reference

Mistral Large 4 reviewed the first draft of SPEC-011 Sharing a Store-Truth Folder against the code on 2026-11-18. It ran read-only through the vibe CLI (--agent ask) in the main checkout, and was asked for defects only. It reported 12 findings. Claude checked each against the code: 11 are right, and 1 is partly right. All are folded into the draft. No code exists yet, so none of this could be checked by a test. Each finding was checked by reading the code it cites.

Sources: the answer and the brief, kept outside the repository in /home/daniel/.cache/agent-usage/vibe/2026-11-18-0042-spec011/ (ANSWER.md, brief.md; cost $0, Mistral Large 4). code: crates/sophia-hash/src/merkle.rs, crates/sophia-store/src/types.rs, logic.rs, folder.rs, sqlite.rs, crates/sophia-interface/src/edit.rs.

#findingright?what was done
1Record files are named by component, but a member’s id is member_digest(component, index). The envelope wrote no id and no index, so foreign records could not be placed.right (merkle.rs member_digest)The envelope writes id; the index is the record n number; the import checks id = member_digest(component, index) for every record (§3.1, §5 step 2).
2A log entry import is no LogOp. Every decoder would fail on it.right (types.rs LogOp, LogOp::parse in the decoders)A new LogOp::Import, stated as a format change (§5 step 5).
3A fact’s relation inputs carry a log position. After an import, the round trip changes bytes and is_stale misreports.partlyThe bytes do not change, because a fact is written as stored and its id covers the version. But such a fact reads as stale after an import. That is now stated as a loss of the rebuild (§3.2, §6).
4ROOT in the tree is never verified, and it conflicts in every merge.rightThe root digest is computed, never stored in the tree (§3.3).
5fact-external/ was called derivable. It is not: only it records that a fact predates every file.right (folder.rs, fact_external_key and its use in replace_in)Exported as the set facts-external (§2, §3).
6Annotation and fact ids were not recomputed on import.rightRecomputed with Annotation::new and Fact::new; a mismatch is refused (§5 step 2).
7Fact and annotation inputs were not checked for existence.rightAdded to §5 step 3.
8A union merge of lineages can give one id two predecessors, or a cycle, with no conflict.right (put_lineage keys on (lineage, id))The import refuses both (§4, §5 step 4); added to acceptance 3.
9Scanning for <<<<<<< refuses a source file that contains that text.rightMarkers are detected only where a line is expected, never inside an envelope’s length-delimited block (§5 step 1); acceptance 3 tests a source that contains them.
10A namespace may contain / (inside backquotes), and _root can collide with a real namespace.right (validate_path)File names are percent-encoded, with upper case encoded too, which also makes them safe on case-insensitive file systems. The root file is @root (§3).
11The check “every member that comes from files has a manifest” cannot be decided, and “no name claimed by two manifests” would refuse shared generic functions.right (edit.rs adds members in store-truth stores; replace_files allows shared generic functions)Only the direction manifest → live member is checked, and only definition names must have one owner (§5 step 4).
12store/truth-since was set to a digest, where SPEC-010 §3 defines it as a log position.right (folder.rs take_truth)It is set to the import entry’s seq (§5 step 5).

Not reported, as Mistral checked: owner/, member/, defined-in/, fact/ and nominal-index/ can be computed from manifests and nominal entities, and removed_seq is invisible to every read path (removed_seq IS NULL).