feat(routing): give Jev meaningful task briefs and host catalogs

CHANGELOG.md

@@ -1,5 +1,11 @@
 # Changelog
 
+## 0.17.1 — Unreleased
+
+- Supply Jev with controlled task briefs and distinct reviewed profile preferences.
+- Pin host-scoped catalogs; preserve unknown measurements and exclude unsupported profiles.
+- Keep model identities, arbitrary catalog prose and provenance local.
+
 ## 0.17.0 — Unreleased
 
 - Add schema-v5 dynamic routing with control-based approval and no study requirement.

docs/contributing-agents.md

@@ -147,7 +147,7 @@ workers or claim synthetic savings as observed results. See the
 
 ## Distribution and versioning
 
-The current package version is 0.17.0; increment all three manifests for further
+The current package version is 0.17.1; increment all three manifests for further
 installer-visible changes. Existing native Claude agents/commands/hooks remain
 preserved. Never rewrite historical plans to claim newer evidence. No repository
 change implicitly installs personally, publishes a release or edits the separate

docs/reviews/2026-09-27-task-routing-inputs.md

@@ -0,0 +1,16 @@
+# Task-level routing inputs
+
+Jev receives a controlled assignment brief and distinct profile preferences,
+with provenance and exact model identities retained locally. Catalog snapshots
+are bound to policy approval and compiled against current worker controls. Null
+measurements stay valid; partial observations never become complete metrics.
+Unknown context cannot meet an enforced context requirement. Claude starter
+aliases use documented qualitative descriptions; Codex discovers its roster from
+the active host, not a universal hard-coded model list.
+
+Offline fake-provider tests inspect the actual payload and demonstrate different
+same-role decisions for routine/intensive tasks on all three hosts. They verify
+privacy, injection rejection, catalog drift, unsupported settings, unknown
+capacity and measurement completeness. These tests establish plumbing, not live
+recommendation quality or token savings. Native and portable validation pass;
+no live models or personal configuration were used.

plugin/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "guildhall",
-  "version": "0.17.0",
+  "version": "0.17.1",
   "description": "The Guildhall \u2014 a gathering place for adventurers. A TDD-ordered coding agent harness for Claude Code, tuned for Opus-tier orchestration (Opus 5 recommended seat). The /quest slash command runs Mordain the Guildmaster, who writes a durable plan file, then dispatches 18 specialist adventurers across three tiers: Opus (architecture-reviewer, security-reviewer, reliability-reviewer, migration-safety-reviewer), Sonnet (test-author, feature-implementer, ui-test-author, docs-writer, pr-author, prototype-builder, debug-investigator, observability-reviewer, performance-reviewer, ops-readiness-reviewer, accessibility-reviewer), and Haiku (refactorer, plugin-validator, fog-cartographer). Post-green reviews fan out in parallel \u2014 two always-on (security, docs) plus six gated production-readiness reviewers (observability, reliability, performance, ops-readiness, migration-safety, accessibility) that fire only when their trigger matches the diff. Rook (pr-author) closes the quest with a platform-agnostic PR draft and folds the runbook into the body. Integrates with IDD-framework specs.",
   "author": {
     "name": "GrillerGeek"

plugin/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "guildhall",
-  "version": "0.17.0",
+  "version": "0.17.1",
   "author": {
     "name": "GrillerGeek"
   },

plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json",
   "name": "guildhall",
-  "version": "0.17.0",
+  "version": "0.17.1",
   "author": {
     "name": "GrillerGeek"
   },

plugin/portable/references/task-routing.md

@@ -0,0 +1,71 @@
+# Task-level routing inputs (schema v5)
+
+Dynamic routing is approved selection, not benchmark-proven improvement. Use
+`routing_catalog.py` to compile a reviewed catalog and actual host controls on
+JSON stdin: `{"catalog": <catalog-v1>, "host": <request-v5 host>}`. It returns
+supported candidates, excluded IDs and their `catalog_revision`; it does not
+write files, activate routing or call models. Set the policy's candidates and
+revision to this snapshot. Changes require a new policy preview and approval.
+
+## Catalogs and provenance
+
+The [catalog schema](../resources/schemas/catalog-v1.schema.json) uses complete
+host-scoped candidate records. Start with the matching resource:
+[Codex](../resources/catalogs/codex-skill.json),
+[native Claude](../resources/catalogs/claude-native.json), or
+[standalone Claude](../resources/catalogs/claude-skill.json).
+Codex's starter is deliberately empty: populate exact model/effort pairs and
+permitted values from the currently callable worker tool and current host metadata.
+Do not copy the setup author's personal model roster into another installation.
+No paid discovery or another installed CLI is necessary or authoritative.
+
+Claude's aliases are descriptive starting points. Their routine/extended/intensive
+and efficiency descriptors are qualitative priors based on the official
+[model configuration documentation](https://code.claude.com/docs/en/model-config),
+reviewed 2026-09-27. Verify actual supported aliases, forced settings and provider
+mappings on this host. These descriptions are not measured cost, capacity or
+quality. Context and capabilities remain unknown until supported by host metadata
+or documentation. Native frontmatter defaults stay unchanged. Standalone Claude
+must discover its own tool controls. Keep Fable excluded.
+
+Each `routing_profile` has controlled work types, reasoning depth, complexity,
+risk and efficiency preferences plus local source/revision. Use `documented`
+only when the linked source supports the description; use `user_preference` for
+reviewed judgments. No universal model ranking ships. Ask for the missing preference
+when metadata gives no meaningful distinction; do not label all profiles the same
+and promise useful routing. Presets do not authorize any models or roles.
+
+`facts_source` records where capacity/capabilities came from. Unknown facts remain
+null/empty and cannot satisfy an enforced context/capability constraint. A task
+with no established hard context minimum can use `context_bucket: unknown`;
+never change a known requirement to unknown just to make a candidate eligible.
+Unsupported settings are excluded. Recompile and obtain review when actual
+controls or the catalog changes. Catalog updates do not silently widen policy.
+
+Measurements remain optional and role/category-scoped. Schema v5 adds source,
+sample_count, completeness and revision; partial observations cannot satisfy
+numeric ceilings or become provider metrics. Unknown cost/quota remains unknown.
+Raw subscription tokens do not establish weighted allowance or monetary savings.
+
+## Task brief and privacy
+
+From the worker's permitted handoff, set role/category, ambiguity, risk, context
+bucket, reasoning depth, change breadth, expected output and verification needs.
+Each field has a controlled vocabulary in the
+[request schema](../resources/schemas/request-v5.schema.json); use `unknown` where
+needed. Test-author facts come only from its allowed Spec/API/test inputs. Do not
+read an implementation to classify that assignment. No extra classifier model
+call is needed. The same specialist may receive very different briefs.
+
+The `categories-v2` outbound contract sends these fields, objective, capability
+count, known capacity, complete scoped measurements and each profile's controlled
+preferences/basis. Each candidate gets different descriptive criteria when its
+reviewed preferences differ. Only request-local `p0`, `p1`, … labels leave the
+host; model names, catalog IDs, provenance URLs, paths, revisions and evidence
+hashes stay local. The router's own model selector is necessarily sent to Jev.
+Optional summary mode still needs approval of the exact text for each task.
+
+The TypeSafe [choice API](https://docs.typesafe.ai/introduction/quickstart) supports
+text state and per-choice criteria (reviewed 2026-09-27). Returned confidence is
+recorded, not treated as calibrated coding success. Receipts retain local task
+facts and catalog revision; do not invent a provider rationale or savings claim.

plugin/portable/resources/catalogs/claude-native.json

@@ -0,0 +1,183 @@
+{
+  "schema_version": 1,
+  "host": "claude-native",
+  "profiles": [
+    {
+      "id": "haiku",
+      "host": "claude-native",
+      "model": "haiku",
+      "effort": null,
+      "roles": [
+        "accessibility-reviewer",
+        "architecture-reviewer",
+        "debug-investigator",
+        "docs-writer",
+        "feature-implementer",
+        "fog-cartographer",
+        "migration-safety-reviewer",
+        "observability-reviewer",
+        "ops-readiness-reviewer",
+        "performance-reviewer",
+        "plugin-validator",
+        "pr-author",
+        "prototype-builder",
+        "refactorer",
+        "reliability-reviewer",
+        "security-reviewer",
+        "test-author",
+        "ui-test-author"
+      ],
+      "categories": [
+        "docs",
+        "pr",
+        "implementation",
+        "tests",
+        "debug",
+        "review",
+        "prototype",
+        "refactor"
+      ],
+      "capabilities": [],
+      "context_tokens": null,
+      "quality": null,
+      "latency_ms": null,
+      "cost_usd": null,
+      "usage_tokens": null,
+      "profile_revision": "starter-2026-09-27",
+      "qualification": null,
+      "measurements": [],
+      "facts_source": {
+        "kind": "unknown",
+        "reference": null
+      },
+      "routing_profile": {
+        "work_types": [],
+        "reasoning_depth": "routine",
+        "complexity": [],
+        "risk": [],
+        "efficiency": "low_overhead",
+        "basis": "documented",
+        "source": "https://code.claude.com/docs/en/model-config",
+        "revision": "2026-09-27"
+      }
+    },
+    {
+      "id": "sonnet",
+      "host": "claude-native",
+      "model": "sonnet",
+      "effort": null,
+      "roles": [
+        "accessibility-reviewer",
+        "architecture-reviewer",
+        "debug-investigator",
+        "docs-writer",
+        "feature-implementer",
+        "fog-cartographer",
+        "migration-safety-reviewer",
+        "observability-reviewer",
+        "ops-readiness-reviewer",
+        "performance-reviewer",
+        "plugin-validator",
+        "pr-author",
+        "prototype-builder",
+        "refactorer",
+        "reliability-reviewer",
+        "security-reviewer",
+        "test-author",
+        "ui-test-author"
+      ],
+      "categories": [
+        "docs",
+        "pr",
+        "implementation",
+        "tests",
+        "debug",
+        "review",
+        "prototype",
+        "refactor"
+      ],
+      "capabilities": [],
+      "context_tokens": null,
+      "quality": null,
+      "latency_ms": null,
+      "cost_usd": null,
+      "usage_tokens": null,
+      "profile_revision": "starter-2026-09-27",
+      "qualification": null,
+      "measurements": [],
+      "facts_source": {
+        "kind": "unknown",
+        "reference": null
+      },
+      "routing_profile": {
+        "work_types": [],
+        "reasoning_depth": "extended",
+        "complexity": [],
+        "risk": [],
+        "efficiency": "balanced",
+        "basis": "documented",
+        "source": "https://code.claude.com/docs/en/model-config",
+        "revision": "2026-09-27"
+      }
+    },
+    {
+      "id": "opus",
+      "host": "claude-native",
+      "model": "opus",
+      "effort": null,
+      "roles": [
+        "accessibility-reviewer",
+        "architecture-reviewer",
+        "debug-investigator",
+        "docs-writer",
+        "feature-implementer",
+        "fog-cartographer",
+        "migration-safety-reviewer",
+        "observability-reviewer",
+        "ops-readiness-reviewer",
+        "performance-reviewer",
+        "plugin-validator",
+        "pr-author",
+        "prototype-builder",
+        "refactorer",
+        "reliability-reviewer",
+        "security-reviewer",
+        "test-author",
+        "ui-test-author"
+      ],
+      "categories": [
+        "docs",
+        "pr",
+        "implementation",
+        "tests",
+        "debug",
+        "review",
+        "prototype",
+        "refactor"
+      ],
+      "capabilities": [],
+      "context_tokens": null,
+      "quality": null,
+      "latency_ms": null,
+      "cost_usd": null,
+      "usage_tokens": null,
+      "profile_revision": "starter-2026-09-27",
+      "qualification": null,
+      "measurements": [],
+      "facts_source": {
+        "kind": "unknown",
+        "reference": null
+      },
+      "routing_profile": {
+        "work_types": [],
+        "reasoning_depth": "intensive",
+        "complexity": [],
+        "risk": [],
+        "efficiency": "thorough",
+        "basis": "documented",
+        "source": "https://code.claude.com/docs/en/model-config",
+        "revision": "2026-09-27"
+      }
+    }
+  ]
+}