Retire PR #416's unrun compaction-window pilot - #661

Scope

• Retire draft PR Compaction-window A/B/C preregistration draft (not frozen, not run) #416 with a dated record preserving its user choice, protocol, historical observations, review/test claims and independently measured file digests. Repoint the three living community-practice rows to the record's current-client confirmatory reopen conditions.
• Base commit: 4ced2923063db6a6dcafa9f25af5ee05a4153c75 (origin/main, merged into this branch by d810f0d6f in the review round). The builder worked on the prepared base 9b0b8d6d25f9e3fb8f71770500e774170423315e.
• Lane: lane:foundation
• Owned paths touched: docs/decisions/2026-10-03-compaction-window-ab-retirement.md, docs/community-native-practice.md, manifests/evidence.json.

The builder used the prepared base. During its run origin/main advanced to 4ced2923063db6a6dcafa9f25af5ee05a4153c75 (#636: AGENTS.md and its registry row). The review round merged it through the docs/lanes.md hot-file protocol (merge path, no force push) and refreshed the registry in the last commit; see Review round.

Builder and roles. The builder was GPT-6.1 Sol: codex exec (Codex CLI 0.159.3) through the OmniRoute provider, with model cx/gpt-6.1-sol, effort max and a workspace-write sandbox, per the builder run's own Codex session metadata. Its session started at 2026-10-03T10:27:03Z, its first stamped precondition read was at 10:27:52Z and its last session event was at 11:56:41Z. Both start times are after the custody notice's two-hour objection window closed at 10:21:56Z (notice posted 08:21:56Z). The coordinator's publisher stage made commits 71080d449 and c905678f6, pushed them and opened this PR. By its own returned notes it ran on Claude Sonnet 5.5 and used the task-prescribed Co-Authored-By: Claude Opus 5.5 trailer, so that trailer does not name the model that made those two commits. The review-round commits 0cad642c1, d810f0d6f and f7aab0151 were made by Claude Opus 5.5.

SOTA sources

• anthropics/claude-code, tag v2.1.288, commit 1c229fcd1e1e4e452e29a8f116b45fe4cfe2c528, CHANGELOG.md entries 2.1.284–2.1.288; read 2026-10-03 through gh api. Installed version/help corroborates client 2.1.288 and the supported controls.
• Anthropic's model configuration, default thresholds and environment-variable reference, read 2026-10-03. They retain the approximately 967K native 1M Anthropic-API default; the per-model save change is version drift rather than an earlier-window recommendation for that route.
• Pinned #416 source, including the four blueprints/compaction-window-ab/ files, its historical test file, evidence manifest and blind-checkout source. Each complete original blob was independently hashed and counted; all six file registrations in its manifest diff match.
• Community sweep at the retirement base, the upstream-harness rule, Gate A's re-aim and the native-companion architecture; exact immutable citations are in the record.
• #488 citation source, blueprints/gate-b-gpt6-route/PREREGISTRATION.md, preregistration.json and tests/test_gate_b_gpt6_route_preregistration.py; all twelve citation sites remain resolvable through Compaction-window A/B/C preregistration draft (not frozen, not run) #416's preserved head.

Evidence-class table

Local commands run

Builder sandbox output (pre-commit, prepared base 9b0b8d6d2)

The builder's sandbox could not write git state, so this block shows its working tree before any commit; the publisher made the commits afterwards. Python commands used RTK's proxy to preserve complete returned output. The six consistency checks and mandatory registry methods below are the final-content repeats. The whole-suite run preceded the final pointer-only correction; its failed result is preserved. No trial, model comparison or new host acceptance is claimed.

$ rtk proxy python3 scripts/validate.py
exit 0; passed; 69 components, 9420 hashed files, 4 profiles, 187 receipts

$ rtk proxy python3 scripts/host_receipts.py validate
exit 0; passed; 182 receipts

$ rtk proxy python3 scripts/validate_convergence.py --all-recorded --root . --json
exit 0; valid true; all 26 recorded contracts valid, no errors

$ rtk proxy python3 scripts/component_matrix.py --check
exit 0; checked; 32 rows

$ rtk proxy python3 scripts/new_host_grand_list.py --check
exit 0; passed; 32 layers, 66 winners

$ rtk proxy python3 scripts/build_ecosystem.py --check
exit 0; passed; existing architecture-pin drift remains reported

$ rtk proxy python3 -m unittest
exit 1; Ran 9919 tests in 2222.786s
FAILED (failures=484, errors=194, skipped=863)

$ rtk proxy python3 -m unittest tests.test_osv_lockfile_coverage.LockfileInventoryTests.test_every_tracked_lockfile_and_manifest_is_listed tests.test_blind_checkout.RepositoryClassificationTests.test_every_blueprint_value_under_a_label_key_is_classified tests.test_workflow_security_coverage.NewWorkflowSecurityCoverageTests.test_all_published_workflows_are_listed_and_covered -v
exit 0; Ran 3 tests in 0.222s; OK; no skips

$ rtk proxy python3 -m unittest tests.test_adaptive_market_research.MarketResearchCredentialFilePermissions.test_0700_and_0755_owned_parent_directories_still_pass tests.test_windows_terminal_defaults.OverlayTests.test_the_installed_client_knows_no_notification_type_without_a_decision tests.test_pre_push_gate.PrePushGateTests.test_unset_tmpdir_defaults_to_var_tmp -v
exit 1; Ran 3 tests in 3.832s
FAILED (failures=2, errors=2)

$ rtk git diff --check
exit 0

$ rtk git diff --name-only origin/main...HEAD
exit 0; empty; no builder commit exists under the addendum

$ rtk git status --short
exit 0
 M docs/community-native-practice.md
 M manifests/evidence.json
?? docs/decisions/2026-10-03-compaction-window-ab-retirement.md

At builder time, the hot-file protocol used main's complete manifest at the prepared base. After the final minimal pointer correction, the supported register_file command, component-matrix --write, new-host-grand-list --write and validate all exited 0 in order. Only the community document's registry entry changed; no generated committed report changed. Validate did not require registration of the dated record. Current-main integration and another hot-file refresh were still needed at that point; the publisher published first, and the review round did both (last block below).

The full-suite failures include temporary-path guards, filesystem ownership/writability limits and an installed-client notification-table mismatch (auth_storage_failure). Existing targeted tests reproduced representative failures; all implementation/test files inspected are unchanged from the prepared base. The builder neither skipped those checks nor repaired unrelated paths. These observations do not classify all 678 failure/error reports, and no clean-baseline whole-suite replay is claimed.

Final exit-time live #416 checks at 2026-10-03T11:55:15Z returned OPEN, CONFLICTING, the unchanged pinned head and preserved branch, zero reviews and only the custody notice. All 273 returned #608 comments were examined; only the existing custody census mentions #416, with no competing lane claim. All 17 main docs/scripts #416 lines are covered in record section j. No contract stop condition was hit.

Publisher, post-commit (head c905678f6, merge base 9b0b8d6d2)

Re-run in the review round against the published head, with shell exit codes:

$ git show --stat 71080d449
exit 0; docs/community-native-practice.md (6 +-) and docs/decisions/2026-10-03-compaction-window-ab-retirement.md (538 +)

$ git show --stat c905678f6
exit 0; manifests/evidence.json only (2 +, 2 -)

$ git diff --name-only origin/main...c905678f6
exit 0
docs/community-native-practice.md
docs/decisions/2026-10-03-compaction-window-ab-retirement.md
manifests/evidence.json

$ git diff --check origin/main...c905678f6
exit 0

The publisher's own runs on c905678f6: python3 scripts/validate.py exit 0; the three registry methods exit 0 (Ran 3 tests in 0.207s; OK); git push exit 0 with the pre-push gate passing.

Review round (merged head f7aab0151)

$ git ls-remote origin refs/pull/416/head refs/heads/claude/compaction-window-ab-prereg-20260927
exit 0; both 452f7b14715ae2a038c1d4a98e292b5de7812c87 (2026-10-03T13:08:54Z)

$ gh pr view 416 --json state,isDraft,headRefOid,updatedAt,comments,reviews
exit 0; OPEN, draft, head 452f7b14, 1 comment (the custody notice), 0 reviews; no stop condition

$ git merge --no-ff --no-commit origin/main   # 4ced29230
exit 0; clean auto-merge of AGENTS.md and manifests/evidence.json

$ git checkout origin/main -- manifests/evidence.json
exit 0; git diff --quiet origin/main -- manifests/evidence.json then exited 0 (main's copy exactly); merge committed as d810f0d6f

$ register_file one-liner from docs/lanes.md for docs/community-native-practice.md
exit 0; one row, 6a867561…/56050 -> 6e7ebbe6…/56411, matching sha256sum and wc -c of the file

$ python3 scripts/component_matrix.py --write
exit 0; written; 32 rows; no committed file changed

$ python3 scripts/new_host_grand_list.py --write
exit 0; written; 32 layers, 66 winners; no committed file changed

$ python3 scripts/validate.py
exit 0; passed; 69 components, 9420 hashed files, 4 profiles, 187 receipts

$ python3 scripts/validate_convergence.py --all-recorded --root . --json
exit 0; 26 records, valid true

$ python3 scripts/host_receipts.py validate
exit 0; passed; 182 receipts

$ python3 scripts/component_matrix.py --check
exit 0; checked; 32 rows

$ python3 scripts/new_host_grand_list.py --check
exit 0; passed; 32 layers, 66 winners

$ python3 scripts/build_ecosystem.py --check
exit 0; passed; existing architecture-pin drift remains reported

$ python3 -m unittest tests.test_osv_lockfile_coverage.LockfileInventoryTests.test_every_tracked_lockfile_and_manifest_is_listed tests.test_blind_checkout.RepositoryClassificationTests.test_every_blueprint_value_under_a_label_key_is_classified tests.test_workflow_security_coverage.NewWorkflowSecurityCoverageTests.test_all_published_workflows_are_listed_and_covered -v
exit 0; Ran 3 tests in 0.204s; OK; no skips

$ git diff --check origin/main...HEAD
exit 0

$ git diff --name-only origin/main...HEAD
exit 0
docs/community-native-practice.md
docs/decisions/2026-10-03-compaction-window-ab-retirement.md
manifests/evidence.json

$ git push origin HEAD:foundation/pr416-compaction-window-retirement-20261003
exit 0; c905678f6..f7aab0151, fast-forward; pre-push gate Ran 3 tests, OK

Every Python command ran with TMPDIR set to a session directory outside the repository, and each exit code was read from the shell, not through a pipe. A grep of git diff origin/main...HEAD (74,005 bytes) for home paths and the user name exited 1: no matches.

Decision record

docs/decisions/2026-10-03-compaction-window-ab-retirement.md

Retirement preserves the native default and the user's “test before changing” decision. Keeping/re-freezing the old descriptive pilot does not supply confirmation or close the launch gates. A reviewed, frozen, current-client confirmatory Harbor/Inspect comparison with quality non-inferiority and a preregistered cost criterion, or a primary Anthropic default/recommendation change, reopens the question.

Host evidence

No new or changed host receipt. Historical observed/relayed labels remain historical. The independent read-only Opus/max fact review ran on c905678f6 after publication and returned repair; one repair round followed (see Review round), and an Opus re-check of its diff is pending. The sandboxed builder was prohibited from invoking a reviewer.

Review round

The independent read-only Opus/max review of c905678f6 returned repair with 12 findings: 4 should-fix and 8 minor, three of them verification gaps. One repair round on Claude Opus 5.5 applied all 12; none was rejected. Each changed claim was re-read at its source in this round.

Carry-over: git diff --stat c905678f6 f7aab0151 lists AGENTS.md (main's #636), the record (20 insertions, 16 deletions) and manifests/evidence.json (main's AGENTS.md row only; the community row equals c905678f6's). On the reviewed files, git diff c905678f6 f7aab0151 -- docs/decisions/2026-10-03-compaction-window-ab-retirement.md docs/community-native-practice.md shows only the record repairs above. An Opus re-check of that diff is pending.

Checklist

• New/changed GitHub Actions are pinned to a full commit SHA with a version comment; no actions changed.
• New/changed workflows declare top-level permissions: contents: read or a justified narrower addition; no workflows changed.
• No secrets are printed, logged or committed; no new required secret is added.
• No new paid hosting, subscription or billing surface is introduced.
• Peer-owned files and worktrees are preserved.
• Current-main integration and final hot-file refresh complete: merge d810f0d6f of 4ced29230, registry in f7aab0151.
• Whole-suite acceptance on the final head and the Opus re-check of the repair round. CI validate passed the whole suite on c905678f6; the f7aab0151 run and the re-check are pending. The builder's sandbox whole-suite result was FAILED.
• Required hosted checks pass on the final reviewed head.

🤖 Generated with Claude Code