and provenance; T1 templates/create/validate/submit.

  • X1 search; N1 settings/persistence/notifications; I1 routes/deep links and integration.

F1 + F2 unlock live preview and feature screens. F3/F4 unlock runs/chat UI. F5 provides a separate authenticated preview URL, without production changes. Parallel tasks own separate feature subdirectories and tests. Manager owns main, shell registration, pubspec, this queue and integration. Small shared-file edits are resolved at landing, not used to serialize independent features.

Each run pins Task v6 with a whole-run openai/gpt-6-luna override (all four stages), app/ as workspace, its own worktree under tmp/flutter-fN, and an individual background pipelines wait. No workflow definitions were modified. The CLI reports a client/server build mismatch; admission succeeded and the server echoed the expected model/scope. No shared service was rebuilt/restarted.

Corrected admission scope

Run 24 failed before edits: --write-paths app grants an exact path, not a directory subtree. Runs 25–28 were stopped once the common admission mistake was confirmed, and their agents settled. Partial shell/run/chat files remain in the same worktrees. New runs 29/31/32/33/30 continue those retained files with app/ itself as workspace and default workspace edit scope. Retry preserves the old workspace/write scope, so could not correct this admission mistake. No policy override, server edit or shell-write bypass was used by the manager. Allowed pipeline models: Luna or DeepSeek V4.1 Flash; current runs use Luna.

Latest integration check

After Runs landing, integrated main:375 Flutter tests pass; analysis and normal web build pass. Manager reran browser probe against that normal bundle:18 PNGs, deep link/back/forward/drop/reconnect pass,5 subscriptions/5 lists/7 detail reads. Preview8096 serves the rebuilt build/web without any service restart. Rollback bundle retained in approved temp as preview-before-runs-1792600010. LAN/tailnet /runs reach the authenticated login correctly. Authenticated production data rendering not tested; fixture evidence is not a claim of live-data verification.

Visual-audit export correction prepared (not installed over active work)

Run34 UIReview3 native source is complete:71 messages,4,196,297 decoded bytes. Its projection duplicated47 inline screenshot URIs (3,752,634 bytes) already stored in complete native JSON, exceeding the unchanged8MiB HTML cap. Direct infrastructure correction c303756 on exact-running-b4627e2 base (main cherry-pick ceffd5d) summarizes ONLY data: URIs whose exact bytes are also present in the message's retained native record. Native JSON is unchanged and escaped; absent native copy falls back to full output. No limits raised and no partial evidence declared complete. New large-inline/fallback/security regressions, full Go suite, build/vet and focused race tests pass. Exact original visit scratch export now 4,676,544 bytes/71 messages/untruncated. Staged binary is in approved temp as pipelines-server-inlinefix-c303756. Original production evidence untouched. Install and exact export-only repair34/3 automatically at a quiescent server window; do not interrupt active37/39/42/50/58 work. Review's rejection is legitimate: Summary/Source tab labels break mid-word under scaled text; once the audit gate is repaired, the same run follows its recorded Code edge to fix that actual defect. On schedule-list58 completion manager checked the safe window:37 Review, 42 UIReview and50 Code still active;34/39 visual audits export-waiting. Therefore no restart was performed. Do not launch additional model stages ahead of this short necessary deployment window; finish/check current candidates meanwhile.

At804e531 manager ran the integrated main Flutter suite: 98tests passed and flutter analyze was clean. These cover landed source/components only, not the six unreviewed export-blocked feature branches or visible screen delivery.

Export blocker RESOLVED — direct manager recovery

User explicitly directed manager to perform routine unblock/install without waiting for further decisions. Confirmed ZERO active runs and quiescent original Code sessions, backed up SQLite via backup API plus installed binary/plist under ~/.local/share/pipelines/backups/exportfix-b4627e2 (private directory), atomically replaced only the server binary and restarted exact launchd job gui/501/net.awan.pipelines.server. No web/ntfy/OpenCode service was restarted. Installed ping commit is b4627e29755c3496f5caee5a9c316d8bf24f0713; SHA256 f70b4e17324262cf5275927c40f617838fefc0d9217946eeaf6d3a320c786fac. Rollback binary remains saved, DB/session/workspaces untouched.

Export-only repair of34/37/39/42/49/50/58 visit1 ALL returned complete, untruncated, 2.25–4.28MiB. Original Code session IDs and output hashes match the pre-install checkpoint, with one Code visit each; all seven moved to Review, export_pending_sequence0. Evidence in approved OpenCode temp: export-recovery-before.json, export-recovery-repair-results.json and export-recovery-after.json. EOF/reset errors on old background waits were the expected restart disconnect, NOT run failures. Reattached seven waits using a private CLI built from the sameb4627e2 source; installed global CLI left alone. No coding-stage replay, evidence overwrite or gate bypass. Network preview still responds302 to protected login at http://10.0.0.249:8096/. Implementation/reviews now proceed in parallel; the prior install-approval notes below are historical.

The fix is now merged: rendered file cap8MiB, source cap8MiB unchanged, within ordinary artifact allowance. Pipeline read-only scratch exports of all six exact sessions were complete at2.25–4.28MiB, establishing the fix is sufficient. Run53 automatic delivery failed in pre-existing TestHumanRunAttentionNotification because the test inherited the running service's PIPELINES_WEB_PUBLIC_URL. Manager reproduced that same failure on unchanged main with a harmless example URL; no notification source/test was edited. All packages pass on candidate with only that test-process env var unset; build/vet/diff pass. Manager landed checked source directly (no coding-stage replay), and refreshed installed managed Pipelines skill from current source via go run ./cmd/pipelines skill install. Installed server is unchanged. Deployment safety inspection found57/58/59 active; no restart or binary replacement performed. Need explicit operator approval for installation/restart after active work settles before exact export-only repairs. Old CLI service status cannot resolve its expected launchd label; do not guess a service or restart unrelated processes. Inspect authoritative supervisor first. Preparation completed after user asked what manager was doing: authoritative launchd label is net.awan.pipelines.server (not old CLI's net.a-wan.pipelines), running installed binary under ~/.local/share/pipelines/bin with original DB and OpenCode endpoint retained. Built a deployment candidate from the exact running9ee2e00 revision plus ONLY export fix22801a9: branch deploy/export-limit-only, commitb4627e2. All six new export regressions pass on that exact candidate. Staged binary in approved OpenCode temp directory is pipelines-server-exportfix-b4627e2; rollback copy is pipelines-server-before-exportfix-66779. Installed binary/launchd registration remain untouched. This avoids deploying unrelated newer main changes. The temporary launch pause ended after the direct, safe server installation. Graph/model/layout lanes have progressed, and independent small tasks continue alongside the seven restored feature reviews. No repeat approval gate remains for routine authorized recovery/integration.

User asked about failed runs; explicit status inspection found runs34/37/39/42/ 49/50 are waiting, not failed: each Code visit1 computed success and is quiescent, but export is partial/truncated at exactly2097152 bytes. Exact recorded native session identities remain retained. pipelines repair-session-export 34 1 was attempted once; it returned session export repair incomplete: partial at the same cap. No full stage reruns, evidence replacement, cancellation or gate bypass was performed. Remaining waits stay active for eventual export-only recovery. User subsequently approved the narrow server export-limit exception and asked for continued small parallel tasks. Run53 owns ONLY bounded export-limit code, necessary verification/repair regressions and a focused note. Existing web and mobile source remain off-limits. Deployment remains a separate operator gate under internal/server/AGENTS.md; no shared service restart is delegated. After authorized installation, recover the original exact visits EXPORT-ONLY. Do not burn repeated repair attempts against the unchanged deterministic size limit.

Parallel work while E1 runs: R2c/R2d/S1a are separate worktrees, each with its own focused existing Task v6 run and background wait. No new workflow definitions.

At 160ab45: 29 Flutter tests and 6 Node proxy fixtures pass; flutter analyze, flutter build web and git diff --check pass. Real Chrome phone/desktop captures are in the approved OpenCode temporary directory, named f2-shell-*.png. The preview process is running separately on loopback 8096, authenticating via the unchanged existing web proxy on 8095. Verified unauthenticated page redirects to login, JS denies with 401, and proxied login responds 200. The served app is currently a shell with explicit unimplemented bodies, NOT a feature-complete app. No Android/iOS build or device verification has occurred.

Network preview update: per user request, the preview now binds 0.0.0.0:8096, reachable at http://10.0.0.249:8096/ on the current LAN. Verified listener *:8096, HTML login redirect 302 and unauthenticated JS 401 via the LAN address. Existing proxy/API services remain unchanged. Bind command uses --host 0.0.0.0 explicitly; the source default remains loopback for safety.

HTTPS access fixed for the user's Tailscale URL: https://alans-macbook-pro.tail6bf4dc.ts.net:8096/. Port8096 originally spoke only plain HTTP; added tailnet-only Tailscale Serve HTTPS8096 → http://127.0.0.1:8096. Existing Serve443→8787 and8443→8095 routes preserved. Verified trusted certificate (curl ssl_verify_result0), protected root302 to login, login200, unauthenticated JS401; LAN HTTP still302. No Funnel/public-Internet exposure or auth disabling. Prior Serve config saved in approved temp tailscale-serve-before-preview8096.json; remove ONLY this route with tailscale serve --https=8096 off if rollback needed.

Manager integration findings

  • F3 candidate omitted lib/features/runs/data/runs_repository.dart: root runtime-data ignore matched source. Main 718442d adds an app-only exception; candidate branch has it merged. Source is preserved, awaiting corrective commit and expanded lifecycle/decision regression tests before landing.
  • F2 UIReview was auto-skipped despite rendered shell changes (existing server classifies workspace-relative lib/ paths as nonvisual). Do not change server; manager obtains browser evidence directly and keeps rendered acceptance open.
  • Direct Luna correction agents are scoped to the idle F2/F3 worktrees. No recovery workflow or pipeline version was created. P1 is independent and already started while corrections proceed.

Completion loop

Manager role: check evidence/diffs, verify, merge, update preview and start next ready items. Per latest user steering, substantial implementation and corrections belong to existing pipelines, not manager-written replacements. Already-running direct correction work is preserved; do not discard it or duplicate it.

Use pipelines wait for every started run. Inspect failure evidence before fixing, resuming or retrying; preserve partial work. On success inspect actual diff and checks, integrate, verify ancestry, then launch the next ready item. Do not rerun completed work just because session export or landing failed. Keep source, tests, landing and preview states separate. No production cutover in this phase.