Return the replacement in the tool's own output shape

[lightningpixel]

Every compression was being discarded. This is the PostToolUse:Grep hook warning that has been open all session, and the answer is worse than expected.

What the transcript says

Claude Code validates a hook's updatedToolOutput against the schema of the tool it replaces. Read, Grep and Bash all return objects. The hook handed back a bare string:

PostToolUse hook returned updatedToolOutput that does not match Grep's output shape;
using original output. [
  { "expected": "object", "code": "invalid_type", "path": [],
    "message": "Invalid input: expected object, received string" }
]

The model never saw a compressed result. Orvix cannot tell — nothing is reported back to the hook — so it recorded the saving in gain.jsonl anyway.

On the session where this was found, all four compressions were rejected: three Grep and one Bash. orvix gain reported 3 864 tokens saved. The real figure was zero.

The shapes, taken from a recorded session

tool	result
Read	{ file: { content } }
Grep	{ mode, numFiles, filenames, content, numLines, totalLines, appliedLimit? }
Bash	{ stdout, stderr, interrupted, isImage, noOutputExpected }

The replacement now goes back in the shape it arrived in, mirroring extractText case for case so the field written is the one that was read. Grep states how many lines it returned, so that count is updated with it.

Why nothing caught it

Both the tests and the benchmark replayed tool_response as a bare string — a shape no tool actually produces. The hook was correct against its fixtures and wrong against reality, and the fixtures were the only thing looking.

Both now use the real shapes and fail loudly if the hook ever returns something the tool would not accept.

The benchmark ratio is unchanged at 0.699. The arithmetic was always right; it was measuring a path that did not exist. It measures the real one now.

Verification

• npm run build, npx tsc --noEmit, npm test — 376 tests pass
• npm run benchmark — ratio 0.699, success_delta 0, both fixtures preserved
• The new test was written first and failed with expected 'string' to be 'object' — the same complaint Claude Code made