@iowarp/clio-coder 0.4.3 → 0.4.4

package/src/domains/providers/models/local-models/clio-coder-local-coding-targets.yaml

@@ -1715,15 +1715,16 @@
         card's dial is reasoning strength with levels low, medium, high, and
         xhigh, set through the system prompt as "Reasoning strength: <value>",
         and the router passes it as chat_template_kwargs.reasoning_strength
-        pinned to xhigh. There is no off level on that scale, and no
-        ThinkingMechanism in this engine emits that key:
-        resolveRequestCapability populates chat_template_kwargs only with
-        enable_thinking or, on the harmony path, reasoning_effort. Encoding this
-        family as on-off would put a flag on the wire that this template does
-        not read while the TUI showed a working off/on dial, which is the
-        failure qwopus3.5-9b-v3 above was reclassified to stop. A real dial for
-        this family needs a per-family kwarg key in the engine, not a catalog
-        edit.
+        pinned to xhigh. There is no off level on that scale, so the family
+        stays always-on: encoding it as on-off would put an enable_thinking
+        flag on the wire that this template does not read while the TUI showed
+        a working off/on dial, which is the failure qwopus3.5-9b-v3 above was
+        reclassified to stop. The strength dial itself is driven per request:
+        thinking.chatTemplateKwargs.byLevel below maps Clio's thinking level to
+        chat_template_kwargs.reasoning_strength on llama.cpp, and LM Studio,
+        which drops chat_template_kwargs, reports that kwarg as undeliverable
+        instead of sending it. No per-level output has been measured yet, so
+        the always-on classification above stands until one is.
 
         Sampler provenance: temperature 1.0, top_p 0.95 and top_k 64 are the
         card's Best Practices values and the router preset uses the same three.
@@ -1743,7 +1744,7 @@
     gpuTiers:
       "32gb": "Agent-role target on a 32 GB card: UD-Q4_K_XL weights, q8_0 KV at ctx=131072, parallel=1, with the Q8_0 perception encoder and the DFlash drafter both resident."
     runtimePreference:
-      llamaCpp: "OpenAI-compatible chat completions with --jinja. Set the reasoning budget at launch with --chat-template-kwargs {\"reasoning_strength\": \"...\"}; there is no per-request field Clio can send for it."
+      llamaCpp: "OpenAI-compatible chat completions with --jinja. Clio drives the reasoning budget per request through chat_template_kwargs.reasoning_strength, mapped from the thinking level by thinking.chatTemplateKwargs.byLevel; a --chat-template-kwargs launch value is only the server default for requests that send none."
       openaiCompat: "Use when a gateway fronts the same llama.cpp chat-completions surface. Expect a reasoning trace on every response whatever the configured thinking level says."
     llamaCpp:
       ctxSize: 131072
@@ -1762,7 +1763,7 @@
       mechanism: always-on
       chatTemplateKwargs:
         byLevel:
-          # Card: reasoning strength low|medium|high|xhigh (scratchpad/catalog/models/unsloth-Muse-Glimmer-30B.md:35). The per-family kwarg key the note above asked for.
+          # Card: reasoning strength low|medium|high|xhigh (scratchpad/catalog/models/unsloth-Muse-Glimmer-30B.md:35). The per-family kwarg key the note above describes.
           key: reasoning_strength
           values:
             low: low
@@ -1774,9 +1775,10 @@
       guidance: |
         Muse Glimmer reasons on every turn and its strength scale has no off
         level, so Clio shows the dial as forced and budgets for a reasoning
-        preamble ahead of every answer. Change the amount of thinking at the
-        server, with the reasoning_strength chat-template kwarg; a per-request
-        thinking level cannot reach this model.
+        preamble ahead of every answer. The thinking level still sets how much
+        it thinks: on llama.cpp Clio sends reasoning_strength (low, medium,
+        high, xhigh) as a chat-template kwarg per request. LM Studio drops that
+        kwarg, so there the strength stays at the server's own default.
     serving: "Meta Superintelligence Lab Muse-Glimmer-30B, a dense 29.6B causal transformer with a ~1.8B ViT-G/14 perception encoder, interleaved local and global attention with a 2048 sliding window, and a DFlash block-diffusion drafter that proposes 16 tokens per forward pass. Card context length is 131072. Input is text plus image; the mini router additionally tags video, which the card does not claim, so the catalog follows the card."
 
 # ThinkingCap is a brevity finetune of Qwen3.6-27B and not the base model, so it

package/src/domains/providers/runtimes/claude/claude-code.ts

@@ -23,6 +23,15 @@ const claudeCodeRuntime: RuntimeDescriptor = {
 	headlessCommand: "claude -p --output-format stream-json",
 	outputParser: "claude-code-stream-json",
 	defaultCapabilities: claudeCodeCapabilities,
+	// The Claude CLI runs its own tool loop. Clio observes only the stream, so a
+	// declared per-tool budget is recorded as external-one-shot, never enforced.
+	externalAgentLoop: {
+		tools: "externally-governed-unobserved",
+		network: "externally-governed-unobserved",
+		budget: "external-one-shot",
+		generatingRetry: "allowed",
+		modelCatalog: "static",
+	},
 	synthesizeModel(target: TargetDescriptor, wireModelId: string, kb: KnowledgeBaseHit | null): Model<Api> {
 		return synthesizeClaudeDelegatedModel({
 			target,

package/src/domains/providers/support.ts

@@ -95,7 +95,7 @@ export function supportGroupLabel(group: ProviderSupportGroup): string {
 
 function classifyGroup(runtime: RuntimeDescriptor): ProviderSupportGroup {
 	if (runtime.id === "openai-codex") return "featured";
-	if (runtime.externalAgentLoop !== undefined) return "external-worker";
+	if (runtime.externalAgentLoop !== undefined && runtime.auth !== "claude-cli") return "external-worker";
 	if (runtime.id === "alcf") return "cloud-api";
 	if (runtime.auth === "oauth" || runtime.auth === "claude-cli") return "subscription";
 	if (catalogProviderForRuntime(runtime.id) || (runtime.auth === "api-key" && !runtime.probe)) {

package/src/domains/providers/types/runtime-descriptor.ts

@@ -171,7 +171,7 @@ export interface RuntimeDescriptor {
 		tools: "externally-governed-unobserved";
 		network: "externally-governed-unobserved";
 		budget: "external-one-shot";
-		generatingRetry: "forbidden";
+		generatingRetry: "forbidden" | "allowed";
 		modelCatalog: "live-authoritative" | "static";
 	};
 	/**

package/src/domains/safety/action-classifier.ts

@@ -84,6 +84,7 @@ function isUnderPrefix(abs: string, prefix: string): boolean {
 
 function baseClassify(tool: string): ActionClass | null {
 	switch (tool) {
+		case ToolNames.Evidence:
 		case ToolNames.Read:
 		case ToolNames.Grep:
 		case ToolNames.Find:
@@ -110,6 +111,12 @@ function baseClassify(tool: string): ActionClass | null {
 		// or survives the session, so it is reversible, local, and never gated
 		// behind a confirmation.
 		case ToolNames.Panes:
+		// limitation appends a typed receipt to the session ledger and touches
+		// nothing else, so it is never gated behind a confirmation.
+		case ToolNames.Limitation:
+		// decide appends the model's own design decision to the session
+		// decision board. One ledger append, no workspace effect, never gated.
+		case ToolNames.Decide:
 			return "read";
 		case ToolNames.Write:
 		case ToolNames.Edit:

package/src/domains/safety/default-path-policy.ts

@@ -67,8 +67,10 @@ export const DEFAULT_DAMAGE_CONTROL_PATH_POLICY: PathPolicyInput = {
 		// widen its own permissions in two tool calls. Both are authored by the
 		// operator through the CLI (`clio-coder verifiers author` lands nothing
 		// before `--yes`), so model-side writes and deletes are refused.
+		// Skills also refuse model-side writes and deletes; installs go through `clio-coder skills install`.
 		".clio-coder/safety.yaml",
 		".clio-coder/verifiers.yaml",
+		".clio-coder/skills/",
 		"/etc/",
 		"/usr/",
 		"/bin/",

package/src/domains/safety/finish-contract-registration.ts

@@ -4,16 +4,22 @@
  * The chat-loop fires `turn_end` with the final assistant text and this
  * registration emits an advisory `inject_reminder` (severity "warn") when the
  * turn mutated workspace state but recorded no validation evidence and no
- * explicit limitation. The trigger is the observed mutation, not the wording of
- * the assistant's text. The chat-loop's generic effect application renders the
+ * `limitation` receipt. The trigger is the observed mutation, not the wording
+ * of the assistant's text, and the text never enters the assessment. The chat-loop's generic effect application renders the
  * notice, persists the session entry, and flushes the reminder into the next
  * model request. Every decision is also written to the audit ledger.
  */
 
 import { VERIFICATION_SCRIPT_FAMILY_HINT } from "../../core/verification-scripts.js";
 import type { MiddlewareEffect, MiddlewareHookInput, MiddlewareHookRegistration } from "../middleware/index.js";
+import type { UserTaskAcceptance } from "../user-tasks/acceptance.js";
 import type { CompletionContractAuditInput } from "./audit.js";
-import { assessFinishContract, type FinishContractAssessment } from "./finish-contract.js";
+import {
+	assessFinishContract,
+	DEFAULT_RECENT_ENTRY_LIMIT,
+	type FinishContractAssessment,
+	recentEntries,
+} from "./finish-contract.js";
 import type { Rigor } from "./rigor.js";
 
 export const FINISH_CONTRACT_REGISTRATION_ID = "assessor.finish-contract";
@@ -21,15 +27,15 @@ export const FINISH_CONTRACT_REGISTRATION_ID = "assessor.finish-contract";
 /**
  * High-rigor re-prompt directive. Injected dynamically via effects (never added
  * to the static system prompt) so the prompt prefix stays byte-stable. It tells
- * the model to validate with a verification-family command or to state what
- * could not be verified before claiming done. The command hint mirrors the
+ * the model to validate with a verification-family command or to record a
+ * `limitation` receipt before claiming done. The command hint mirrors the
  * vocabulary `detectValidationCommand` and `isVerificationScriptName` accept.
  */
 export const HIGH_RIGOR_REVALIDATION_MESSAGE =
 	`[Clio Coder] high-rigor finish gate: this completion claim has no validation evidence. ` +
 	`Before claiming done, run a verification command (the ${VERIFICATION_SCRIPT_FAMILY_HINT} family, ` +
-	`e.g. "npm run test", "npm run lint", "npm run build") or explicitly state what could not be verified ` +
-	`and why. Do not end the turn until you have validated or recorded the limitation.`;
+	`e.g. "npm run test", "npm run lint", "npm run build") or call the limitation tool with the scope ` +
+	`and reason of what could not be verified. Do not end the turn until you have validated or recorded the limitation.`;
 
 export interface CreateFinishContractRegistrationOptions {
 	/**
@@ -45,6 +51,7 @@ export interface CreateFinishContractRegistrationOptions {
 	 * limitation instead of merely warning.
 	 */
 	resolveRigor?: () => Rigor;
+	readActiveAcceptance?: (window: ReadonlyArray<unknown>) => UserTaskAcceptance | undefined;
 	/**
 	 * Record the contract's decision to the audit ledger. Optional; when wired
 	 * (production passes the safety audit sink), every turn_end decision — each
@@ -59,7 +66,7 @@ export function createFinishContractRegistration(
 ): MiddlewareHookRegistration {
 	return {
 		id: FINISH_CONTRACT_REGISTRATION_ID,
-		description: "advise when a turn mutated files without validation evidence or an explicit limitation",
+		description: "advise when a turn mutated files without validation evidence or a limitation receipt",
 		hooks: ["turn_end"],
 		evaluate(input: MiddlewareHookInput, context): ReadonlyArray<MiddlewareEffect> {
 			if (input.hook !== "turn_end") return [];
@@ -70,8 +77,9 @@ export function createFinishContractRegistration(
 			// finalAssistantStopMessage.
 			const stopReason = input.metadata?.stopReason;
 			if (stopReason !== undefined && stopReason !== "stop") return [];
-			const assistantText = input.text?.trim() ?? "";
-			if (assistantText.length === 0) return [];
+			// An empty final text is not a completion claim. The assessor reads
+			// only ledger receipts, so the text is otherwise unused here.
+			if ((input.text?.trim() ?? "").length === 0) return [];
 			let entries: ReadonlyArray<unknown> | null;
 			try {
 				entries = options.readSessionEntries();
@@ -79,21 +87,28 @@ export function createFinishContractRegistration(
 				return [];
 			}
 			if (entries === null) return [];
+			const rigor = options.resolveRigor?.() ?? "normal";
+			const activeAcceptance =
+				rigor === "high"
+					? options.readActiveAcceptance?.(recentEntries(entries, input.turnId ?? null, DEFAULT_RECENT_ENTRY_LIMIT))
+					: undefined;
 			const assessment = assessFinishContract({
-				assistantText,
 				sessionEntries: entries,
+				workspaceRoot: process.cwd(),
+				rigor,
+				...(activeAcceptance ? { activeAcceptance } : {}),
 				assistantTurnId: input.turnId ?? null,
 			});
-			const rigor = options.resolveRigor?.() ?? "normal";
 			recordDecision(options, input.turnId ?? null, assessment, rigor);
 			if (assessment.kind !== "engage") return [];
 			if (rigor === "high") {
 				// Withhold the completion and force a re-prompt: a continuation
 				// request carries the turn onward, and the paired reminder gives
 				// the directive its own visible system-reminder line.
+				const message = activeAcceptance?.verification.length ? assessment.message : HIGH_RIGOR_REVALIDATION_MESSAGE;
 				return [
-					{ kind: "request_continuation", message: HIGH_RIGOR_REVALIDATION_MESSAGE },
-					{ kind: "inject_reminder", message: HIGH_RIGOR_REVALIDATION_MESSAGE, severity: "warn" },
+					{ kind: "request_continuation", message },
+					{ kind: "inject_reminder", message, severity: "warn" },
 				];
 			}
 			return [{ kind: "inject_reminder", message: assessment.message, severity: "warn" }];