Skip to content

CI reliability: the census names the child that left its unit (#433); Assistant keeps the finished reply in view - #533

Merged
andersalm merged 3 commits into
developfrom
fix/433-census-flake
Nov 5, 2026
Merged

andersalm merged 3 commits into
developfrom
fix/433-census-flake

Conversation

@andersalm

@andersalm andersalm commented Nov 4, 2026 •

Copy link
Copy Markdown
Contributor

Deepen notes

  1. Expert view: keep the census diagnostic change. For the Assistant failure, distinguish a delayed scroll from a lost follow and preserve readers who scroll up.
  2. Raw material: census: scripts/ci-installed-journeys.py:468-687 at fa4d905 and the original failed receipts. Assistant at d497f78: scripts/assistant-reply-layout-browser-smoke.mjs:199-236 already polls for five seconds; capsules/assistant/browser/agent-stream.js:2078-2087,2924-2946,3139-3151 controls the final scroll; agent-harness.js:469-487 changes composer clearance. CI run 41933322796 and the unrelated G5 Private Browser: Jetson support #90 remote gate show the same -774 px failure.
  3. Methods: retain the census snapshot/redaction checks. Repeat the Assistant smoke with pinned Chromium at 4x CDP CPU throttling, then control final-render scheduling around a composer resize and measure the settled scroll position. Change the product only if the original code fails that browser check.
  4. Independent check: census fixtures: 124 passed; Claude review found the redaction issue in 3ee171a, corrected in d497f78. Assistant: 50 original runs at 4x CPU throttling and 15 with four CPU workers passed. A controlled layout clamp during deferred finalization failed on the original product 3/3 times: presentation_done, atEnd=false, reply 66 px behind the composer after 5 s. Check both saved-pin comparisons against the current reachable scroll position. Local candidate fe0073c: controlled regression 10/10 PASS, reply 36 px above the composer; behaviour suite 2055/2055 PASS; full source gate PASS; diff check PASS. Both saved-pin checks now use the current reachable scroll limit; generation and reader controls remain. Lesson 923 records the failure. The exact 774 px CI offset remains unreproduced.

Problem

#433: the installed journey in source-home-linux (ubuntu-24.04) fails intermittently with "finite journey live child left its unit census" on changes that do not touch it (3 Nov develop, #526, #532). The receipts do not say which child left its unit or where it went, so a real escape and an early census sample cannot be told apart.

Change

Diagnostics only; the census assertion and cleanup checks are unchanged (no retry, wait or skip). On failure the receipt records, for the missing child and its parent: pid, start tick, state, role, executable hash, pidfd state and current cgroup, plus a re-read of the expected unit's cgroup.procs. Public cgroup paths keep systemd unit names and replace other components with <name>; full arguments go to a 0600 private log. The first failure snapshot is kept before cleanup.

Next failure decides product bug (escape) or census bug; #433 stays open until then.

Local Assistant follow-up (awaits publication)

Commit fe0073c47 preserves the final scroll follow when layout clamps the saved scroll position while the Markdown render is pending. The browser regression holds that render, increases viewport height from 800 to 1600 under 4x CPU throttling, then checks the completed presentation with the existing end-position and composer-clearance assertions. It restores the desktop viewport for the remaining reader-scroll checks.

The commit is local, directly after d497f7892. The published PR head remains the census candidate. Anders owns publication; current-candidate independent review and CI confirmation remain before integration.

Also: Assistant keeps the finished reply in view (fe0073c)

The Assistant reply layout smoke failed twice on 4 Nov (remote gate for #531, CI for this PR): after a streamed reply finished, the newest reply was left above the composer. Codex reproduced a real product bug: when the layout shrinks the scroll range, the saved scroll pin is clamped and Assistant wrongly concluded the reader had scrolled up, so it stopped following. The guard now compares against the reachable scroll limit; a reader who really scrolled up still keeps control. Throttled regression: 3/3 fail before, 10/10 pass after. The exact 774 px interleaving from CI was not reproduced (50 throttled and 15 loaded runs). Reviewed by Main (Claude).

Keep the first missing-child snapshot through cleanup. Record retained and current PID, parent, start tick, command, cgroup and pidfd observations. Keep raw arguments in a private runner log and safe fields in the uploaded receipt. Preserve the immediate failure and existing cleanup gates.

Verification: 35 finite-journey tests pass. The diagnostic fixture fails on the base; the first-capture regression fails before its keep-first guard and passes after.
Compare saved scroll pins with the current reachable scroll limit, so a layout clamp keeps the follow through deferred Markdown finalization. Keep the generation guard and reader scroll controls.

The pinned Chromium regression holds the idle render, resizes the viewport under 4x CPU throttling, and checks the settled reply with the existing assertions. Before: 3/3 failures, reply 66 px behind the composer. After: 10/10 passes, reply 36 px above it. The exact 774 px CI interleaving was unreproduced in 50 throttled and 15 CPU-load runs.

Checks: elastos-heavy ./scripts/test-behaviour.sh (2055 passed); elastos-heavy ./scripts/test-source-gate.sh (exit 0); git diff --check. Failure recorded as lesson 923.
@andersalm andersalm changed the title CI: the installed-journey census names the child that left its unit (#433) CI reliability: the census names the child that left its unit (#433); Assistant keeps the finished reply in view Nov 5, 2026
@andersalm
andersalm merged commit 27ea103 into develop Nov 5, 2026
16 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant