Archived
Lucas WintherandClaude Opus 5
3d6067f4fa
Writes docs/specs/001-search-engine-api.md covering the search engine and HTTP API: collections managed at runtime, per-file membership supplied to Recoll at index time, content-hash document identity with canonical paths, an in-process indexing job queue, scoped search, and authentication. Twenty-four numbered requirements, each with an acceptance criterion, plus the constraints the design must hold: read-only documents, per-file membership, content-hash identity, and no silently unscoped search. The spec is status: draft. It has had its self-review pass but has not been signed off, so no plan and no code may be written against it yet. Self-review changed three things. Deleting a collection now states that it enqueues reindexing of the affected canonical paths, because the previous wording claimed no index work while leaving the deleted id in the index. The search scope parameter is specified as collection ids rather than names, since names are mutable and would break saved queries on rename. And the reported match count is documented as approximate, because Xapian estimates it for large result sets. Three known gaps are recorded rather than left implicit: there is no way to read a matched document until spec 002, the indexing overhead figures come from small text files rather than real PDFs, and every authenticated user sees every collection until spec 004. docs/TODO.md gains a "Resume here" section stating where the workflow stands, what the next action is, and the traps a resuming agent would otherwise hit -- chiefly that the extended-attribute membership design was superseded mid-session and survives in the append-only decision log. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>