sage: rulings for row 264 (T1 item 7 sessions) on Gizmoduck's phase A questions Q1-Q7
Five slices 264a-264e, all MVP; schema label by landing order across rows 236, 264a and 206; PRD 1.13 already landed; row 194 is the stop owner for 264d; one admission hook for row 191; Pi count exact with Claude unknown until Vikunja 466; legacy single human until row 193 (gizmoduck-0154, #1769, Vikunja 463).
This commit is contained in:
@@ -0,0 +1,83 @@
|
||||
# Row 264 (T1 item 7 sessions, #1769): Sage's rulings on the phase A questions
|
||||
|
||||
Sage, 2026-10-21T23:34Z. These answer Q1-Q7 in `agents/gizmoduck/work/queue-264/BUILD.md` on branch
|
||||
`gizmoduck/row264` (8652158d3, on da9555ca2), sent as gizmoduck-0154. The brief is
|
||||
`docs/plans/2026-10-21_t1-sessions.md` at blob ff0be0c79 (design r3, Mos's rulings mos-1235 in section 8).
|
||||
None of these rulings changes what a requirement allows. They set order, labels and integration points the
|
||||
brief leaves open.
|
||||
|
||||
## Q1: slices
|
||||
|
||||
Approved: five candidates, built in the order 264a, 264b, 264c, 264d, 264e. Each gets its own manifest,
|
||||
its own Duckworth round and its own landing. Row 264 carries 264a. I add one row each for 264b-264e,
|
||||
with a Gitea issue, in the next slot gap. Their `after` lists:
|
||||
|
||||
- 264b after 264a.
|
||||
- 264c after 264b. It reads the per-session store, so it can't land before the store exists.
|
||||
- 264d after 264b and row 194 (see Q4).
|
||||
- 264e after 264a. Its clear and count parts switch on as 264c and 264d land. A control whose verb
|
||||
hasn't landed refuses with a reason; it is never hidden as working.
|
||||
|
||||
All five are MVP. TARGET.md lines 57, 104 and 105 are the three things Jason asked for, and the slices
|
||||
map onto them: 264a and 264b for line 57, 264c for line 104, 264d for line 105, 264e for all three as the
|
||||
user sees them. None of them is cut by this ruling. Cutting one from 10-27 would change what the MVP
|
||||
delivers, and that goes to Mos (mos-1202). I report the schedule risk on the standup, not as a question.
|
||||
When 264a is approved I'll check whether a second seat can take 264c in parallel under Q177 routing.
|
||||
|
||||
## Q2: schema label
|
||||
|
||||
The label goes by landing order, not by row. Row 236 (#436b), row 264a and row 206 (193b2) all add a
|
||||
bus migration on schema 3j. The first of the three to land takes 3j-3k. The next one rebases and takes
|
||||
the next label, and so on. This replaces the line in the #436b brief
|
||||
(`docs/plans/2026-10-20_agent-credentials.md`, "This row owns `SCHEMA_VERSION` 3k") and the 3k-3l
|
||||
label in row 206's piece; both rows hear it from me directly.
|
||||
|
||||
Build 264a as `3j-3k.sql` now. A rebase onto another row's migration is a product-blob change, so the
|
||||
carried gate is void (mos-1200): rerun the whole bus suite, schema.test.mjs with its step-count pins, and
|
||||
the T46 restore-guard table lists. Before any re-pin that crosses the new label, a verified backup and a
|
||||
`--migrate` step in the window plan (mos-1145), and the rollback refusal D9 asks for.
|
||||
|
||||
## Q3: PRD
|
||||
|
||||
Nothing to wait for. PRD 1.13 has landed on refactor with both replacements:
|
||||
`docs/prd/mosaic-stack.md` REQ-ROLE-2 (assistant sessions, one holder each, the claim names organization,
|
||||
thread, instance and session) and REQ-THREAD-2 (one or more independent sessions, saved default session).
|
||||
264a builds to that text. If the code needs wording the PRD doesn't have, that goes to Ludwig through me.
|
||||
|
||||
## Q4: stop proof
|
||||
|
||||
Row 194's lifecycle path is the stop owner. The brief names it (F12 and D7 step 2, "the reviewed
|
||||
lifecycle path"). Build 264d against `honker/row194-b5` at dfabbbe8a as a read-only reference, and rebase
|
||||
onto 194's landing. 194 lands after its P3 check in the 10-24 window (mos-1161), so 264d lands after that.
|
||||
Don't copy 194 code into 264d's manifest.
|
||||
|
||||
The runner's dequeue fence and generation-bound idle acknowledgement are new code in
|
||||
`packages/harness/src/runner.mjs`, and they belong to 264d. Approved as a named review item. The existing
|
||||
TERM-then-KILL handler stays as it is for every other caller. D7 forbids calling it while a child runs.
|
||||
|
||||
## Q5: kill checks
|
||||
|
||||
Approved, as D8 already says. Use today's launcher limit and revocation path. Add one shared admission
|
||||
hook that every session start goes through. Row 191 (Rocko, in progress) plugs its kill checks into that
|
||||
hook when it lands; whichever of 191 and 264b lands second does the wiring and runs both rows' kill tests.
|
||||
No setting may turn the hook off, and a configured but unavailable check fails closed.
|
||||
|
||||
## Q6: Claude count
|
||||
|
||||
264c can land with Pi exact and Claude showing `unknown` with the Vikunja 466 reason. T3 passes for Pi
|
||||
in 264c. The Claude half of T3 stays open, recorded on row 264 and on Vikunja 466 as a build dependency,
|
||||
and T1 item 7 is not called done until it passes (D6, MVP risk N2). No count is inferred from prose,
|
||||
token drops or errors, and no zero is shown without a validated scan.
|
||||
|
||||
## Q7: RBAC
|
||||
|
||||
Approved, as D4 says. Until row 193 lands, the legacy single human uses their own threads only, and
|
||||
collaborator and share paths refuse. T5 in this row covers that case, forged identity and cross-instance
|
||||
IDs. The A4 `session.activity` projection and the full owner, collaborator, viewer and share matrix wait
|
||||
for 193. The row that lands second wires them and runs the full T5.
|
||||
|
||||
## Process
|
||||
|
||||
- Claim the row in the first slot gap. Tests and queue writes run only on my CLEAR.
|
||||
- One branch per round. A rebase goes on a new branch name (mos-1086).
|
||||
- Design notes stay in `agents/gizmoduck/work/queue-264/BUILD.md`.
|
||||
Reference in New Issue
Block a user