JSON output
Every command prints one JSON object with --format json, versioned by schema_version and carrying the same meta block, so an agent can read any answer as data.
tessri scan --format json prints one object, versioned by schema_version (currently 1; adding fields does not bump it), with kind: "scan". Top-level keys: summary (sizes, by_status, orphans, test_only, reachability_coverage), languages, directories (with owners and ownership per directory), ownership (authors, bus_factor, no_active_owner, top_authors, needs_owner_files, codeowners; {"available": false} without git), review (window, totals with bulk, unowned, areas, authors, reviewers, largest_changes; {"available": false} without git or with review_days = 0), largest_files, hotspots, cleanup (orphan_files with owner, test_only_files), analysis, caveats, and with --full a per-file files table.
tessri graph --format json and tessri impact --format json print their own objects (kind is module-graph, file-graph, or file or directory with a target for an impact) with the same schema_version. tessri why <path> --format json prints the same facts the text shows, as data: file, verdict (status, confidence, reachability, entry_points, notes), git, owners, graph (reached_via, dependents, config_refs, doc_refs, depends_on, stranded_if_removed, and kinds: the kind of evidence behind each edge, as impact names them, for reached_via as one entry per hop and for dependents and depends_on as a map from path to kind) and advice; git, owners and graph are null when the analysis had nothing to say. tessri brief --format json prints kind: "brief": its own sections (change with the full base and head, the commits on the way and the uncommitted files in scope; attention; consequences, with not_analysed; validation, with candidates, ran, report and coverage; people; open) and, whole, the triage and check objects it was composed from.
The MCP server returns these same objects as each tool's structuredContent, unchanged.
Every JSON object also carries a meta block with the same names everywhere: tool, generated_at, repo (name, root), and git with available, reason (null, or not-a-repository, no-commits, git-not-found, git-error, not-requested), head (the checkout's HEAD, resolved on its own), history_rev (the commit history was read up to: HEAD for a scan, the base for triage and check), shallow and commits. All ids are full object ids. triage, check and brief add meta.change (since, as given or the default branch chosen when none was, base, head, uncommitted), and check and brief add meta.analysis, which says skipped with reason no-code-changes when nothing was compared, so an empty findings list is never mistaken for a clean comparison. The older top-level fields (git in a scan, the short change.base and change.head of triage and check) keep their meaning. A caught error (an unknown base, a directory that is not a repository) exits 2 and prints nothing on stdout; exit 1 is reserved for check --fail-on, a result that was computed and failed the chosen gate.
Source: docs/guide/json.md, Markdown in the repository; this page is built from it. Edit it on GitHub.