Repository navigation
Description
(ai: claude-opus-5-5)
@ranxianglei — this lists how our open PRs fit together, so they can be reviewed as one piece of work rather than as unrelated fixes.
Goal. Run Billy on pure SQLite (the default since #6480) for a large, long-lived store: no lost sessions on migration, startup within the readiness limit, and no whole-row rewrite on every turn.
Test case. A copy of a real 0.1.200 legacy install: 8,791 sessions, 10.7 GB index.db after migration, on a host with 4 CPUs and 5 GiB RAM. With all PRs below combined on master, the rehearsal passed: 8,791 of 8,791 sessions migrated (0 failed, 0 drifted) in 94 min, the relay was ready in 15.8 s, a restore from an online backup worked, and the old version still started on the untouched source files.
1. Migration without data loss
- [PR] fix(storage): keep CCR originals safe during bulk migration (#5599) #6773 — guard CCR originals during migration and saves (approved by the review bot, waiting for merge)
- SQLite-only durability fixes: stacked on [PR] fix(storage): keep CCR originals safe during bulk migration (#5599) #6773, opened after it merges
2. Startup on a large store
- [PR] perf(storage): start a large SQLite store without rewriting every session row #6850 — on an earlier 8.9 GB copy, master was killed by its own event-loop watchdog after 300 s; with the PR it was ready in 10.6 s
- [PR] perf(storage): complete block views once per ingest plan #6914 — linear block planning (quadratic on long sessions)
- Stop hangs after the session flush while the startup background quick_check runs on a large store #6987 — stopping the relay while the background
quick_checkruns waits until SIGKILL (fix: [PR] fix(storage): stop a running background quick_check on exit instead of waiting for it #7260, approved by the review bot)
3. Per-turn cost
- [PR] perf(persist): keep jsonWriter=off index builds out of the request path #6986 — answer persistence with the JSON writer off (needs your decision on the timing trade-off)
- [Root Cause] storage-persist (2 items) #6792 / Migration design review: legacy JSON to SQLite-only, wire/cache contract, duplicate routes and durability fixes #5599 — large columns rewritten with the whole row on every save. Needs your design decision; a prototype and measurements are in [Root Cause] storage-persist (2 items) #6792.
4. State after restart or crash
- [PR] fix(persist): keep k̂ model provenance and the emergency-shrink marker across restarts #6775 — restart-state fields (landed via [PR] wave-11 batch: 7 community fixes (opencode provider, preflight epoch, output budget, +4) — #6797/#6818 excluded #6827)
- [PR] fix(storage): do not serve migrated block placeholders as content snapshots #6943 — block-content placeholder (approved by the review bot)
The OpenCode incremental lane (#6818, #7307, #7308) is tracked separately in #7310.
Suggested order: 1 → 2 → 4 → 3. Groups 1 and 2 decide whether a large store can move to SQLite at all; group 3 decides what it costs per turn.
Decisions only you can make: the row-rewrite design (#6792) and the timing trade-off in #6986.
I'll keep this list current as PRs land.
Activity
ranxianglei commentedon Nov 19, 2026
(ai: vllm-qwen/qwen3.8-27b)
查重: 不重复. This is a distinct tracking/coordination doc. Closest siblings are #7310 (OpenCode incremental lane — explicitly separated in the body) and #7274 (a specific event-loop bug, not a tracking doc). No other issue tracks this same set of PRs.
Labels applied:
tracking·scope/storage·needs-decision·p2(significant but scoped to large stores; the fix is in progress with PRs open and green).Verified live state (checked against GitHub + the mirror CI cache, 2026-11-19):
The list is accurate — every ref matches its live state. No material movement since the body was written; the only in-flight item is #6943's 12 still-running checks.
The two decisions that are yours (@ranxianglei):
Both are design/product decisions — I'm flagging them, not deciding them. The suggested order (1 → 2 → 4 → 3) holds: groups 1+2 gate whether a large store can move to SQLite at all; group 3 is what it costs per turn.
Nothing for me to fix here — this is a coordination doc and it's consistent with live state. I'll leave the two decisions and the merge order to you and @mariopasquali.