For AI

Use the same engine findings as the editor. This reference covers 0.35.0. For connection setup, see AI Agents.

Check before you write

Call find-similar with a proposed snippet, or a path, start_byte and end_byte. It can find an existing implementation even when that implementation appears only once in the repository. For existing duplication, use duplicates, then cluster-by-id to inspect the finding.

Read the returned kind and source ranges:

`kind`	Meaning and action
`identical`	Source text matches apart from permitted whitespace differences. Reuse the existing implementation where appropriate.
`nearly_identical`	Renamed, parameterized or lightly edited copies. Inspect differences before extracting shared code.
`loosely_similar`	Substantial copied work with larger edits. Read both ranges.
`same_behavior`	Optional embedding-based match. Similarity does not prove interchangeability.
`structural_only`	Informational shape match, not a clone. It contributes nothing to duplication figures.

A cluster's mass ranks impact, not confidence or extraction safety. A cluster carries no pair similarity measurements. Request an explicit comparison with both endpoints when you need evidence about a particular pair.

If the MCP server is unavailable, use the CLI

Check that the editor's LSP server is running and that MCP uses the same workspace root. A not-ready response or a connection error is not evidence that no similar code exists. A build mismatch must be resolved by using matching bundled binaries.

For a separate scan:

deslop . --notext --nohtml --no-color

Read .deslop/deslop-report.json. The CLI cannot query an unwritten snippet: inspect the baseline report before a change, scan again afterwards, and check the changed paths in clusters[].occurrences[]. Reuse or consolidate confirmed copies, then rescan. Unchanged files can reuse cached parse work, but analysis still considers repository-wide relationships.

The CLI can also compare two existing occurrences. If neither surface is available, report that limitation rather than guessing.

Read the JSON

Use the report's embedded schema_doc; over MCP, request schema-doc once per session. Parse JSON rather than text or HTML.

Field	Meaning
`metrics.duplication_percent`	Covered clone lines divided by analysed lines. Not detector accuracy.
`metrics.threshold.breached`	Whether the configured duplication ceiling was exceeded. The CLI writes reports before exiting `3`.
`clusters`	Clones in descending mass order, then informational findings.
`cluster.kind`	The classification above; current reports use `kind`, not the older `bucket` field.
`cluster.mass` / `cluster.rank`	Clone impact and position. Informational findings have zero mass and no clone rank.
`cluster.severity`	Diagnostic level, independent of mass and duplication percentage.
`occurrences[].hidden`	An occurrence hidden by report policy; it does not contribute duplicated lines.

Byte ranges, not line numbers

Ranges are [start_byte, end_byte), with the end excluded. They refer to the scanned source version. Refresh the analysis after edits before reusing offsets.

Cluster IDs are stable

Use the cluster ID to identify a finding across surfaces. Rank changes with the report order; IDs can change when the underlying clone content changes.

Configuring a repository

exclude removes a file before analysis. report_hide retains analysis but hides its occurrences. Built-in rules already cover dependency directories, build artifacts and common generated files. Thresholds opt into failing CI; they do not change what constitutes a clone.

Operating rules

Read the actual occurrences before editing. Do not treat a label as proof that a refactor preserves behaviour, or weaken thresholds to conceal a finding. If duplication is deliberate, record why. See Accuracy Transparency for measurement limits and the agent recipe for a reusable project rule.