feat: MCP prompts capability — authoring method as static text (ADR-0008)

CHANGELOG.md

@@ -2,6 +2,17 @@
 
 All notable changes to tagged releases are documented here.
 
+## Unreleased
+
+- Add the MCP `prompts` capability (ADR-0008): three non-normative authoring-method prompts —
+  `author_pack`, `test_pack`, `fix_pack` — served as static, versioned text that the client's model
+  executes with the client's key. The server still calls no model, holds no key, and opens no
+  network connection; serving a prompt is read-only, like `get_schema`. Method content distills the
+  expressiveness studies: the resolution-model shapes that avoid conflicts, the `onUnknown`
+  discipline, the decimal-string rule for ordered comparisons, the prepared-facts ledger, and the
+  instance-matrix logic probe. Following a prompt does not make a pack conformant, and no prompt
+  interprets any policy.
+
 ## 0.2.0 - 2026-07-27
 
 - Add an EXPERIMENTAL evaluator (ADR-0007): `judgment-pack experimental evaluate` and the

docs/adr/0003-mcp-integration-and-testing-surface.md

@@ -12,6 +12,10 @@ deciders: Brian Jin
 > **Amended by [0007](0007-experimental-evaluator.md):** the surface gains exactly one explicitly
 > experimental evaluation tool (`experimental_evaluate`); the validation tools still evaluate
 > nothing.
+>
+> **Amended by [0008](0008-mcp-prompts-authoring-method.md):** the surface gains the protocol's
+> `prompts` capability — non-normative authoring-method text served statically; the client's model
+> executes it.
 
 ## Context and problem statement
 

docs/adr/0008-mcp-prompts-authoring-method.md

@@ -0,0 +1,87 @@
+---
+status: proposed
+date: 2026-07-27
+deciders: Brian Jin
+---
+
+# Serve authoring method as MCP prompts; the intelligence stays in the client
+
+> **Proposed.** The MCP server gains the protocol's `prompts` capability with three non-normative
+> method prompts — `author_pack`, `test_pack`, `fix_pack` — served as static, versioned text that
+> the client's model executes with the client's key.
+
+## Context and problem statement
+
+Authoring a judgment pack is judgment-laden translation, and by [ADR-0006](0006-authoring-lifecycle-in-the-client.md)
+it lives in the client: the runtime is a stateless oracle that never authors, decides, or holds a
+key. But the expressiveness studies produced hard-won *method* knowledge that currently lives only
+in prose documentation a filesystem-less client cannot reach: the single-outcome/detector
+architecture that §8's conflict rule forces authors to rediscover, the `onUnknown` discipline and
+its fallback-blocking behavior, the decimal-string rule for ordered comparisons, and the
+prepared-facts ledger that RFC 0007 makes basic authoring hygiene. Every new authoring client
+currently relearns these by tripping on them.
+
+An "encode this policy" *tool* was considered and rejected earlier in the project: it would put a
+model, a key, and nondeterminism inside the runtime, and would make the reference implementation
+the canonical interpreter of what policies mean — spec-by-implementation one layer up.
+
+MCP has a third primitive that threads this: **prompts** are static text the server serves and the
+*client's* model executes. The server never calls a model.
+
+## Decision drivers
+
+- Equip the client-side authoring loop without violating keyless / offline / stateless.
+- Turn study findings into method every author inherits, instead of folklore rediscovered per pack.
+- Keep interpretation plural: the runtime may teach *how to work*, never *what a policy means*.
+
+## Considered options
+
+- **A. MCP `prompts` capability** serving method guidance as versioned static text.
+- **B. Documentation only** (status quo): the method stays in `docs/`, unreachable from a
+  filesystem-less client's tool surface.
+- **C. An authoring tool** that produces packs server-side. Rejected again for the reasons above.
+
+## Decision outcome
+
+Chosen option: **A**. Serving a prompt is a read-only operation in the same class as `get_schema`:
+static bytes, versioned with the binary, no model, no key, no network, no state. It *strengthens*
+ADR-0006 rather than bending it — the prompt's entire content is instructions for the client-side
+loop against the oracle tools.
+
+Settled constraints:
+
+- **Three prompts, method-only:** `author_pack` (the guided create → validate → evaluate loop, the
+  §8-shape guidance, the decimal-string and `onUnknown` rules, the prepared-facts ledger),
+  `test_pack` (the instance-matrix probe: per-outcome, conflict, unknown, missing-evidence,
+  not-applicable, forced-outcome, ordered-comparison rows), `fix_pack` (diagnostics-driven repair
+  in carrier → structural → semantic order).
+- **Non-normative, and each prompt says so in its own text:** following the method does not make a
+  pack conformant; only validation decides conformance; the produced document belongs to the
+  client. No prompt implies the runtime blesses, stores, or interprets anything.
+- **No policy interpretation:** prompts teach the format's mechanics and the loop's discipline.
+  They never state what any domain's policy means. Argument values (a policy text passed by the
+  caller) are echoed into the rendered prompt verbatim, not interpreted.
+- **Versioned with the binary** like the embedded fixtures: prompt text changes ship as releases
+  and appear in the changelog.
+
+This amends [0003](0003-mcp-integration-and-testing-surface.md) a second time: the surface is now
+tools + prompts. The tools' posture is unchanged.
+
+### Consequences
+
+- Good, because a filesystem-less client gets the authoring method at the same place it gets the
+  oracle, and study findings stop being folklore.
+- Good, because the division of labor becomes explicit and inspectable: the runtime ships the
+  method, the client ships the mind, the tools stay the oracle.
+- Bad, because the reference runtime's prompts become the de facto authoring *style*; mitigated by
+  the non-normative marking, by the prompts being small inspectable text rather than behavior, and
+  by teaching mechanics rather than interpretation.
+- Bad, because client support for MCP prompts is uneven; degradation is graceful (tools are
+  untouched) but the guidance reaches only clients that surface prompts.
+
+## More information
+
+Method content sources: the expressiveness studies in `judgment-pack-evaluator-experiments`
+(studies 001–003) and spec RFC 0007. Protocol: MCP `prompts/list` and `prompts/get`, capability
+`prompts` advertised at initialize. Follows ADR-0002 (one core), ADR-0003 (MCP surface), ADR-0006
+(authoring in the client), ADR-0007 (the experimental evaluator the `test_pack` prompt drives).

docs/adr/README.md

@@ -36,3 +36,4 @@ when** it became that way. architecture.md links out to the relevant ADR rather
 | [0005](0005-single-jps-diagnostic-code-prefix.md)        | Use a single `JPS-` prefix for diagnostic codes                                 | accepted |
 | [0006](0006-authoring-lifecycle-in-the-client.md)        | The authoring lifecycle lives in the client; the runtime is a stateless oracle  | accepted |
 | [0007](0007-experimental-evaluator.md)                   | Ship an experimental evaluator behind an explicit experimental surface          | accepted |
+| [0008](0008-mcp-prompts-authoring-method.md)             | Serve authoring method as MCP prompts; the intelligence stays in the client     | proposed |

docs/mcp-clients.md

@@ -27,6 +27,22 @@ gitignored in this repository. Copy a snippet, don't commit one.
 None of these evaluate, decide, or authorize anything, except `experimental_evaluate`, which
 evaluates experimentally and says so in every payload.
 
+## The prompts
+
+The server also serves three **method prompts** (MCP `prompts` capability) — static, versioned,
+non-normative guidance that your client's model executes with your key. In Claude Code they appear
+as slash commands (`/mcp__judgment-pack__author_pack`, `…test_pack`, `…fix_pack`); other clients
+surface them differently or not at all, and everything works without them.
+
+| Prompt | Guides |
+| --- | --- |
+| `author_pack` | Encoding one policy decision: the create → validate → evaluate loop, the resolution-model shapes that avoid conflicts, the decimal-string and `onUnknown` rules, the prepared-facts ledger |
+| `test_pack` | Probing a pack's logic with an instance matrix (per-outcome, conflict, unknown, missing-evidence, not-applicable, forced-outcome, ordered-comparison rows) |
+| `fix_pack` | Repairing a non-conformant pack from the validator's diagnostics, in carrier → structural → semantic order |
+
+Following a prompt does not make a pack conformant — only validation decides that — and the
+documents you produce are yours; the runtime stores and interprets nothing (ADR-0008).
+
 ## Claude Code
 
 One command, from any directory: