CLI and MCP capability coverage¶

Verified source implementation: 2026-10-17. All 46 UI coverage IDs are accounted for: 36 implemented equivalents and 10 explicitly policy-restricted capabilities. Restricted rows have working permitted reads/local routes; they do not claim full MCP parity. gravetrail capabilities and MCP get_capabilities return this versioned inventory, including scope, aggregate effects, required grants and current launch authorization. Follow next_cursor; availability does not mean the current caller has permission.

These 46 IDs group UI capabilities; they are not command/tool counts. The registered interface has 98 CLI command routes, 29 CLI groups, 78 MCP tools, 2 concrete resources, 4 resource templates and 1 prompt (verified 2026-10-18). The CLI and MCP guides list every route/tool, including basic discovery commands outside this UI capability matrix. Documentation tests compare those references with actual registrations and inputs.

Use CLI guide for command options/examples and MCP integration for typed schemas and transport setup. Every workspace operation selects an explicit ID. Use --home at launch and keep one DuckDB owning process per home. Stop that owner before opening the same home with another process; separate test homes may run concurrently.

Read-only is the default. Write, configuration, outbound AI and application authority are independent startup grants. --allow-admin is local CLI only, requires unrestricted scope and cannot be accepted by MCP. Workspace settings can revoke access immediately. MCP may tighten privacy/access settings; it cannot expand its own policy. Application writes need unrestricted --allow-write --allow-config-write --allow-application-config.

Profile/routing reads and saves, chunk/review previews, reports and deterministic graph queries make zero provider calls. Discovery, connection tests and explicit AI requests need --allow-llm, exact --allowed-llm-origin and approved --allowed-credential-ref names, in addition to write/source/privacy gates where applicable. The catalog aggregates effects for each capability; an ordinary read within a mixed capability does not require its write grant. Never put a credential value in arguments, patches, profiles or MCP payloads. Use a read-only named launch variable, hidden per-invocation CLI input, or UI session input. Runtime bindings are scoped to workspace/profile, expire on profile changes and clear on close; UI credentials are separate. Fallback profiles resolve their own approved references. No environment mutation is used for session credentials. Redirects are refused.

ID	UI capability	CLI routes	MCP tools	State / restriction
P01	List/get/create/edit/remove provider profiles; name/provider/URL/model/reference/auth/timeout/retries/max_tokens/context_chars/temperature/structured/Azure API version	`providers list` `providers get` `providers save` `providers remove`	`list_provider_profiles` `get_provider_profile` `save_provider_profile` `remove_provider_profile`	Implemented
P02	Session connect/status/clear; reference compatibility and restart behavior	`providers credentials status` `providers test` `ai request`	`connect_runtime_reference` `get_runtime_connection_status` `clear_runtime_connection`	Restricted: MCP binds only an approved named launch reference. Raw credentials use hidden per-invocation CLI input or a UI session; never tool arguments.
P03	Model discovery with sanitized results	`providers models`	`discover_provider_models`	Implemented
P04	Actual model connection test and sanitized usage/failure	`providers test`	`test_provider_connection`	Implemented
P05	Default profile, each of five task routes and ordered fallback profiles	`providers routing get` `providers routing update`	`get_model_routing` `update_model_routing`	Implemented
A01	Selected evidence analysis: SQL lineage, SAS lineage, lineage validation, column usage, code explanation	`ai request`	`request_ai_assistance`	Implemented
A02	Offline statement chunk preview, selected chunk/range and budget information	`ai chunks`	`preview_ai_chunks`	Implemented
A03	Unresolved warning/error review queue, recommended source context, custom bounds and explicit AI review	`ai review list` `ai review preview` `ai review request`	`list_review_candidates` `preview_diagnostic_review` `request_diagnostic_review`	Implemented
A04	Saved AI result list/get, outcomes/quarantine, source-job/hash warnings, advisory download and AI activity/usage/duration	`ai results list` `ai results get` `ai results export` `ai activity`	`list_ai_results` `get_ai_result` `export_ai_result` `get_ai_activity`	Implemented
A05	Explicit validated lineage merge as inferred and graph refresh; no advisory merge	`ai merge`	`merge_ai_claims`	Implemented
S01	Default SQL dialect, per-folder/per-extension overrides	`settings schema` `settings get` `settings update`	`get_settings_schema` `get_workspace_settings` `update_workspace_settings`	Implemented
S02	Case-sensitive identities, wildcard expansion, CSV sample rows	`settings schema` `settings get` `settings update`	`get_settings_schema` `get_workspace_settings` `update_workspace_settings`	Implemented
S03	SAS library mappings, static macros, default job/session and folder job assignments	`settings schema` `settings get` `settings update`	`get_settings_schema` `get_workspace_settings` `update_workspace_settings`	Implemented
S04	Saved minimum confidence, include_inferred, diagnostic visibility, reanalysis mode	`settings schema` `settings get` `settings update`	`get_settings_schema` `get_workspace_settings` `update_workspace_settings`	Implemented
S05	max_file_mb1–100, max_files1–100000, parse_workers1–8	`settings schema` `settings get` `settings update`	`get_settings_schema` `get_workspace_settings` `update_workspace_settings`	Implemented
S06	AI workspace concurrency1–4, process cap8, chunk_chars500–100000	`settings schema` `settings get` `settings update`	`get_settings_schema` `get_workspace_settings` `update_workspace_settings`	Implemented
S07	external_llm, allow_full_source, max_context_chars500–100000, redact_secrets, llm_history, retention_days1–3650	`settings schema` `settings get` `settings update`	`get_settings_schema` `get_workspace_settings` `update_workspace_settings`	Restricted: MCP may tighten privacy. Enabling transmission/full source/history, increasing disclosure/retention budgets or disabling redaction requires local CLI --allow-admin or UI.
S08	Application logging DEBUG/INFO/WARNING/ERROR, persisted per home	`application logging get` `application logging update`	`get_application_logging` `update_application_logging`	Implemented
S09	Agent enabled/write/source permissions, default detail, max_chars512–100000, audit_reads	`settings schema` `settings get` `settings update`	`get_settings_schema` `get_workspace_settings` `update_workspace_settings`	Restricted: MCP cannot enable disabled access/source/write, increase detail/response budgets or disable read auditing. Use local administrator CLI/UI.
S10	Actual local MCP-handler check, available schemas and generated Codex/Claude stdio/HTTP examples	`agent inspect` `agent self-test` `agent client-config`	`get_agent_setup` `test_local_agent_access`	Implemented
W01	Rename existing workspace and edit description	`workspace update`	`update_workspace`	Implemented
W02	Export allowed configuration and create new workspace from v1/v2 configuration; exclude profiles/secrets/privacy/source/history	`workspace config export` `workspace config import`	`export_workspace_configuration` `import_workspace_configuration`	Implemented
W03	Storage locations/home status and current browser home selection equivalent	`runtime info`	`get_runtime_info`	Restricted: No live server home switch. Relaunch CLI/MCP with an authorized --home; UI changes its current browser session only.
W04	Analysis and ingestion run lists/details, inventory file_size/encoding/metadata and import diagnostics	`runs list` `runs get` `ingestions list` `ingestions get` `source list` `source metadata`	`list_analysis_runs` `get_analysis_run` `list_ingestion_runs` `get_ingestion_run` `search_source_code` `get_source_metadata`	Implemented
W05	Baseline/comparison run diff and JSON download	`runs compare` `export report`	`compare_analysis_runs` `export_report`	Implemented
V01	Eighteen theme presets, detected/configured/loaded theme and explicit apply	`application theme list` `application theme get` `application theme apply`	`list_theme_presets` `get_application_theme` `apply_application_theme`	Implemented
V02	Saved graph node limit10–1000, traversal depth1–30, tree expansion0–30, upstream/downstream/combined, collapse temporary	`settings schema` `settings get` `settings update`	`get_settings_schema` `get_workspace_settings` `update_workspace_settings`	Implemented
V03	Graph labels compact/full/hidden, semantic/neutral palette, width1–5, source complete/relevant and line numbers	`settings schema` `settings get` `settings update`	`get_settings_schema` `get_workspace_settings` `update_workspace_settings`	Implemented
V04	Interactive graph/tree/edge/node focus, expand/collapse, direct column usage and bounded column transformation lineage with roles, source evidence and coverage	`lineage graph` `columns usage` `columns catalog` `columns lineage`	`get_graph_data` `get_column_usage` `get_transformation_columns` `get_column_lineage`	Implemented
R01	Data dictionary: declared SQL types, sampled CSV types, fallback direct references, provenance and unknowns	`dictionary list`	`get_data_dictionary`	Implemented
R02	Declared schema columns plus original DDL evidence	`assets schema`	`get_asset_schema`	Implemented
R03	Overview hotspots, isolated/temporary assets, cyclic groups and parsing-status counts	`workspace health`	`get_workspace_health`	Implemented
R04	Standalone stored-lineage evidence check (does not analyze or repair)	`lineage validate`	`validate_stored_lineage`	Implemented
R05	Dependency paths in both directions, max_paths1–500, depth, temporary view, boundaries and exports	`lineage paths` `export paths`	`get_dependency_paths` `export_dependency_paths`	Implemented
R06	Step-wise lineage rows/logical dependency order with evidence; no inferred runtime order	`lineage steps`	`get_lineage_steps`	Implemented
R07	Code search matching line numbers/snippets and report export, beyond current metadata-only file search	`source search` `export report`	`search_source_code` `export_report`	Implemented
E01	Interactive offline graph HTML and graph visualization JSON with appearance	`export graph`	`export_graph`	Implemented
E02	Relevant original source ZIP, selected original file download and historical evidence versions	`export sources` `source download`	`export_source_bundle` `export_original_source`	Implemented
E03	Inventory/dictionary/column/impact/code-search/run-comparison CSV/JSON/HTML reports matching available UI formats	`export report`	`export_report`	Implemented
C01	Cleanup counts/estimates and all existing workspace-category previews/applies/reset/deletion	`cleanup inspect` `cleanup sources` `cleanup lineage` `cleanup columns` `cleanup ai` `cleanup history` `cleanup jobs` `cleanup reset` `cleanup workspace` `workspace reset`	`inspect_cleanup` `preview_cleanup`	Restricted: MCP cleanup is read-only. Guarded local CLI apply needs a fresh preview, typed name and full-home backup.
C02	Shared parser/Streamlit caches, orphan blobs and full recognized application-data reset	`cleanup inspect` `capabilities`	`inspect_cleanup` `preview_cleanup`	Restricted: Shared cache/orphan/application apply is UI-only. Authorized unrestricted previews expose estimates, never apply tokens or paths.
C03	History retention preview, archive eligibility, archive list, local archive and recovery restore	`history preview` `history archive` `history archives` `history restore`	`preview_history_archive` `list_history_archives`	Restricted: MCP may preview/list history archives only. Archive/restore is an explicit guarded local CLI/UI action.
C04	Stale staging preview, selected cleanup, age24–8760h, recovery list/restore	`staging preview` `staging clean` `staging recoveries` `staging restore`	`preview_staging_cleanup` `list_staging_recoveries`	Restricted: MCP may preview/list staging recoveries only. Clean/restore is an explicit guarded local CLI/UI action.
C05	Automatic expired staging preference; only app-created files,24h,100 files/25MB per visit	`settings get` `settings update` `staging clean`	`get_workspace_settings` `update_workspace_settings` `preview_staging_cleanup`	Restricted: MCP may read/disable the automatic preference, never enable or execute it. Automatic cleanup runs only on the existing opt-in UI event.
M01	Database integrity/storage, checkpoint, full-home backup, backup list/download, new-home restore	`maintenance inspect` `maintenance backup` `maintenance checkpoint` `maintenance restore` `maintenance backups` `maintenance download`	`inspect_database_maintenance` `list_runtime_backups`	Restricted: MCP exposes authorized unrestricted metadata only. Checkpoint/backup/restore and backup download are local CLI/UI; restore requires a new home.
I01	Browser ZIP/folder uploads and file inspection	`ingest` `source list` `source metadata` `source download`	`ingest_sources` `search_source_code` `get_source_metadata` `export_original_source`	Implemented

Artifacts use managed IDs, home/workspace binding, exact source versions, SHA-256 and bounded base64 reads (at most 16 KiB before encoding and the response budget). Files cap at 64 MiB. Source authorization is checked again on retrieval, including saved exports after revocation. CLI downloads create a new destination exclusively and verify the assembled hash. Unsupported kind/format pairs fail; no arbitrary file read/write or backup retrieval MCP tool exists.

Settings/profile patches require expected_revision; refresh stale state before retrying. AI requests pin original hash/job and explicit line bounds. Schema/quote failures stay quarantined. Only validated lineage tasks may merge, always as inferred relationships; repeat merge is idempotent. A committed merge with failed graph refresh reports retry required. Synchronous requests have bounded concurrency/timeouts; busy requests refuse without queuing. After an ambiguous disconnect, inspect saved activity before retrying: provider billing is not exactly once. Summary pages omit source-bearing details; exact evidence is an explicit, separately authorized request. Respect analysis_current, complete, boundaries and cursors.

General cleanup requires a fresh preview, explicit local confirmation and successful full-home backup. History/staging journals are specialized recovery, not general undo. Restore full-home archives only into a new destination. Shared cache/orphan/application removal stays UI-only; MCP maintenance stays read-only. Reading settings/recoveries never triggers automatic cleanup. Runtime reset and schema removal remain separate operations.

The checked-in capability fixture is tested against actual Typer routes/help, MCP schemas and authorization. The testing guide describes parity checks for transports, the running UI, both real ZIPs and live providers.