fix(sandbox): charge each file ceiling once per execution, not once per source

[icecrasher321]

Two ceilings on a Function run's sandbox files were charged per source rather than per execution. Both came out of review comments on #7251; the behavior is real in each case, though not for the reason the comments gave.

Mounts — a storage key mounts once

planUserFileMounts assigned a path per array element, so one storage key named by two sources became two mounts.

The comment pointed at collectSandboxFileMountRefs appending a file once per marker position. That half was already handled upstream — resolveSandboxFilePathReference reuses the marker standing for a given file.key, so <block.file.path> twice in one block yields one marker. And the stated consequence, "the key-based path map makes every reference use the last mount", is harmless on its own: both mounts hold identical bytes.

What is actually reachable is the other input to the same list. mountedUserFiles comes from the files param, which is visibility: 'user-or-llm' and deduped nowhere. A model repeating a file id, or passing a file the code also references with <block.file.path>, produces a real duplicate — costing a presign, a second transfer of identical bytes, and a second charge against both the mount byte budget and MAX_BLOCK_MOUNTED_FILES, either of which then refuses a request that fits.

So the fix goes at planUserFileMounts rather than in the collector: it is the one place the whole mount set is assembled, so it covers the explicit list, the markers, and the overlap between them in a single pass. Keyed on the storage key, since that is the object's identity and already what mountPathsByKey looks paths up by. First occurrence wins, so the name listed first is the one the code sees.

No keyless branch: userFileSchema declares key: z.string().min(1), and the resolver drops a file without a key before it can become a marker, so there is no empty-key case to carry.

This only ever removes work. Fewer files reach MAX_BLOCK_MOUNTED_FILES and the byte budget, so it cannot newly refuse anything; the surviving mount is the first rather than the last, which is the un-suffixed name.

Exports — one 20-file ceiling per run, not per source

MAX_SANDBOX_OUTPUT_FILES is documented as how many files one execution may export "whether declared by path or discovered by harvesting the output directory". collectExportedFiles already runs the byte ceiling that way — one running total across both, with a declared path inside the harvest directory dropped from the discovered set rather than billed twice. The count ceiling did not follow: listOutputDirectoryFiles checked the harvest alone, so a request could export 20 declared plus 20 discovered.

Declared and discovered are now counted together, and the declared-path filter moved into the listing so count and bytes apply the same rule.

Scope of the behavior change: execute-request sets outputSandboxDir only when no output declares a sandboxPath, so on every call it makes, declaredPaths is empty — the comparison and the error message are exactly what they are today. The copilot doc tools never set the directory at all. The combination is admissible through the contract (outputs.files alongside the legacy top-level outputSandboxPath), which is what makes this worth closing rather than leaving to the next caller, but nothing on a normal path changes.

The naive form of this fix would over-refuse: a run whose 20 harvested files include the one declared path would count as 21. A test pins that case.

Tests

• planUserFileMounts: one storage key mounts once across sources.
• Conformance (both providers): a run that declares and harvests spends one ceiling; a declared path inside the harvest directory is not charged twice.
• Fixture keys in sandbox-mounts.test.ts were identical across files the tests meant to be distinct — the collision-suffix and inline-budget tests both passed only because their fixtures shared one key. They now differ, which is what those tests always claimed to set up.

bun run test (function-execution, remote-sandbox, executor/variables): 823 passed, 40 skipped. bunx turbo run type-check: 26/26. bun run check:api-validation clean.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
docs	Skipped		Aug 30, 2026 9:17pm

[greptile-apps]

Greptile Summary

This PR makes sandbox file ceilings apply once per execution rather than independently per source.

• Deduplicates mounted user files by storage key before assigning mount paths.
• Counts declared and harvested outputs together while excluding declared files from harvested results.
• Adds conformance coverage for combined output counting and duplicate mount sources.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
apps/sim/lib/execution/remote-sandbox/index.ts	Combines declared and harvested files under one execution-wide output count and avoids re-harvesting exact declared paths.
apps/sim/lib/function-execution/sandbox-mounts.ts	Deduplicates input mounts by storage key while preserving deterministic first-occurrence naming.
apps/sim/lib/execution/remote-sandbox/conformance.test.ts	Adds provider conformance tests for combined ceilings and exact-path deduplication.
apps/sim/lib/function-execution/sandbox-mounts.test.ts	Adds duplicate-key coverage and corrects fixture keys for tests representing distinct files.
apps/sim/executor/variables/resolver.ts	Clarifies that marker reuse limits request variables while mount planning enforces storage-key deduplication.

_{Reviews (2): Last reviewed commit: "fix(sandbox): spend each file ceiling on..." | Re-trigger Greptile}

[cubic-dev-ai]

All reported issues were addressed across 4 files

<sub>Reply with feedback, questions, or to request a fix.

Fix all with cubic | Re-trigger cubic</sub>

[icecrasher321]

@greptile

[icecrasher321]

@cubic review

[cubic-dev-ai]

@cubic review

@icecrasher321 I have started the AI code review. It will take a few minutes to complete.

[cubic-dev-ai]

All reported issues were addressed across 5 files

<sub>Reply with feedback, questions, or to request a fix.

Fix all with cubic | Re-trigger cubic</sub>