Command and capability reference

This page is the adopter-facing summary. The authoritative mapping ships inside the pack, because it is what the agent reads:

• references/capability-map.md — every intent, its exact command, and what that command returns.
• references/gaps.md — a fourteen-point assessment against a general code-intelligence contract.
• references/evidence.md — how provenance, confidence and completeness are carried into an answer.
• references/investigation-patterns.md — the five patterns and their stop conditions.
• references/example-prompts.md — worked examples across debugging, feature impact, architecture, refactoring and modernization.

Two surfaces

Wicked Estate ships a command-line tool and an MCP server. The pack drives the CLI by default, for two reasons: the server advertises 30 tool schemas that stay resident in the agent’s context for a whole session, and the CLI is the larger surface — 42 names, several with no MCP equivalent.

Surface	Maturity	Why
CLI	validated	Exercised end to end against a real index; tests pin the returned shapes of resolve, blast-radius (with --depth and confidence summary), path, lineage, traverse, rank (with --seeds), rules-inventory (top-level keys; per-engine entry fields are source-read), rules-recall, and other verbs on a small fixture
MCP	contract-complete	Mapped from upstream’s registered tools and conformance schemas. Never executed by this pack

The memory, knowledge, and proposal domains are MCP-only. Register the server only if you need them, and prefer --readonly. The richer MCP response shapes (summary.top_by_pagerank on BlastRadius, Communities summaries, ContextBundle) are also MCP-only. With --readonly, a store path that does not exist is not created: the server serves that store as an empty in-memory store and says so on stderr. If every answer from one store is empty, check its path: the graph path comes from --db, else WICKED_ESTATE_DB, else .wicked-estate/graph.db relative to the server’s working directory, and the memory, knowledge, and cross-store paths come from WICKED_MEMORY_DB, WICKED_KNOWLEDGE_DB, and WICKED_XEDGE_DB, else from WICKED_HOME, else $HOME/.wicked.

Intent to command, in brief

Intent	Command
Resolve a name to a stable ID	wicked-estate resolve <name> --json
Fetch source	wicked-estate source --symbols <id> --json
What depends on this	wicked-estate blast-radius <name> [--depth N] [--cursor C] --json
What this depends on (forward)	wicked-estate lineage <symbol> --json (a name or a symbol id)
Walk by direction and edge kind	wicked-estate traverse <symbol> [--direction ...] [--edge-kinds ...] --json
Route from A to B	wicked-estate path <from> <to> [--max-depth N] --json
Bounded neighbourhood	wicked-estate graph-view --focus <name>
Load-bearing symbols	wicked-estate rank [--seeds ...] [--limit N] --json
Rules inventory	wicked-estate rules-inventory --json
Rules recall	wicked-estate rules-recall [--severity ...] --json
Subsystems	wicked-estate clusters --json
Task context	wicked-estate context <name> --budget <chars> --json
Graph size and freshness	wicked-estate stats --json (indexed commit, index time, commits behind)

Traps worth knowing: lineage accepts a name or a symbol id, but an unknown or ambiguous name exits non-zero with nothing on stdout, so resolve first when a name may be shared. The CLI parses arguments strictly: a malformed value, a surplus operand, a repeated flag, or a --db with no path exits 1 with nothing on stdout, which is a command error and not an empty answer. Most read commands refuse a --db path that does not exist or is empty with no graph at <db> and create nothing; clusters, context, semantic, semantics, drift, cross-graph and correspond instead turn it into an empty graph and return an empty result. Before running one, check that stats --json --db <path> shows a non-zero nodes count for every database it will read. semantic and clusters --weight semantic also return an empty result when the index has no embeddings, with the reason only on stderr, and until tfstate has imported live state, drift lists every declared resource as undeployed, which is not drift. rank --seeds biases a graph-wide ranking — seeded output includes symbols the seeds cannot reach. To rank a dependent set, keep only the rows whose symbol is in your set and report missing members as unranked.

What the CLI reports about completeness

• blast-radius --json returns unresolved, truncated_dependents, searched_depth, depth_horizon_reached, node_cap_reached, next_cursor, page {from, count, total}, and a confidence {min, avg, edge_count} summary. One reply stays under 25,000 characters; truncated_dependents counts the rows on later pages. Pass --cursor <next_cursor> until next_cursor is null to get the whole resolved answer. A refused cursor (exit 1, nothing on stdout) means the answer changed, so start again from the first page. An accepted cursor says nothing about freshness; stats --json does. A true depth_horizon_reached means more dependents lie beyond the depth searched; raise --depth (default 12, maximum 24). A true node_cap_reached means the node budget cut the walk, and no flag raises it. Each blast-radius row carries a per-row depth (1 is a direct dependent) and no per-row confidence.
• lineage --json pages the same way, with next_cursor and --cursor. Its default depth is 24, which is also its ceiling, so a true depth_horizon_reached there has no deeper setting to try.
• path --json returns found, depth_bounded, node_bounded and unresolved. found: false means no route exists only when unresolved is null and both bound flags are false. A misspelled name sets unresolved and still exits 0. With a bound flag true, a route may lie beyond the search. path follows every dependency edge kind, so read each hop’s kind. Endpoint line counts from 0; use line_1based.
• Commands served by the same retrieval tools as the MCP server (lineage, traverse, rank, rules-inventory, rules-recall) write STALENESS: to stderr under --json; in text mode they print it on stdout after the result (source-read from tool_bridge.rs; pinned on lineage --json). They print N commit(s) for each root that is behind, unknown for each root whose freshness could not be determined, and a single 0 commits line when every root was checked and none is behind (source-read from main.rs). CLAMPED:, the line printed when a limit such as --depth above 24 is lowered to its ceiling, follows the same channels. blast-radius and path suppress STALENESS: under --json, and text stats stays silent when freshness is unknown, so check freshness with wicked-estate stats --json: the graph is current only when staleness.commits_behind is 0 and staleness.unknown is empty. query, clusters and context print commits-behind lines on stdout before their JSON, even under --json.

What does not exist on the CLI

• A ranked-dependents summary on blast-radius. MCP BlastRadius adds summary.top_by_pagerank, which the CLI lacks. Its rows carry a per-row depth, and --depth 1 still returns only direct dependents; see evidence.md for the split.
• Memory, knowledge, and proposal domains. These have no CLI verb.

Commands that write

Read-only work needs none of these, and the pack asks before any of them: index, scip, evidence, tfstate, import-telemetry, annotate, semantics, compact, watch, supports retract, and clusters --annotate.