MCP ↔ CLI tool-name mapping

Two surfaces, two naming conventions:

MCP (Model Context Protocol) tools live in a flat namespace shared with every other MCP server an agent has loaded. Names like code_search, ledger_append, scratch_write use a code_ / ledger_ / scratch_ prefix so they don't collide with search, append, write from other servers.
CLI subcommands live under asd with hierarchical nesting: asd search, asd ledger append, asd scratch write. The hierarchy matches --help output and human shell habits.

These two conventions cannot collapse to one without breaking either MCP namespace hygiene or CLI ergonomics. This document is the canonical map.

Direct 1:1 names (same name on both surfaces)

MCP tool	CLI command
`annotate_commit`	`asd annotate-commit`
`architecture`	`asd architecture`
`callers`	`asd callers`
`callees`	`asd callees`
`checklist`	`asd checklist`
`context_for`	`asd context-for`
`dead_code`	`asd dead-code`
`effect_declare`	(MCP-only — no CLI verb)
`effects`	(MCP-only — no CLI verb)
`endpoints`	`asd endpoints`
`health`	(MCP-only — HTTP healthcheck)
`impact`	`asd impact`
`investigate`	`asd investigate`
`map`	`asd map`
`prepare_change`	`asd prepare-change`
`recipe_classify_test_migration`	`asd recipe classify-test-migration`
`recipe_migrate_stale_tests`	`asd recipe migrate-stale-tests`
`references`	`asd references`
`reindex`	(MCP-only — library call)
`scopes_list`	`asd scopes list`
`scorecard`	`asd scorecard`
`search`	`asd search`
`since`	`asd since`
`status`	`asd status`
`sync`	`asd sync`
`task_close`	`asd task-close`
`test_summary`	`asd test-summary`
`traces`	(MCP-only — execution traces)
`trust`	`asd trust`
`verify_effects`	`asd verify-effects`

MCP tool	CLI command	Why prefix
`code_search`	`asd search`	"search" collides with other MCP servers
`code_read`	`asd read`	"read" collides with file-read tools
`code_query`	`asd search` (via filters)	"query" too generic for the flat namespace

The CLI also accepts the MCP-era names as aliases (Plan D t-003), so asd code_search foo and asd search foo both work.

These MCP tools collapse a CLI subcommand-of-subcommand into one underscore-joined name because MCP has no nested-command concept.

MCP tool	CLI command
`audit_tail`	`asd audit tail`
`audit_verify`	`asd audit verify`
`conclusions_list`	`asd conclusions list`
`conclusions_export`	`asd conclusions export`
`conclusions_import`	`asd conclusions import`
`feedback_list`	`asd feedback list`
`feedback_mark`	`asd feedback mark`
`feedback_promote`	`asd feedback promote-as-truth`
`invariant_add`	`asd invariant add`
`think_speculate`	`asd think speculate`
`think_model`	`asd think model`
`think_failed`	`asd think failed`
`think_question`	`asd think question`
`think_list`	`asd think list`
`invariant_list`	`asd invariant list`
`ledger_append`	`asd ledger append`
`ledger_approve`	`asd ledger approve`
`ledger_find`	(MCP-only — no `asd ledger find`; on the CLI reads go through `asd list ledger` / `asd conclusions list`)
`ledger_get`	(MCP-only — no `asd ledger get`; on the CLI read a symbol's ledger via `asd read <sym>`)
`ledger_rebind`	`asd ledger rebind`
`ledger_reject`	`asd ledger reject`
`ledger_supersede`	`asd ledger supersede`
`ledger_withdraw`	`asd ledger withdraw`
`scratch_clean`	`asd scratch clean`
`scratch_discard`	`asd scratch discard`
`scratch_list`	`asd scratch list`
`scratch_promote`	`asd scratch promote`
`scratch_read`	`asd scratch read`
`scratch_update`	`asd scratch update`
`scratch_write`	`asd scratch write`

No _of suffix anywhere — callers_of / callees_of / effects_of / traces_of were retired in Plan A t-002. Use the bare form on both surfaces.
MCP names always use snake_case (e.g. task_close, conclusions_export).
CLI names always use kebab-case at the top level (task-close, prepare-change) and lower_snake for subcommands (ledger append, conclusions export).
code_ prefix is reserved for read-side tools that would collide with other MCP servers (code_search, code_read, code_query). New write-side tools that need disambiguation should use a verb prefix (recipe_classify_test_migration), not code_.

crates/agentstatedeveloper-mcp/src/mcp_server.rs::tool_name_regression — locks the renames + asserts the _of suffix never returns + verifies every tool above is registered.
The audit comment at the top of the #[tool_router] impl in mcp_server.rs (line ~668) restates these conventions for future editors.