Establish routine activity, review and authoritative usage coverage #4639

Activity

1. EHotwagner

   added sub-issues on Oct 23, 2026

   • Reconcile actor, review and available usage coverage for real Delivery #4702 #4706
   • Finite usage/review census, 2026-10-19 to 2026-10-23 (child of usage coverage) #4823
2. EHotwagner commented on Oct 23, 2026

   

   EHotwagner

   on Oct 23, 2026

   ContributorAuthor

   More actions

   fdev rescope (user-delegated): one finite settled census now, as a child; whole-programme coverage stays here

   #4823 — finite census (sub-issue of #4639).

   • Population: every PR merged in FS-GG org repositories, and every Project 3 item closed, in [2026-10-19T18:19:36Z (Unify continuous board clearing, intake and complete-cohort inspection #4711 merge), 2026-10-23T00:00Z). The member list is frozen at census start and its digest recorded.
   • Per member:
     • repository/PR, head and merge SHA, route marker;
     • review evidence: an independent verdict at the exact head, or "unsampled";
     • required-check conclusion;
     • Actions machine time (Jobs API started_at→completed_at);
     • elapsed open→merge, labelled as including overlapping queues (never idle or critical-path time).
   • Native usage:
     • Codex rollouts and Claude Code transcripts on the coordinator host, read through the maintained strict readers at .github 61526e94 (#4802); allowlisted identities and counters only.
     • Tokens are attributed only where session, branch or PR evidence joins; everything else is reported as unattributed.
     • Other hosts (including the Home agent) are Unknown. Billed cost is Unknown; a list-price estimate is labelled as an estimate.
   • Effect admission: coherent113, which carries the Claude reader, is unpublished. fdev admits one source build of the usage reader (the CLI at 61526e94) for the census, read-only. The census does not wait for coherent113.
   • Output: a report under docs/reports/ plus an Atlas pointer, through a routine PR with one independent review.

   #4639 stays the Outcome and keeps whole-programme activity, review and usage coverage, with its original honesty rules. Pre-2026-10-19 history stays Unknown under #4639, not out of scope. #4706 and #4823 are accepted samples at their own scopes. Installed capture adoption stays with #4640.

   — fdev
3. EHotwagner commented on Oct 23, 2026

   

   EHotwagner

   on Oct 23, 2026

   ContributorAuthor

   More actions

   Root process findings and owned disposition at the existing clearing boundary:

   • The resumed full80/83 read retained complete population/membership coverage and zero writes, but refused final issue/source joins for4650/4820/Coordination1032/4778 because I updated those owning bodies while the read was running. This is a failed affected read, not an all-current/Verified result. I retained the original, completed the writes, refreshed the actual owning issues and qualified complete83-item native planning snapshots before further decisions. No original effect or Unknown hold was released. Finish owned body/field writes before a wide determining barrier, then refresh any independently changed joins beside their action.
   • Two manually assembled private preparation calls refused before effect: a copied request digest did not match the frozen bytes, and a reused owner-text predicate did not match the current six-field owner wording. Originals were preserved; correction used the actual sealed input digest and exact current issue-body hash. These are root construction defects, not supplier failures. Future invocation should derive final digest and arguments from the same frozen bytes in one failure-gated path; a failed setup check must prevent dependent calls. No new recovery tool, competing ledger or gate is selected.
   • PR763 native CI caught an unannotated F# test helper overload that source-only review had not compiled. The same owner repaired one line on the same PR, retained the first logs, and obtained actual successful native compilation and required checks before normal source delivery1e897a8c. Pure checks and source review are not compiler evidence.
   • Idle-agent messages do not start a review. Review completion was counted only after actual resumed reviewer work and an exact-source verdict; follow-up dispatch is required when an existing reviewer is idle.

   Root owns these dispositions and the concise Atlas capture gap; no replacement worker or retrospective Delivery is created. Accepted child source/census scopes and full parent promises remain unchanged.439 inherited plus46 later Unknown holds remain unreleased.
4. EHotwagner commented on Oct 24, 2026

   

   EHotwagner

   on Oct 24, 2026

   ContributorAuthor

   More actions

   Root disposition of new process findings since the previous disposition. These are internal measurement and source-check results; the accepted #4706/#4823 samples and this issue’s whole coverage acceptance remain unchanged.

   • RSS measurement construction — repaired prospectively. The first input-acquisition attempt refused at a 512 MiB cap using inherited ru_maxrss (945,156,096 bytes), which was not current process RSS. A separately admitted successor kept the cap and used sampled current parent/child RSS with process-birth identities: 148 samples, 30,834,688-byte observed peak, three verified archives and closed responses. The first refusal remains failed; the successor does not reconstruct its actual RSS.
   • Supported SDK context — corrected, with limits. Root incorrectly treated the private SDK401 FSI subset as the ordinary full SDK. The corrected context uses full SDK400 for ordinary SDK resolution and explicit SDK401 FSI/Core7516 for the compiler role. Original08 passed its five scoped identity/generator/check phases. That establishes those checks, not a complete build or full formal qualification; earlier SIGABRT causes remain unproved. Map06 did not retain the exact refused mapping path, so that original remains Unknown and cannot be reclassified from later policy or the previously accepted-path list.
   • Compiled mapping collection — still owned work. On the same PR1059, retained compiled01 stopped at the raw mapping-row bound. Compiled02 aggregated classified observations but stopped at the 2 MiB semantic snapshot bound. Accounting and locked restore passed; build-process exit zero did not establish completion of the ten-project build or the following thirteen receipt tests. Compiled02’s refused real-file row matched its expected source hash: this was a collection-bound failure, not evidence of substitution or a compiler defect. review_bar_rootless, under Programme/root, owns the bounded representation repair. Reconsider acceptance only after independent review preserves exact real-file checks and the first refusal, followed by a freshly admitted representative complete build and receipt-test result. These failures do not justify discarding observations, waiving caps or replaying an uncertain original.
   • Launch/retirement ordering — source repaired; original uncertainty retained. Persistent logs, report handles and the selector now open before Popen; pure causal controls reproduce the old unmonitored-child path and verify that corrected prelaunch log failure forks no child. Postlaunch reporting failures enter finite retirement without overwriting the first cause. This is prospective source evidence. The earlier named hold 4816-pure-wrapper-log-open-unknown-child-birth still lacks original PID/birth, direct wait and EOF evidence. A later empty process scan or closed lock does not release it.
   • Fixture/setup errors — separated from product results. Under umask 077 the stale-mode fixture started at 0600, so its intended 0644→0600 change was a no-op; the affected method passed separately with the intended synthetic setup. The original seventeen-test run remains failed. Separately, DOM’s oversize fixture raised Errno27 while creating the 16,777,217-byte file; its first successor exposed missing FSGG_MANAGED_GROUP_OWNER. A second, correctly bound successor passed twenty tests on unchanged source bytes. Each original is retained; neither setup correction supplies runtime or retirement evidence.

   Root retains the Atlas capture gap for the next natural source delivery. The 439 inherited plus 46 later Unknown holds, and the separately named 4816 hold, remain unreleased. These findings do not impose a global stop on ordinary authorized work, change this issue’s Done when, or establish whole-programme acceptance. No new issue, reporting ledger or replacement Delivery is selected.
5. EHotwagner commented on Oct 24, 2026

   

   EHotwagner

   on Oct 24, 2026

   ContributorAuthor

   More actions

   Root decision — 2026-10-24 ~15:10 UTC, fdev (Programme/root), under the repository owner's explicit delegation in the interactive session ('you decide the rest').

   Scope narrowing of unsupported counters. Counters the provider does not support (billed cost, other-host and unsupported usage counters) are reported permanently as Unknown and are no longer acceptance prerequisites. Whole-population coverage is accepted on supported counters with truthful Unknown reporting. No attribution, baseline or cost is invented. The rest of Done when is unchanged.
6. EHotwagner commented on Oct 24, 2026

   

   EHotwagner

   on Oct 24, 2026

   ContributorAuthor

   More actions

   fdev (Programme/root coordination, accountable owner of the remaining coverage promise): scoped acceptance evidence for root, under the 2026-10-24 scope-narrowing decision. I have not closed this issue and have not changed Status.

   Delivery. PR #4863, merged as ce38f444c5f6daf996781f869bb09c549b61bc31 through tools/routine-delivery.py (outcome delivered; routine eligibility passed). The delivered tree c14d8fe0ff5fb22692ad5b686ba1da609700aaf4 equals the tree the required CI tested. All 33 checks passed on the final head 54811ca3.

   • Report: docs/reports/2026-10-24-usage-census-4639.md. Evidence: docs/reports/evidence/2026-10-24-usage-census-4639.json (26 KB).
   • Independent review (fsgg-critic-normal): REQUEST CHANGES at 70310ef4 (two review rows misclassified, two disclosures missing); APPROVE at d8a57a66 after repair; APPROVE at 54811ca3 (a merge of main only; the report files byte-identical). The reviewer recomputed the population and digest by a different method, re-adjudicated every E, A and note row, all V rows and the N rows with review wording, and reproduced the totals.
   • Same method, reader (61526e94, read-only) and redaction as the accepted #4823 census. No new store or collector. No transcript recapture, invented attribution, baseline or cost.

   What the two reports now cover. [2026-10-19T18:19:36Z, 2026-10-24T18:00:00Z), contiguous, with no gap and no overlap (zero Actions runs shared).

   Measure	#4823	This continuation	Together
   Merged PRs (frozen by digest)	134	88 (220982cb…2f7a)	222
   Project 3 items closed	34	12 (77b9233f…e6c0)	46
   Exact-head independent verdict	44	43	87; 135 unsampled
   Actions job time, PR phase / merge pushes	283.6 h / 97.6 h	160.8 h / 70.5 h	444.4 h / 168.0 h
   Claude Code tokens (responses)	1,777,676,869 (7,330)	2,270,492,275 (9,597)	4,048,169,144 (16,927)
   Codex tokens (responses)	5,304,222,463 (32,788)	2,813,873,591 (13,938)	8,118,096,054 (46,726)
   Tokens joined to a member	65,125,639 on 3 PRs	25,072,784 on 1 PR	90,198,423 on 4 PRs; the rest unattributed

   Clause reconciliation against the full #4639 Done-when and the owning plan.

   #	Acceptance clause	Status	Evidence or reason
   1	Routine activity recorded against the settled population	Met	PRs, closed items, route markers, required checks and Jobs API time per member, for both segments.
   2	Review recorded against the settled population	Met	Every member has V or a stated unsampled reason (A, E or N).
   3	Available native usage recorded	Met under narrowing	Strict-reader counters, deduplicated by response ID, for the coordinator host. Other hosts and unsupported counters stay Unknown.
   4	Explicit missing and unsupported observations	Met	New counter register in the report (Supported versus permanently Unknown); Gaps section.
   5	Honest time denominators	Met	Job wall time is not billed time; the rule for jobs that end before they start (0 s, with signed sums) is now stated; 581 force-pushed-head runs (15.1 h) are disclosed outside the totals.
   6	Honest token denominators	Met	Deduplicated; attributed share stated (1.1% of this window's Claude Code tokens); the runtime change and the overlap of the Codex and Claude sessions are stated.
   7	Honest cost denominator	Met under narrowing	Billed cost is Unknown; no estimate.
   8	Unsampled work is not presented as reviewed	Met	135 of 222 are labelled unsampled with reasons; independence rests on the record's own assertion (one account posts every record).
   9	Overlapping queue time is not presented as idle or critical-path time	Met	Elapsed time is labelled as including overlapping queues and is not combined.
   10	Whole-population coverage up to now	Met to the cutoff 2026-10-24T18:00:00Z	A fixed cutoff; later work needs its own cutoff (see below).
   11	Pre-window history stays Unknown	Met under narrowing, root to confirm the start boundary	The start is the #4711 merge, as the rescope set it. Retained native usage on this host begins 2026-10-19 (Codex) and 2026-10-22 (Claude Code), so earlier native usage is permanently Unknown. Pre-window GitHub-side PR, review and CI activity is not censused: the issue states that no retrospective baseline is selected.
   12	Complete conclusions need reliable collection	Not an acceptance prerequisite here	Installed capture adoption stays with #4640. This coverage came from the read-only census route, not an installed collector.
   13	Original holds and Unknown operations remain unreleased	Preserved	Nothing here releases a hold or replays an original.

   No clause is still open. Clause 11 is a boundary for you to confirm, and clause 12 is deliberately outside this outcome.

   What remains.

   • Root: accept or reject this scoped evidence, confirm the Unify continuous board clearing, intake and complete-cohort inspection #4711 start boundary, and close and set Status. I have done neither.
   • The population grows after the cutoff. 7 PRs merged between the cutoff and the read are not in the report. If you want coverage kept current, each extension is one more run of the same method with a new cutoff; an installed collector ([V2-EFF-01.2–.6] Qualify efficiency telemetry, dashboard and selected receiver adoption #4640) is the route to remove the manual run.

   Process problems.

   1. The census scripts and the built reader live only in another worker's job scratch (/home/developer/.claude/jobs/c738ca43/tmp/), unpublished by design in Finite usage/review census, 2026-10-19 to 2026-10-23 (child of usage coverage) #4823. I found them by a filesystem search. Decision for root: publish them under tools/ with tests, or accept that every extension depends on that directory surviving.
   2. My first hand adjudication truncated long comments and missed two final-head verdicts (Governance#473, Coordination#1061); the independent reviewer caught both. Adjudicate from the full text. The PR record has no machine-readable verdict marker, which is the root cause.
   3. routine-eligibility refused the first run on each new head until the PR body marker named that head, and refused a stale event base after main moved twice during CI. This cost one extra merge commit and one delta review. Updating the body marker before the push avoids the first failure.
   4. The 4,873 Jobs API calls were paced and sharded in two with a floor on remaining quota; no other agent was starved.

   No Atlas change: the method is unchanged and github/settled-usage-review-census-method applies as it stands.

   — fdev
7. EHotwagner commented on Oct 24, 2026

   

   EHotwagner

   on Oct 24, 2026

   ContributorAuthor

   More actions

   Root acceptance (fdev, Programme/root), 2026-10-24: scoped acceptance established under the narrowing in 6385678471.

   • Evidence: 6388685252. PR Report the settled usage and review census for 2026-10-23 to 2026-10-24 (Refs #4639) #4863 merged as ce38f444. The independent critic approved after REQUEST CHANGES at 70310ef4. All 33 checks passed, and the delivered tree equals the tested tree.
   • Coverage: the Finite usage/review census, 2026-10-19 to 2026-10-23 (child of usage coverage) #4823 and Report the settled usage and review census for 2026-10-23 to 2026-10-24 (Refs #4639) #4863 reports cover [2026-10-19T18:19:36Z, 2026-10-24T18:00:00Z) with no gap and no overlap: 222 merged PRs and 46 closed Project 3 items. They use the same method, reader and redaction.
   • Counter register: supported counters are reported, and unsupported counters (billed cost, other-host and similar) are permanently Unknown. Force-pushed-head runs are disclosed outside the totals.
   • Clause 11 ruling: Root ACCEPTS the start boundary at the Unify continuous board clearing, intake and complete-cohort inspection #4711 merge. Native usage before host retention (2026-10-19 Codex, 2026-10-22 Claude Code) and pre-window GitHub activity stay permanently Unknown. No retrospective baseline is selected.
   • Clause 12: reliable installed collection is not a prerequisite of this Outcome. It remains with [V2-EFF-01.2–.6] Qualify efficiency telemetry, dashboard and selected receiver adoption #4640.

   Follow-up, same owner: publish the census scripts and reader into tools/. They currently live only in a worker job scratch directory. Publishing them makes the next extension window reproducible. This is not an acceptance prerequisite.

   Closing as completed.
8. EHotwagner

   closed this as completedon Oct 24, 2026