Context-efficient MCP results#
Current behavior#
Terminal MCP calls to tact_edit, tact_check, and tact_apply_patch accept detail="auto" (the default) or detail="full".
autoreturns a compact, hash-bearing receipt for routine all-EXACTsuccesses.fullreturns the complete structured result.Healed matches, ambiguity, refusal, detected drift, and new diagnostics always retain full detail, even when
autowas requested.Reads, healing, delta reads, and navigation keep rich output because that evidence is the point of the operation.
Compact receipts are correlation handles, not diff proofs.
Successful writes report the result hash; read-only checks report the source and hypothetical result hashes.
To inspect a write later, retain the source hash from tact_read or request detail="full", then pass that source hash to tact_delta.
Passing the result hash after a write only confirms that the file has not changed again.
Direct Python calls keep their native return values; GUI clients may render richer payloads.
Tact can control the size of the result it produces, not the lifetime of a result retained by an MCP client. Context deletion, compaction, and any later summary are client responsibilities.
When to choose each mode#
Use tact_edit directly for ordinary writes.
Use tact_check when you need an explicit preview, a risky multi-edit inspection, or a dry run.
Ask for detail="full" when you are diagnosing a non-routine outcome or machine-reading the legacy detailed representation.
Dated evidence#
The following figures are frozen measurements from the August 2026 compact-receipt evaluation; they describe the evaluated implementation and inputs, not a current benchmark service.
Case |
Legacy → compact bytes |
Compact share |
|---|---|---|
Edit, one change |
247 → 88 |
35.6% |
Edit, ten changes |
1,670 → 90 |
5.4% |
Check, one change |
422 → 150 |
35.5% |
Check, ten changes |
2,033 → 152 |
7.5% |
Patch write, two files / ten changes |
1,887 → 159 |
8.4% |
Patch check, two files / ten changes |
2,393 → 286 |
12.0% |
The evaluation used cl100k_base only as a stable token proxy; byte reduction and token reduction are not interchangeable, particularly for SHA-256 values.
The historical validation recorded 1,043 passing tests and 4 skipped opt-in live-harness tests.
These counts are frozen evidence, not a claim about the current suite.