The user's administrator repair request was “please fix the large project board issue.” The change reduces the cost of rendering all matching issues while preserving all cards, priorities, recorded details and normal interactions. No issue mutation, schema/data migration, account/capacity change or new budget was made.
Closed cards previously built rich audit subtrees and repeated expensive issue/console copies. The shared board now mounts a rich body only while its disclosure is expanded; closing releases it. Rich issue records are shared using Arc across quadrant/card computations. Existing Markdown, comments, attachments, requirements, audit entries and financial reporting remain available.
Reviewed implementation: 7f3db3724f2da74e63198e25ea042c2e1ad4b79b, PR 6817. Parent and independent source review were recorded as agent review, without claiming human acceptance. Normal integration into dev: 1a9743b1d886200773643fd2cfebbcaaa1a2b5ff. Main integration: dd3524deb1fd0a3147c82a87fdbc8e89ce3c2279. All three source trees match e442f39343dac1fcbb1bcc5820402f408a5d30dc.
Branch validation 4145 passed. Immutable candidate 4146 passed all four actual jobs/tasks. Candidate logs show 480 library tests, 967 main tests, 24 Gitea MCP tests and one Kubernetes test passed, all four configured database shards passed, and backend/frontend builds succeeded. The separate worker browser regression passed 300 rich synthetic cards with 40 comments each, refresh, priority persistence/rejection, native/finding identity collisions, source focus, search and 20 reopen cycles. Closed fixture cards mounted zero audit bodies.
TEST: autocoder-test/autocoder-test-engine and autocoder-test/codecartography-test-frontend; production: autocoder-engine/autocoder-engine and autocoder-engine/codecartography-frontend. All four deployments were Ready 1/1 at source 1a9743b1d886200773643fd2cfebbcaaa1a2b5ff, candidate 1a9743b1d886200773643fd2cfebbcaaa1a2b5ff-4146-1.
sha256:62e59b403e4029bd6b6d9548ff6aa4ae5c5f67dc4d692ab13159ec54c10f82a4.sha256:d957261fabc3fdeeff945892e130040a9b35ba51430c616a24a00e49864c1a76.Both origins returned HTTP 200 for health and the exact frontend revision. Authenticated automation status reports global, implementation and requirements automation paused; canonical shared instructions verified at digest 8a7099fa2c9ba56acd3e12332c62ebb7299c2aafcf02b81ca9e5fcce3dc72a12. Runtime effort remains high. DEV backend/frontend remain at replicas 0. No paused automation was resumed and no running workload environment was edited ad hoc. No worker image publication was necessary for this UI change.
Old production baseline source 67dc0d4896dd6c5206b72c72ad8269c6c32eee6d and the repaired production source were exercised on worker01 with the existing immutable acceptance image sha256:b007f8db6907eba646c98bdb4e3db00383c460d09f8aa15e081b6ebe4b8f2750, fresh 2 GiB browser pods and the same 1440×1000 viewport. Normal navigation selected SuperProject AutoCoder, Issues, All issues. All 225 native AutoCoder-Engine cards retained their quadrant totals 10/67/0/53/95.
Measurements use the complete browser pod cgroup peak, actual mounted DOM elements and the exported WebAssembly linear memory buffer. The WASM download size and JavaScript heap are different measurements. Both old and repaired runs exercised card reopening, search/filter restoration and reload. The new run additionally compares real audit, comment and financial section counts to the server's recorded source.
The historical browser OOM was not reproduced in the old baseline; the verified finding was excessive memory use. Old peak was 1,417,240,576 bytes, DOM 82,032 elements, WASM linear memory 97,779,712 bytes, with zero OOM events. This receipt does not claim proof of the original crash conditions.
The browser used authenticated read routes, stubbed only write-capable startup repository discovery (using a real read-only SuperProject list) and chat-session initialization, and blocked other mutations. Those startup writes were not validated. No real provider/model call was made by these checks. Credentials were supplied in memory and not recorded.
The final fresh browser run passed at the exact served revision. All 225 unique native card identities and all five quadrant totals survived 20 reopen cycles, title search, type/state/stage filtering and reload. Closed cards consistently mounted zero rich bodies; expansion mounted only the selected body and collapse released it. Real issue #1574 retained its 16 audit detail entries, comments, original-request section and one financial component with unknown cost status, matched against the actual API response before and after reload. The stored raw audit counter was zero despite 16 entries; this pre-existing source inconsistency was preserved rather than rewritten by a UI repair.
| Measurement | Old production baseline | Repaired production |
|---|---|---|
| Matching issue cards | 225 | 225 |
| Mounted DOM elements, closed board | 82,032 | 2,521 |
| WASM linear memory | 97,779,712 bytes | 63,045,632 bytes |
| Browser pod cgroup peak | 1,417,240,576 bytes | 679,219,200 bytes (647.8 MiB) |
| OOM / OOM-kill events | 0 / 0 | 0 / 0 |
Peak memory fell 52.1%; DOM elements fell 96.9%. No JavaScript/WASM errors, console errors, failed API reads or attempted issue mutations occurred. No repeated board request cycle occurred while idle. The existing account-status poll (source poll_codex_accounts_status, every three seconds) remains operational.
The final script initially required corrections to its audit-counter assumption, response field path, non-unique selector and idle assertion overlooking that existing account poll. Those attempts were classified as acceptance-script defects, preserved separately and did not trigger product edits. The successful final comparison uses a fresh pod, so earlier attempts and synthetic fixtures do not contribute to its peak.
The original dirty/conflicted checkout was not edited. The isolated source worktree is clean; source branches and history were preserved. Original baseline/fixture pod UID 2774f5c9-1a9c-478f-bde0-f86254134b25 was removed and absence verified. Each failed acceptance pod was removed using its exact UID precondition. Final acceptance pod autocoder-engine/large-issue-board-acceptance-20260930, UID 11b313c1-859e-44cf-9c0c-f56f553377fb, was deleted after receipt/memory capture and absence verified. The owned isolated Docker build container 788bac1978ee4fae85356b8f2146ec3d69e74783db17505f2ad84f3905714d39 was removed after verifying task label and no active execs. No other workload, user data or cached build container was removed.
Restricted task evidence remains under /home/ax_julian/temp/large-issue-board-20260929-evidence, including baseline, successful/failed browser receipts, cgroup measurements, CI logs, deployed digests and cleanup receipts. Credentials and browser storage were not saved. Failed Playwright diagnostics may contain a selected issue title; the evidence directory and files are owner-restricted. This Wiki page is the operational change record.
Existing automation pauses, DEV stop, account throttling and high reasoning effort remain in effect. No additional financial or token ceilings were introduced. No rollback was required.