MCP reference

The five tools and six prompts at the agent boundary — inputs, outputs, and what MCP will never decide.

`health`

Project identity, Git revision, working tree, provider status, and coverage. A missing Codebase Memory is an explicit degraded result, never invented coverage.

`prepare_change`

Input: target (path or path::symbol) and the agent’s real intent. Optional: max_tokens (default 2400), max_nodes (12), max_relationships (16), mode: baseline to skip intent-conflict detection.

Output: an operation_id and a packet with

critical_constraints and decisions that govern the target, each with authority, provenance, and how it matched;
relationships explained by Records, with derived state and the path when indirect;
structure: the selected nodes and edges around the target, each tagged with its role and any explaining Record ids;
relevant_code, intent_conflicts, known_risks, known_unknowns, warnings, the consistency snapshot, and budget_overflow when authoritative context exceeded the ceiling.

The budget is a ceiling, never a target: an empty neighborhood produces a small packet, and governing knowledge is never cut to fit.

`explain_target`

Why a target exists: the Records that govern it by exact binding, its Subject, and what is known versus unknown. It uses the same matcher as prepare_change, so the two tools never disagree.

`finalize_change`

Input: operation_id, summary, and candidates. Each candidate requires kind, statement, rationale, durability, and bindings; it may add relationships, supersedes, severity, id, risks, evidence, and subject. A relationship is {source, kind, target} with path::symbol endpoints and a kind such as calls, uses, writes, imports, or tests; once resolved, it also anchors the Record. The diff base is the declared base_revision, else the HEAD that prepare_change saw, else HEAD.

Output: committed Records, discarded candidates with a reason, conflicts, superseded Records, the derived state of relationships the new Records explain, observed signals, and the mechanical capture of the change.

Discard reasons include transient, duplicate, missing_rationale, rationale_restates_statement, mechanical_noise, no_meaningful_binding, and legacy_contract (a pre-1.0 call without candidates).

`resolve_conflict`

Input: conflict_id, decision (keep_pinned or adopt_new), and human_answer — the person’s literal answer, kept for audit. Without human_answer the tool refuses. adopt_new requires the Git actor to be declared in .rationale/config.yaml.

Prompts

prompts/list returns six actions from the same source as the Claude Code shortcuts:

Prompt	Purpose	Arguments
`preflight`	Prepare context and state governing Records before editing.	`target`, `intent`
`explain`	Explain a possible Chesterton fence.	`target`
`capture`	Close a change with durable candidates.	optional `statement`
`conflicts`	Present conflicts with pinned Records and hand the decision to a person.	none
`health`	Diagnose MCP, provider, Git, and canon health.	none
`protocol`	Load the master protocol.	none

A retired prompt such as review answers with the action that replaced it. An unknown prompt is a JSON-RPC error and never ends the session.

Since v1.1.0, prompt and tool descriptions are written in English and ask the agent to reply in the user’s language, keeping identifiers verbatim.

Transport

Newline-delimited JSON-RPC over stdio, protocol 2024-11-05. One persistent provider session per server process. Diagnostics go to stderr; stdout carries only MCP messages.