[finding] `--union-into` pushes an ABSOLUTE `packages.items[].path` while `turbo ls` emits repo-relative ones, so the document it hands the shard partitioner is heterogeneous

[os-steve]

Found while auditing every field of the turbo ls document for #10046 (that card fixed the sibling defect in the same writer, packages.count). Filed unassigned. Inert today, and unlike the count case the correct fix is a judgment call rather than mechanical — which is why it is a separate card and was deliberately not folded into #10046's PR.

The divergence

turbo ls --output=json emits packages.items[].path as a repo-relative path. Measured on turbo 2.10.10 against this repo, all 77 packages:

$ pnpm exec turbo ls --output=json | node -e "...
items with ABSOLUTE path: 0 of 77
sample paths: packages/apps/account, packages/cli, packages/client

scripts/check-cross-package-test-inputs.mjs's unionInto() appends an absolute one:

items.push({ name, path: join(REPO_ROOT, dir) });

Measured with a changed-file list of exactly scripts/sync-template-versions.mjs:

{"name":"@objectstack/spec","path":"/home/user/objectstack-issue-10046/packages/spec"}
{"name":"create-objectstack","path":"/home/user/objectstack-issue-10046/packages/create-objectstack"}

So a single document can carry both conventions in one array — turbo's entries relative, the unioned entries absolute.

Why it is inert today

The only consumer is scripts/partition-test-shards.mjs, which does readdirSync(it.path) via countTestFiles(). That resolves a relative path against process.cwd(), and CI runs the script from the repo root, so both conventions land on the same directory. Measured, same package, same tree:

relative (turbo's convention): shard 1/1: 1/1 packages, weight 414
absolute (union's convention): shard 1/1: 1/1 packages, weight 414

Why it is still worth closing

countTestFiles() returns 0 for a path it cannot read — a deliberate, documented choice (// package path missing locally -- weight 0, still assigned). That is what keeps this quiet: a path resolved against the wrong base does not throw, it produces weight 0, and the LPT partitioner absorbs a zero weight without complaint. Measured — same relative-path document, run from a different cwd:

relative, cwd = /: shard 1/1: 1/1 packages, weight 0

The package is still assigned, still tested; only the load balancing silently degrades. So the failure mode of this class is not a red step, it is a shard matrix that quietly stops balancing.

A future consumer is the sharper risk: anything reaching for the obvious join(REPO_ROOT, it.path) — correct for every entry turbo wrote — produces a garbage path for exactly the unioned entries, which are the cross-package scans #7802 exists to keep running.

Why this is not a one-line fix

Both normalisations are defensible and they trade against each other:

• Normalise the union to relative — the document becomes a faithful turbo ls payload, which is what partition-test-shards.mjs's assert-the-shape-loudly posture wants. But relative paths are cwd-dependent, and the consumer's own weight function fails silently (weight 0, above) when the cwd assumption breaks.
• Normalise everything to absolute — robust against cwd, but the document is then no longer what turbo ls emits, and it would need rewriting turbo's own entries, i.e. mutating fields the union step does not author.

There is also a third option: leave the data alone and have the consumer resolve it.path against the repo root explicitly, which fixes the cwd fragility for both conventions and makes the heterogeneity harmless rather than merely inert.

⚠️ Whichever is chosen, note the entry-gate consequence: it.path feeds countTestFiles() feeds the package weight feeds shard placement. A change here can move packages between shards, so it is not a comment/data-shape change — which is precisely why #10046's PR left it alone (that card's claim declared no change to which packages get sharded).

Refs: #10046 (sibling defect in the same writer, and where this was measured), #7802 (the failure family the union step exists for).

---

Generated by Claude Code

[os-elon]

Claim: PM loop round R1 (folded member — chain head #10149)
Session: session_0185KTgmREZA4sY5pnRbubXj
Branch: claude/issue-10149-shard-weight-chain (shared with the chain head; this card has no branch of its own by design)
Worktree: objectstack-issue-10149
Domain: domain:devx
File surface: scripts/check-cross-package-test-inputs.mjs, scripts/partition-test-shards.mjs (stop on breach; explain in the report)
Container & model: M, mode:subagent, model: opus
Clause-②: no
Serial constraints cleared: see the chain head's claim — full check lives there, including the declared residual risk.

Why this card is folded into #10149

#10149 measured, on two 2026-08-20 merge_group CI runs, the exact failure mode this card predicts as its own consequence:

So the failure mode of this class is not a red step, it is a shard matrix that quietly stops balancing.

Measured spreads (Run this shard's tests step, one run): Test Core 8m18s / 4m54s / 16m28s; a second run produced a Test Core (2/3) leg of 3 seconds. Both violate #4859's acceptance thresholds (Test Core ≤ ~7 min, Dogfood ≤ ~5.5 min) by ~2.1–2.4×.

⚠️ This does not assert that this card is the cause. This card's "inert today" measurement (CI runs from the repo root, so both conventions resolve to the same directory) and #10149 §3's reading (the test-file-count proxy has decayed) are mutually exclusive explanations of the same symptom, and neither has been tested in the merge_group environment specifically. Adjudicating them is the dispatch's first deliverable, and either card's premise may be falsified — that is a valued outcome, not a failed run.

The two are landed together because this card's own entry-gate warning makes them inseparable:

⚠️ Whichever is chosen … it.path feeds countTestFiles() feeds the package weight feeds shard placement. A change here can move packages between shards

The three normalisation options enumerated in this card's body are carried into the dispatch verbatim as options, not pre-decided — including the third (leave the data, resolve against the repo root in the consumer), which the dispatch flags as the one that also fixes the cwd fragility this card measured (relative, cwd = /: weight 0).

Claimed cross-lane via the maintainer direct-dispatch channel; the domain:devx seat is untouched. Full authorisation record on #10149.

---

Generated by Claude Code