ci: Turbo seeds vs the 10 GB Actions cache pool — measure the pool, then shrink or retire the per-sha seeds #25452

Filing gate: ② ruling-directed sibling card — batch #5 item 1, letter C (#25373 comment 6366946328, maintainer 「同意」 2026-10-23T14:13:20Z). Reader: the domain:devx execution seat, dispatch step. Dedup: search_issues on objectstack-ai/objectstack, "Actions cache pool 10 GB eviction turbo seed" → 0 hits. Session: cse_01LKHDniRQSWjRJyDtacwpEd.

One-sentence problem

Every push to main writes about 12.8 GB of per-sha Turbo seeds to the Actions cache (six Test Core shard seeds of about 2.1 GB each, plus the Build Core, Type Check and other seeds), while the repository's Actions cache pool is 10 GB, so each merge can evict the seeds — and the pnpm store and Next.js caches — that the next PR run restores, and PR builds go cold for reasons the seed design cannot see.

Evidence (file lines at objectstack commit 2b7c331)

• Per-sha seed keys, saved on every push event, including cancelled runs (if: always() && github.event_name == 'push'): ci.yml:1397-1402 (Test Core, per shard), ci.yml:2242-2247, ci.yml:2749-2754 (Build Core), lint.yml:5928-5930 (typecheck). Keys end in ${{ github.ref_name }}-${{ github.sha }} (ci.yml:606, :1786, :2087, :2320, :2529, :2974; lint.yml:5140, :5753), so no two merges share a key and nothing is overwritten — every merge adds a full set.
• Cancelled push run 39923516471 saved all six Test Core shard seeds at about 2.1 GB each (the ci: the push-to-main CI run is cancelled by the next merge 9 times in 13, so the Turbo cache seed that only that run saves goes stale and every PR shard and lint lane rebuilds cold (Build Core 51 s → 4m30s, shard closures Σ 20–31 runner-min per run) #25373 card thread's falsification, comment 6364570166).
• Plain actions/cache@v6 steps (restore and post-save on every event, PRs included) sit in 12 workflow files; ci.yml:2785-2798 saves the pnpm store and apps/docs/.next/cache on PR runs too. PR-scoped entries count toward the same pool.
• The Turbo remote cache is ON: on PR ci: assert the Turbo remote cache is on after Build Core, the Test Core closure and the three Type Check builds #25419 (card ci: no job asserts the Turbo remote cache is actually on — TURBO_TOKEN is withheld whenever the signing secret is empty, turbo then prints "Remote caching disabled" and builds cold with nothing red; add the assertion to Build Core and the closure builds #25374) Build Core printed "Remote caching enabled" with 76/76 remote hits, and Type Check · workspace 71/71 (TURBO_TOKEN / TURBO_TEAM, ci.yml:749-750). For the tasks the remote cache serves, the per-sha .turbo/cache seed in the Actions pool may be redundant weight.
• Pool usage is NOT measured yet. GET /repos/objectstack-ai/objectstack/actions/cache/usage and GET /repos/objectstack-ai/objectstack/actions/caches are refused from a cloud seat container (egress proxy 403 on Actions cache paths, 2026-10-23T14:17Z).

Governing text

• ci.yml:597-610 and lint.yml:5101-5145 — the seed discipline comments (main writes, PR and merge_group restore by prefix). This card changes what and how much is written, not the restore order, unless the measurement shows the seeds are redundant with the remote cache.
• AGENTS.md lines 371–372 are about the local worktree cache, not CI.
• Ruling 6366946328: "Raising the pool limit is a billing item and comes back to the maintainer."

Change (in order; the dev pins what the measurement supports)

1. Measure first. Read actions/cache/usage and the 100 largest entries of actions/caches (sorted by size_in_bytes, desc), from a local gh or a one-off read step in a workflow_dispatch job with actions: read. Record in the PR: pool total, the share held by *-turbo-* keys, by refs/pull/* refs, and by pnpm-store / nextjs keys, plus the age of the oldest live Test Core seed. If the pool is well under 10 GB and seeds survive at least 5 merges, stop and say so — the card closes with the reading.
2. Then shrink, cheapest first, choosing from: (a) drop the .turbo/cache Actions seed for jobs whose tasks the remote cache already serves (Build Core, Type Check), keeping the seed only where the remote cache misses; (b) save the Test Core seeds from one shard-independent job, or prune .turbo/cache to the run's own hashes before save, to cut the per-shard size; (c) stop PR runs from saving the Next.js cache (restore-only on PRs, as the Turbo seed already is).
3. Never raise the pool limit in this card. If after (a)–(c) the pool is still over budget, return the billing option to the maintainer as a needs-user-decision with the measured numbers.

Acceptance

• The PR records the before-measurement (pool total and the four shares above) and the after-measurement taken after ≥ 5 merges to main.
• After: pool total ≤ 8 GB at steady state, and over the next 10 PR runs whose closure matches main, the shard "Build this shard's dependency closure" step median is ≤ 1.7 min and Build Core's build step is ≤ 2.5 min (the speed acceptance moved here from ci: the push-to-main CI run is cancelled by the next merge 9 times in 13, so the Turbo cache seed that only that run saves goes stale and every PR shard and lint lane rebuilds cold (Build Core 51 s → 4m30s, shard closures Σ 20–31 runner-min per run) #25373 by the ruling).
• Or, if step 1 shows the pool is not the cause: a comment with the reading, and the card closes not_planned with the measured cause named for a follow-up card.

Risks

• Dropping a seed the remote cache does not actually cover makes those jobs colder; the dev keeps the remote-cache hit counts from the assert-turbo-remote-cache lines (ci: assert the Turbo remote cache is on after Build Core, the Test Core closure and the three Type Check builds #25419, once landed) before and after.
• Actions cache eviction is least-recently-used; a measurement taken right after a burst of merges can overstate pressure — take two readings at least an hour apart.

Work estimate

Lane domain:devx · 1 PR (.github/workflows/ci.yml, .github/workflows/lint.yml; possibly .github/workflows/*.yml for the Next.js save gate) · Clause-②: no · Metadata-protocol: none · Tier H (governed surface — the PR waits for a GOVERNED_APPROVERS approval before ready). Size M. Serial: lands after #25373's PR (same ci.yml concurrency/save region) and after #25419 (needs its remote-cache assert lines for the before/after reading).

Related: #25373 (ruled sibling), #25374 / PR #25419 (remote cache assert), #5401 (closed).

Maintainer directive (verbatim, untranslated)

同意

— the maintainer, 2026-10-23T14:13:20Z, on director batch #5 item 1 (letter C = land the staged concurrency change as a coverage fix and file a separate card for the cache pool).

Generated by Claude Code