9.5 KiB
Re-pin 7247d80c7 to 86f95bb27 (rows 251, 244, 254, 202, 273, PRD 1.11-1.14), no migration
Sage, 2026-10-22, for mos-1253 under DECISIONS Q173 and mos-1246. The step layout is
repin-1024-draft1.md, with the copy changed and --migrate dropped. Where a step is unchanged I name it
and say only what differs. Mos checks this file's sha256 before P0, and any change after that means a new
check (mos-0849/0850). This file stays uncommitted until row 191's window ends, because refactor doesn't
move while a rebased candidate's gate runs.
Who, and what nobody does
- Darkwing operates if free. Otherwise I operate. Mos verifies each step's output before the next one, runs P3, and gives the W5 go. Row 191's full window keeps first claim on the slot (mos-1225). Steps that start no test and no stub run alongside it: P1, W0, P4, P5, P6, B0, W1-W4. P2 waits for a gap or CLEAR, because it starts node stubs that a busy check counts.
- Jason has no step. Nothing is pruned. Nobody acts as Jason: no
--by jason, no human shell or tmux pane, every bus read and write carries--tester, andbus.testerstays as it is. If it isn'ttrue, stop at P1. - No pi (#383). No Save in the human Console, no
mosaic name. - Outside reach doesn't change. No
--send, no--sso, no proof pair. Both go on together in the phase 2 window, which still waits for Jason's start (Vikunja 412). The tunnel forwards 7332 only. - dragon-lin only. .53 and .54 take nothing (mos-1198, mos-1167).
Names
- OLD, live:
7247d80c7af7ffa70fc61557406bf491cdf3bfc7, schema 3j, pointerpm: true, activation20261021T215926Z-7247d80c7af7. Console flags at W5 last time:--port 7332 --public-host mosaic.woltje.com --auth-header X-authentik-username --allow-user jason.woltje --tester-socket. P1 re-reads them fromstatus. I don't copy them from this file. - NEW:
86f95bb27f6cb237bf9fa9ba53320bace6c4de0b, 59 commits past OLD (git rev-list --count). The last code landing in it is0bbf0f6d1(row 273). The two commits after that are docs and queue files. WIN=agents/darkwing/work/window-repin-86f95bb27/if Darkwing operates, elseagents/sage/work/window-repin-86f95bb27/. The scripts adapt last window's, inagents/sage/work/window-repin-7247d80c7/, with NEW and OLD changed and every--migrateremoved.
What NEW carries over OLD
Non-test paths changed from OLD to NEW (git diff --name-only, tests and READMEs left out):
| Row | Commit | Paths | Live effect to check |
|---|---|---|---|
| 251, 244 | d37210288, e69260a3c |
scripts/test-release.sh, test-task.sh, scripts/testing/* | Gate tooling. Not in the live path. |
| PRD 1.11-1.14 | ea0b2a16d..db1e7732a | packages/tasks/requirements.json | W9: the pm's requirements read gives 1.14 with 85 ids. |
| 254 | 331597709 |
packages/bus/src/human-cli.mjs (library.sync client timeout 130 s), packages/webui/src/bus.mjs (six library-* STATUS codes) |
No library entry is live, so neither path runs. |
| 202 | ddfb29af7 |
packages/cli/src/access.mjs, cli.mjs, scripts/mosaic | mosaic access list/grant/revoke. Nobody runs them in this window (Q188, mos-1141). |
| 273 | 7991662dc..0bbf0f6d1 | scripts/stack-release.sh | --send and --no-sso. With neither given, the plan keeps the pointer's Console flags. W3 checks that. |
Nothing else in packages/bus/src or packages/webui/src changed (git diff --quiet with those two files left
out). The four service scripts and scripts/stack-copy.sh are unchanged from OLD (git diff --quiet 7247d80c7 86f95bb27 -- on them). package-lock.json is unchanged.
Schema and backups
SCHEMA_VERSIONis 3j at both commits (packages/bus/src/store.mjs:17). No migration step and no schema.sql change in the range. W3's plan must show no migration line. If it asks for--migrate, stop.- B0 is the backup taken first. It's an online copy of the live store through sqlite's backup API, read only
on the source, as in last window's W12. It is laid out the way
rehearse --backup PATHreads it (stack-release.sh:2232-2235):$DATA/backups/b0-<STAMP>/bus/bus.sqlite,SHA256SUMSwritten from insidebus/,SCHEMAreading 3j, andCHECKholding the integrity_check result, the event count and maxSeq. The operator sends Mos the path and the SHA256SUMS sha256 before P3. - W5 takes the stopped-store backup every activate takes, plus the rollback set. That one is the rollback point.
- The grant look-alike count from draft 1 is dropped. It guarded the 3i-3j migration, and there is none.
- Rollback is a pointer flip. OLD and NEW share 3j, so
rollback --planthen the real run, with no--restore. Any restore goes to Mos and ends the window (mos-1226, mos-1253 item 5).
Before the window (nothing stopped)
| # | Who | Step |
|---|---|---|
| P0 | Mos | The sha check of NEW and of this file. |
| P1 | operator | status: pointer, activation, 3j, Console flags, unit flags match. Business file sha256 and business-check.mjs LOAD_OK for OLD now and for NEW after W0. ls -ld $DATA/console is 0700 and tester.sock is 0600 (row 196 made them at the last W5). bus.tester is true. |
| W0 | operator | scripts/stack-copy.sh copy $NEW. The copy has tools/unslop-hook and packages/library/bin/askpass with NEW's mode. |
| P2 | operator, in a gap | The two stub runs from last window's p2.sh against NEW's copy, then render both units from REPO's service scripts and diff them against the live units. The only difference allowed is the copy path. |
| B0 | operator | The backup above. Path and SHA256SUMS sha256 to Mos. |
| P3 | Mos | stack-release.sh rehearse $NEW mosaic-stack --backup <B0 path> --port <scratch>, no --migrate. It must show no migration, the interpreter preflight passing, one receipt line, no library: line, and one pm reply. Mos also reads healthz, /api/config and the two tester-http reads on NEW's Console with the pointer's flags, as in mos-1107. |
| P4 | operator | The tunnel read. |
| P5 | operator | The four service scripts held at NEW, and empty porcelain on them. |
| P6 | operator | The 7332 matrix (a)-(g) on OLD as $WIN/matrix-old.txt, plus the two tester-http reads, for W10. |
The window
| # | Who | Step |
|---|---|---|
| W0b | operator | The business file copy. |
| W1 | operator | Idle check, start stamp from date -u, S1 with $SNAP --held. Open pm tasks: the pm finishes or parks them, or the operator records which piece crosses the window. |
| W2 | operator | status: pointer on OLD, pm: true, 3j, unit flags match. |
| W3 | operator | stack-release.sh activate $NEW mosaic-stack --pm --tester-socket --plan. It must show OLD to NEW, no migration, the backup line, and a Console line with OLD's flags exactly: --tester-socket, no --send, no --sso, no proof pair, no drop line. Any other difference stops here. |
| W4 | operator | The unit reads, as last window. W3 and W4 go to Mos for the W5 go. |
| W5 | operator, on Mos's go | The same command without --plan. Exit 0, result ok, every gate check passing, pm_launched included. Record the activation id, backup path, STAMP and the backup SHA256SUMS sha256. Then status. |
| W9 | operator, after Mos verifies W5 | Last window's W9 list: no Sign-in: line in the webui journal since the W5 stamp (with UTC), /proc/<Console pid>/cwd names NEW's copy, tester.sock 0600 in a 0700 console/, one new peercred-starts.jsonl line, no library: line in the bus journal. The pm's requirements read gives 1.14 with 85 ids. If the pm isn't relaunched by the restart (mos-1230), the read moves to W12, as last time. |
| W10 | operator | The matrix again as $WIN/matrix-new.txt, identical to matrix-old.txt apart from named lines. The two tester-http reads match P6's. No listener on 7333 (ss -ltn). P4 again. A Console POST on the main port with no signed-in session is refused (the mos-1195 baseline). |
| W11 | operator | STATE report in $WIN/STATE.md, posted on P19 386, 412 and 435, and on the Gitea issues for 202 (#1700), 254 (#1759) and 273 (#1778). |
| W12 | operator, after W11 | Row 66's live check as last time: backup first, one Claude Code managed session with a known tell, the RESULT note, the next-turn notice and the unslop.jsonl counts. Pi isn't run (#383). |
After W12 I write draft 3 of the phase 2 plan naming 86f95bb27. That window still waits for Jason.
Steps skipped, and why
Mos allowed skipping a step that only repeats a test row 273's or row 202's gate already ran on the same
code. The blob ids at the gated commit and at NEW are equal for every one of those files (git rev-parse):
scripts/stack-release.sh 7c7676c71 at 0bbf0f6d1 and NEW, and access.mjs a0a7c9364, cli.mjs b14b8fd79,
scripts/mosaic 91c79ab4b and human-cli.mjs 27b67fbf7 at ddfb29af7 and NEW.
- No separate run of the
--send/--no-ssowalk or the rollback flag carry. Darkwing's gate on0bbf0f6d1ran stack-release.test.mjs 218/218 (T53, T57, T58 c0/c1) and test-release.sh rc 0. W3's live plan line is the only check left. - No live
mosaic accessrun. Row 202's gate ran access.test.mjs onddfb29af7. A live grant would also break Q188. - The grant look-alike count, for the reason under Schema. It's a scope drop, not a gate repeat.
Nothing else is skipped. P2 and P3 aren't gate repeats: they check NEW's copy as built on dragon-lin.
Stop conditions and rollback
As in draft 1. The first failure stops the step and goes to Mos. A second failure, or any restore, ends
the window (mos-0823, mos-1253 item 5). Before W5 there's nothing to undo. After W5, rollback is
rollback --plan then the real run, which flips the pointer to OLD with the same Console flags. Keep every
copy. Prune nothing.