Retrospectives

Blender AI tutor: retrospectives

The orchestrator writes one entry after each stage's cleanup, before briefing the next stage (rule "Retrospective between stages" in How the build runs). Newest first.

Each entry has:

• What cost time: failures, reruns, supervisor questions, wrong predictions in briefs or reports.
• Code that keeps changing: files changed again this stage after changing in earlier stages, from git log --format= --name-only across stage merges; candidates for a refactoring lane.
• Rule changes, applied to the plan and the next briefs:
  • removed;
  • modified;
  • added.
• Refactors queued: lanes added to the next stage, or "none".
• Orchestrator's decisions this stage, each with the question posed, the options and the choice, marked "Axlog can reverse".

---

chat-ui-and-mobile, 2026-10-12

What cost time

• A known flake stayed open through all four lanes. The chat visual shots failed now and then on main too (1 px of height, or a blank band under the composer). Three lanes spent runs on it and each asked the supervisor about it, until visual-upkeep fixed the settle step (fit-screen.ts: 5 of 60 shots failed before, 0 of 80 after). Fixing it first, alone, would have saved runs in every lane.
• Supervisor questions: 12. Most were real out-of-path edits named in no brief (the hide-rule layer, biome.jsonc's viewport, chat-thread.tsx, memory-history.tsx, Streamdown's buttons, WebKit sign-in). The brief predicted the shared files but not the edits the lanes needed in them.
• Two merge interactions that no lane could see alone:
  • min-h-11 on Rename chat: a spacing utility beat the role's --target-size and followed the theme.
  • mobile.spec.ts measured Streamdown's off-screen buttons at 0×0 under load: one failed full run.
• Reviewers found a bug the integration's probe missed: switching off the pinned header sets it static, so its fade strip floats one screen down. The probe checked computed properties only; no test switches that flag.
• Parallel lanes duplicated a helper being centralised. mobile-fixes wrote the theme cookie by hand while visual-upkeep exported themeChoiceCookie for exactly that.
• Smoke worker's shell slip sent a stray "x" message (within the cap). The sweep report missed its exact final verdict line.
• Bash-open alerts: 9, all long test runs. Each cost a process check.
• What went well:
  • The merge-tree predictions matched every conflict exactly.
  • The Git deployment fired again (4 s).
  • 248/248 e2e twice on the merge.
  • Every screenshot reviewer said OK.
  • The device projects worked on their first review.

Code that keeps changing

Commits since 2026-10-09 (non-merge):

• src/app/page-roles.css 9, 290 lines: roles for every page in one file. Every UI lane edits it.
  • Watch: accounts adds the account page's roles. If two round-3 or personas lanes need it at once, split it by page area (page-roles.css for the shared roles, chat-page roles beside chat.css).
• src/components/chat/chat.css 7: chat and Markdown roles. It churns with chat features.
• e2e/visual/visual-page.ts 5: settled now that fit-screen.ts holds the screen fit.
• src/components/site-header.tsx 5: 25 lines. Every stage adds a link. Accounts adds Account (and may hide links signed out). Small, so no split.
• src/tutor/reply.ts 16: untouched this stage. Re-measure after accounts, as planned.

Rule changes

Removed

• "Record the chat visual flake as an open risk" (option A): the flake is fixed, so a failing visual shot is a failure again.

Modified

• Touch targets: sized through --target-size, never spacing utilities. Recorded under "Patterns the rules force".
• Interaction probes check what the user sees. A known-interactions step in the integration brief also takes a screenshot or a click({trial:true}) check, not only computed styles.
• Section C's report must end with the exact line Sweep verdict: OK or Sweep verdict: BROKEN. The after-merge workflow passes the sweep's verdict into its return value.

Added

• Each extra's off state has a browser test, written by the lane that builds or tags the extra. The rule is in common.md for accounts and personas; it came from design M1.
• Helpers centralised in a stage replace parallel copies at merge. The integration brief lists the helpers each lane exports for others (themeChoiceCookie, manualSection, PHONE_VIEWPORT). The integration worker converts any parallel lane's hand-written copy, as a merge fix.
• A known flake is fixed first, alone, before the lanes that would trip on it start, or in round 1 of the stage.

Balance: 1 removed, 3 modified, 3 added.

Refactors queued

• chat-ui-followups (accounts round 1, beside accounts-core and use-cases-accounts):
  • pinned-header off state (design M1, rules M1);
  • typed extraTag and the tag test (design M2);
  • mobile.spec.ts on the shared helpers (design M3, rules L2);
  • the untested use-case promises (rules M2);
  • Lows: design L2, L4, L5; rules L3; README's page lists (design L1); use_case_use_on_phone naming Settings.
• Into account-page-and-flags:
  • design L3: decide on a shared "choice" helper (cookie, pre-paint lookup, <html> observer) once flags store per user;
  • design L6: seed the cookie at sign-in with hiddenFlagsCookie;
  • the signed-out Settings link.

Measured

• Smoke reply cache: cacheRead 39.6k, cacheWrite 7.5k of 47.2k input. The reply searched 5 times, so the input was bigger than lane-splitting's 26k.
• Opening: 3.55 s, down from 4.9 s.

Orchestrator's decisions this stage (Axlog can reverse)

All are recorded in the plan under chat-ui-and-mobile ("Supervisor approvals" and the after-merge notes), each with the question posed and the rejected options:

• the hide rules in a last cascade layer extras;
• the failed-tool tag in chat-parts.tsx;
• manualSection shared in e2e/chat-page.ts;
• chat-composer-dock;
• the viewport export allowed in biome.jsonc;
• PHONE_VIEWPORT in e2e/viewports.ts;
• Rename chat unpinned;
• the memory mask class;
• Streamdown's buttons at 1.5rem;
• WebKit sign-in by minted cookie;
• hidden chat names as a split link;
• at merge, Rename chat on --target-size;
• at merge, mobile.spec.ts measuring each control on screen;
• the chat name clamped to one line on phones;
• the follow-up fixes as a third round-1 lane in accounts, rather than waiting for deploy-v1 (the off-state bug is live in production, and round 3 adds an extra that needs the typed tag).

---

lane-splitting (and persona-tweaks), 2026-10-12

What cost time

• A lane ended its run on a failed check. visual-baselines' first run stopped when biome check failed (useValidTestTitle on a template-literal title), with nothing committed; it needed a resume. The fix was ordinary; the cost was the stopped run.
• A wrong "merges cleanly" prediction. copy-per-area's report predicted no conflict in site-header.tsx; theme-polish's edit on the adjacent line conflicted. It cost one supervisor round trip. Reports guessed by reading diffs rather than testing the merge.
• A brief bug. The after-merge brief's pathspec ':!e2e/visual/*-snapshots' didn't exclude the PNGs; the worker caught it and asked.
• Deployment. The push produced no Git deployment at all; the CLI fallback ran. The smoke worker noticed that vercel deploy from the repo would upload .env and .env.test (the CLI ignores only .env.local by default) and deployed from a clean worktree. Good judgement, but nothing in the brief protected against it.
• /reload stopped the workflow controller. The child kept running; relaunching the same script re-attached. No work lost.
• Bash-open alerts (three) were all long test runs. Each cost a check.
• One secret leak into a tool output: the integration worker printed .env.test's test-only passphrase while reading the file. Not copied anywhere.
• What went well: the visual baselines replaced throwaway screenshot scripts and pixel comparisons in one stage; reviewers used them directly; e2e 147/147 twice; the five-branch merge with an old-base branch (theme-polish) needed no spec fixes.

Code that keeps changing

Commits touching each file since 2026-10-09 (non-merge, src and e2e):

• src/ui/copy.ts 18: resolved by copy-per-area.
• src/tutor/reply.ts 16, reply-test-kit.ts 10, reply.test.ts 10: every reply-touching lane edits them. 287 lines; not oversized, but it is the hottest file and multi-persona-chats will reshape it. Watch: re-measure after accounts-and-quotas; if accounts lanes also edit it, queue a split (request parsing / streaming and saving / step logging) before multi-persona-chats.
• src/components/chat-view.test.tsx 16: 263 lines, 3 describes. Churn follows chat features; no split yet.
• src/components/new-chat-form.tsx 15, new-project-form.tsx 13: persona-picker churn; persona-editor and multi-persona-chats replace most of it.
• src/tutor/system-prompt.ts 12, its test 13, src/personas/personas.test.ts 13: queued as persona-parts-move (round 1 of personas-and-prompts).
• src/chats/store.ts 11: persona-store rewires it.

Rule changes

Removed

• "Lanes save screenshots; fresh agents view them", throwaway screenshot scripts and script pixel comparisons (with the RGB note): replaced by the visual baselines.
• The 175k "commit, hand off and end" rule and the context-handoff add-on: replaced by compaction at 175k (Axlog).
• "A fresh session per stage": one long session with compaction (Axlog).
• The visual-baselines lane's "don't import COPY" limit: it ended with the merge.

Modified

• Conflict predictions are tested, not guessed. The integration worker's first step runs git merge-tree --write-tree for each branch against the merge so far, in order, and lists the predicted conflicts before merging; lane reports stop predicting merges.
• Deploy fallback only from a clean worktree of the pushed commit; chat-ui-and-mobile adds a .vercelignore with .env*.
• Diffs that exclude a folder use ':!<dir>/**'.
• Secrets: briefs add "never print .env* files; read single variables with node --env-file or grep -c".

Added

• A failing check isn't the end of a run. Common brief line: "A failing command is information. Fix it within your rules, or ask with contact_supervisor; never end your run on a failed check."
• Visual pages wait for a settled page and real masks. describeVisualPage waits until no loading skeleton (aria-busy) remains, and asserts every mask locator matches at least one element (rules R-L2; the skeleton baselines).
• Mobile is reviewed every stage (Axlog): device projects in the visual tests and a mobile reviewer.

Balance: 4 removed, 4 modified, 3 added.

Refactors queued (into chat-ui-and-mobile)

• theme-choice.ts exports the cookie form (themeChoiceCookie) and visual-page.ts uses it (design M2, rules R-L1).
• Visual specs import copy modules; masks built from copy; non-empty mask assertion; settled-page wait; shared manualSection (rules R-L2, R-L3, design L2, L3).
• themes.ts "ADDING A THEME" step 5 (design M1).
• CSS roughs: memory milestones' double marker and hanging indent; the checkbox gap on the spacing unit (R-L4); the composer's backdrop at 390; dark skeleton colour.
• .vercelignore.
• Left for accounts-and-quotas (it owns use-cases/): design L4 (README wording), L5 (tighter slug check).

Measured

• Smoke reply cache: cacheRead 19.7k, cacheWrite 6.3k of 26.0k input. The baseline for personas-and-prompts' cache-cost check.

Orchestrator's decisions this stage (Axlog can reverse)

• site-header.tsx conflict: take theme-polish's side, then convert COPY (rejected: hand-merging the two lines). Standing rule for the same pattern.
• Diff pathspec: ':!e2e/visual/*-snapshots/**'.
• Compaction numbers: reserveTokens 825000 for Opus 5.5 (trigger at 175k, matching the old limit), keepRecentTokens 30000 (rejected: the default 20000, a little short for lane state).
• Stage order: chat-ui-and-mobile before accounts (rejected: beside accounts, because the pages overlap).
• Request details for personas, prompt parts, multi-persona chats, the interview, flags, the pinned header and mobile: all recorded in personas-and-prompts-rulings and chat-ui-flags-and-mobile, each with the question posed and rejected options.
• Mobile checks added to the running sweep before it started.