Repository navigation
Description
What the code prints
At head 8948d9a, src/core/tools/WriteToFileTool.ts still publishes with the pre-guard argument lists: saveDirectly is called with five arguments at line 135 and saveChanges with two at line 167, so the guarded kinds and the approved-target arguments stay defaulted.
- The prevent-focus arm reads the whole file itself at line 114 and records no observation. guardedWrite therefore finds obs undefined and takes its never-read branch (src/core/tools/guardedWrite.ts lines 674-682, createIfAbsent): it refuses a target that this very arm had just read end to end. The arm holds the complete content and the guard is told nothing about it.
- Neither arm threads the approved outside-workspace identity that saveChanges (src/integrations/editor/DiffViewProvider.ts lines 530-541) and saveDirectly (lines 1689-1704) exist to receive, while the tool computes isOutsideWorkspace at line 88 and puts it only into the approval payload at line 94.
- The diff-view arm behaves as designed and is not part of this unit: open() records a partial preview observation only when none exists (DiffViewProvider.ts 203-206) and a verified empty placeholder as complete (239-259), so a never-read existing file correctly gets the re-read remediation and a new file publishes against the placeholder token.
Measured size of the fix
32 changed executable lines in package extension, measured with the repo gate's own function (scripts/stryker-diff.mjs buildManifest over the drafted working-tree diff; 49 raw lines added, 4 deleted). The ceiling is 500 changed executable lines per package. Drafted source is preserved in the lane capture at .tmp-n202237-cutshape/sec83/WriteToFileTool-draft.ts.
This is why the fix is a unit of its own rather than a commit on pull 2444: that pull is priced by the hosted gate itself at 1315 changed executable lines (log line of job 132565556554) against the 500 ceiling, so the gate refuses there whatever this delta is, and carrying it would buy the change no mutation proof. On its own branch the delta is 32 of 500 and the gate can run on it.
Plan for the unit
- Branch n20/w2f-publish-guard from main at f015dae.
- src/core/tools/WriteToFileTool.ts: bracket the prevent-focus arm's whole-file read with stats and record it as a complete observation only while the bracketing stats agree; name the guarded kind the operation is (update for an existing target, create for an absent one); capture canonicalizeForApproval before the ask for an outside-workspace target and thread it, plus the approved flag, into both publish arms. No knob, no suppression, no directive.
- Add the regression spec at the lowest layer that would have failed: src/core/tools/tests/writeToFileTool.guardedPublish.spec.ts, asserting the arguments the two publish calls receive and the observation the prevent-focus arm records - not merely that a method was called.
- Prove both directions: the new spec red at the pinned head blob, green with the fix; counts printed by each run with the transcripts named and the on-disk hash of the pinned source captured before the run starts.
- Run the local mutation preflight on the unit delta (its own base, not main) before pushing, with no Stryker disable directive added.
- Lint with the shared prune switch left off from a real pnpm install --frozen-lockfile, prove src/eslint-suppressions.json byte identical to its blob, add no changeset and no changelog entry per AGENTS.md.
Gaps this unit does not claim to close
- Host scenario: no recorded extension-host scenario exercises the refusal. Both recorded write_to_file scenarios create new files (fixture ids call_write_to_file_create_001 and call_write_to_file_nested_001, expected payload fragments "operation":"created"), so a green host run on that suite shows no regression and does not clear the refusal path. A host scenario that overwrites an existing file would have to be written and is not claimed here.
- Assertion hole, the same one apply_diff had before its own spec: src/core/tools/tests/writeToFileTool.spec.ts asserts only that saveChanges was called (expect(mockCline.diffViewProvider.saveChanges).toHaveBeenCalled()) and never reads its arguments, so it passes identically with and without the fix.
Where this came from
Named for provenance, with no verdict asked of anyone: pull 2444 (#2444) is where this class of defect was found while repairing that pull's red required leg e2e-mock; the sibling tool was named there and left unfixed. Queue rows q4352 (this unit) and q4344 (the 2444 leg) carry the lane records.
Activity
fix(write_to_file): publish its own whole-file read as a read, and th…
easonLiangWorldedtech commentedon Nov 14, 2026
Commit
4a909332803cf3cf81a3585d59e76c44a9a4adb4carries the fix this issue plans, and it lands inside pull #2444 rather than in a pull of its own.The reason is one line: the guarded kind and the approved-target parameters this unit threads arrive in pull #2444's own commits and are absent from main tip
f015dae5326501ef5eb70616ac698828465b0deb, wheresrc/core/tools/guardedWrite.tsdoes not resolve as a path at all andsrc/integrations/editor/DiffViewProvider.tscarries no line namingGuardedWriteKind,writeKind,canonicalizeForApprovalorobservationRegistry- measured on the blobs: at main tip that DiffViewProvider is blob3c1d33ae283bed12d0c756503985eb0f9d5180ecwith none of those names andsaveChangesat line 326 andsaveDirectlyat line 1142 with no guarded kind or approved target in either signature, while at8948d9a30dguardedWrite.ts is blob9bede974293e8e01ca8a1d5996f4bb5e8d4b4132and DiffViewProvider is blob03b95a15651f1e7328ee582e53c0d20e0e38ee38carrying 18 such lines, andmerge-baseof main tip and the8948d9a30dhead is main tip itself (main tip..8948d9a30dis 5 commits,8948d9a30d..main tipis none). So a pull of this commit against main would fold #2444's own 33 files and +10,309/-263 into this unit's diff and the gate would price it at 1347 changed executable lines, where the unit's own commit against the base it declares is 32.What the commit does, in the shape this issue planned: the prevent-focus arm's whole-file read is bracketed with
fs.stat({ bigint: true })and recorded throughtask.observationRegistry.observeas a complete observation only while the bracketing tokens agree, which is what takes guardedWrite off its never-read branch; both publish arms now name the kind the operation is (updatefor an existing target,createfor an absent one) instead of leaving it to the signature default; and the outside-workspace identity fromcanonicalizeForApprovalis captured before the approval is asked and threaded, with the approved flag, intosaveDirectlyandsaveChanges.Proof run on the pinned source before and after: the new spec
src/core/tools/__tests__/writeToFileTool.guardedPublish.spec.ts(sha256E81B5475CA0E237470518AD2800172562E9B170D4909A5D8C67B5AED8C731A70, 12 tests) printsTests 11 failed | 1 passed (12)against the pinned head blob0c5c80abb91ec21307b40d16f72819cfb7206d9b(on-disk sha256C767216F11E56815AA0D334F772DEA1B280005FB505500D10E0D46AF2752002F) andTests 12 passed (12)against the fixed blobdc01b4edac91cf8b559ccc2bbe3e0cf7b1c7b668(on-disk sha2560A7A478AAF7D6C09390886A390FC5E9D8688AC6BD4CCE97964F9CD50276AA25A). The neighbourhood suites printTests 323 passed | 5 skipped (328). Locally, the mutation gate run on the unit delta (32 changed executable lines in the extension package against the base the commit declares) instrumented it asInstrumented 1 source file(s) with 34 mutant(s)and printed valid 34, killed 30, timeout 0, survived 4, noCoverage 0, ignored 0 with 4 blocking mutants and resultAdvisory findings, with no Stryker disable directive added anywhere. All four survivors sit on the bracket agreement condition at lines 134-135 and would needreadTokendefined whilereadContentis undefined, or the reverse, a state this arm cannot be in because the read is gated on the token existing; the line 136-137 half of that condition has no surviving mutant. Two earlier runs of the same delta are on the record too: the gate tool own preflight could not start Stryker on this host at all (spawnSync ...node_modules/.bin/stryker ENOENT, exit 0, empty digest - evidence in neither direction), and the first manual equivalent reached the 5-minute initial-run default and printedInitial test run timed out!before a second, with--dryRunTimeoutMinutes 20added, printed valid 34, killed 18, survived 15, noCoverage 1 on the then-current commit, which is what the strengthened spec was written against.src/eslint-suppressions.jsonis byte for byte the blobd5b845997a775924f598d6b47793fe226b783c61at both ends of the delta: 264 numeric leaf entries summing to 2866 at the base blob and 264 summing to 2866 at the head blob, with none moved, none risen, none gone and none added.Still open, named rather than banked: the two host scenarios for write_to_file both take the create path (
call_write_to_file_create_001,call_write_to_file_nested_001,"operation": "created") so no host scenario covers the update path this defect sits on, and the assertion hole apply_diff had -expect(mockCline.diffViewProvider.saveChanges).toHaveBeenCalled()inapps/vscode-e2e/src/suite/commands/writeToFileTool.spec.ts- is the same one here. Both are recorded in the queue row for this unit.