@martinloop/mcp 0.7.0 → 0.8.0

package/dist/package-version.d.ts

@@ -1 +1 @@
-export declare const MARTIN_MCP_PACKAGE_VERSION = "0.7.0";
+export declare const MARTIN_MCP_PACKAGE_VERSION = "0.8.0";

package/dist/package-version.js

@@ -1,3 +1,3 @@
 // Auto-generated by build — do not edit by hand.
 // Source of truth: packages/mcp/package.json#version
-export const MARTIN_MCP_PACKAGE_VERSION = "0.7.0";
+export const MARTIN_MCP_PACKAGE_VERSION = "0.8.0";

package/dist/vendor/adapters/claude-cli.d.ts

@@ -22,8 +22,10 @@ import { type SpawnLike } from "./cli-bridge.js";
 export type CliArgsBuilder = (prompt: string, request: MartinAdapterRequest) => string[];
 export type CliStdinBuilder = (prompt: string) => string | undefined;
 export interface AgentCliAdapterOptions {
-    /** The executable to spawn (e.g. "claude", "codex"). */
+    /** Stable provider identity (e.g. "claude", "codex"). */
     command: string;
+    /** Exact executable to spawn when it differs from the provider identity. */
+    executionCommand?: string;
     /** Converts a prompt string into the argv array passed to spawn(). */
     argsBuilder: CliArgsBuilder;
     /** Optional stdin payload for CLIs that accept prompt input via stdin or `-`. */
@@ -57,6 +59,10 @@ export interface AgentCliAdapterOptions {
      * rather than only learning about an overspend after the process exits.
      */
     streamingUsageCap?: boolean;
+    /** Terminate after streamed provider usage reports request.context.remainingTokens was exceeded. */
+    streamingTokenCap?: boolean;
+    /** Provider detail-token fields are already included in input/output token totals. */
+    streamingUsageDetailsIncludedInTotals?: boolean;
     /** Test-only override for subprocess spawning. */
     spawnImpl?: SpawnLike;
 }
@@ -69,6 +75,8 @@ export interface ClaudeCliAdapterOptions {
     label?: string;
     /** Override the model passed via --model flag. */
     model?: string;
+    /** Enforce Claude Code's non-mutating plan permission mode. */
+    readOnly?: boolean;
     /** Extra args appended after core args (before prompt). */
     extraArgs?: string[];
     spawnImpl?: SpawnLike;
@@ -107,6 +115,8 @@ export interface GeminiCliAdapterOptions {
     label?: string;
     /** Explicit model override passed via --model. Omitted to preserve Gemini Auto. */
     model?: string;
+    /** Enforce Gemini's non-mutating plan approval mode. */
+    readOnly?: boolean;
     /** Approval mode for headless Gemini runs. Defaults to yolo for autonomous execution. */
     approvalMode?: "default" | "auto_edit" | "yolo" | "plan";
     /** Enable Gemini sandbox mode when the host is configured for it. Disabled by default. */

package/dist/vendor/adapters/claude-cli.js

@@ -24,24 +24,76 @@ import { createAdapterCapabilities, normalizeStructuredErrors, normalizeUsage }
 // ---------------------------------------------------------------------------
 const BLENDED_INPUT_COST_PER_1K = 0.003; // $/1K input tokens
 const BLENDED_OUTPUT_COST_PER_1K = 0.012; // $/1K output tokens
-// USD per 1K tokens. Claude cache creation uses the documented default
-// five-minute rate; cache reads are priced independently from fresh input.
+const PRICING_SNAPSHOT_2026_10_04 = "official-provider-pricing@2026-10-04";
+// USD per 1K tokens. Rates are standard/on-demand text-token prices from
+// official provider pricing pages as of 2026-10-04. Long-context tiers are
+// selected from the estimated/observed input-token count where providers
+// publish a threshold.
 const MODEL_PRICING = {
-    "claude-opus-4-6": { inputPer1K: 0.005, cachedInputPer1K: 0.0005, cacheCreationInputPer1K: 0.00625, outputPer1K: 0.025 },
-    "claude-sonnet-4-6": { inputPer1K: 0.003, cachedInputPer1K: 0.0003, cacheCreationInputPer1K: 0.00375, outputPer1K: 0.015 },
-    "claude-haiku-4-5": { inputPer1K: 0.001, cachedInputPer1K: 0.0001, cacheCreationInputPer1K: 0.00125, outputPer1K: 0.005 },
-    // OpenAI coding models
-    "codex": { inputPer1K: 0.00125, cachedInputPer1K: 0.000125, outputPer1K: 0.01 },
-    "gpt-5-codex": { inputPer1K: 0.00125, cachedInputPer1K: 0.000125, outputPer1K: 0.01 },
-    "gpt-5.1-codex": { inputPer1K: 0.00125, cachedInputPer1K: 0.000125, outputPer1K: 0.01 },
-    "gpt-5.1-codex-max": { inputPer1K: 0.00125, cachedInputPer1K: 0.000125, outputPer1K: 0.01 },
+    // Anthropic current + supported legacy models.
+    "claude-fable-5-1": { inputPer1K: 0.010, cachedInputPer1K: 0.00025, cacheCreationInputPer1K: 0.0125, outputPer1K: 0.050, pricingVersion: PRICING_SNAPSHOT_2026_10_04 },
+    "claude-opus-5-5": { inputPer1K: 0.004, cachedInputPer1K: 0.0002, cacheCreationInputPer1K: 0.005, outputPer1K: 0.020, pricingVersion: PRICING_SNAPSHOT_2026_10_04 },
+    "claude-sonnet-5-5": { inputPer1K: 0.002, cachedInputPer1K: 0.0002, cacheCreationInputPer1K: 0.0025, outputPer1K: 0.010, pricingVersion: PRICING_SNAPSHOT_2026_10_04 },
+    "claude-opus-4-6": { inputPer1K: 0.005, cachedInputPer1K: 0.0005, cacheCreationInputPer1K: 0.00625, outputPer1K: 0.025, pricingVersion: PRICING_SNAPSHOT_2026_10_04 },
+    "claude-sonnet-4-6": { inputPer1K: 0.003, cachedInputPer1K: 0.0003, cacheCreationInputPer1K: 0.00375, outputPer1K: 0.015, pricingVersion: PRICING_SNAPSHOT_2026_10_04 },
+    "claude-haiku-4-5": { inputPer1K: 0.001, cachedInputPer1K: 0.0001, cacheCreationInputPer1K: 0.00125, outputPer1K: 0.005, pricingVersion: PRICING_SNAPSHOT_2026_10_04 },
+    // OpenAI / Codex. GPT-6-family long-context pricing applies above 272K input tokens.
+    "gpt-6-astra": {
+        inputPer1K: 0.010, cachedInputPer1K: 0.001, cacheCreationInputPer1K: 0.0125, outputPer1K: 0.050,
+        pricingVersion: PRICING_SNAPSHOT_2026_10_04,
+        longContext: { thresholdInputTokens: 272_000, inputPer1K: 0.020, cachedInputPer1K: 0.002, cacheCreationInputPer1K: 0.025, outputPer1K: 0.075 }
+    },
+    "gpt-6.1-sol": {
+        inputPer1K: 0.002, cachedInputPer1K: 0.0001, cacheCreationInputPer1K: 0.0025, outputPer1K: 0.010,
+        pricingVersion: PRICING_SNAPSHOT_2026_10_04,
+        longContext: { thresholdInputTokens: 272_000, inputPer1K: 0.004, cachedInputPer1K: 0.0002, cacheCreationInputPer1K: 0.005, outputPer1K: 0.015 }
+    },
+    "gpt-6-sol": {
+        inputPer1K: 0.002, cachedInputPer1K: 0.0002, cacheCreationInputPer1K: 0.0025, outputPer1K: 0.010,
+        pricingVersion: PRICING_SNAPSHOT_2026_10_04,
+        longContext: { thresholdInputTokens: 272_000, inputPer1K: 0.004, cachedInputPer1K: 0.0004, cacheCreationInputPer1K: 0.005, outputPer1K: 0.015 }
+    },
+    "gpt-6-luna": {
+        inputPer1K: 0.0001, cachedInputPer1K: 0.00001, cacheCreationInputPer1K: 0.000125, outputPer1K: 0.0005,
+        pricingVersion: PRICING_SNAPSHOT_2026_10_04,
+        longContext: { thresholdInputTokens: 272_000, inputPer1K: 0.0002, cachedInputPer1K: 0.00002, cacheCreationInputPer1K: 0.00025, outputPer1K: 0.00075 }
+    },
+    "gpt-5.6-sol": {
+        inputPer1K: 0.004, cachedInputPer1K: 0.0004, cacheCreationInputPer1K: 0.005, outputPer1K: 0.020,
+        pricingVersion: PRICING_SNAPSHOT_2026_10_04,
+        longContext: { thresholdInputTokens: 272_000, inputPer1K: 0.008, cachedInputPer1K: 0.0008, cacheCreationInputPer1K: 0.010, outputPer1K: 0.030 }
+    },
+    "gpt-5.3-codex": { inputPer1K: 0.00175, cachedInputPer1K: 0.000175, outputPer1K: 0.014, pricingVersion: PRICING_SNAPSHOT_2026_10_04 },
+    "codex": { inputPer1K: 0.00125, cachedInputPer1K: 0.000125, outputPer1K: 0.010 },
+    "gpt-5-codex": { inputPer1K: 0.00125, cachedInputPer1K: 0.000125, outputPer1K: 0.010 },
+    "gpt-5.1-codex": { inputPer1K: 0.00125, cachedInputPer1K: 0.000125, outputPer1K: 0.010 },
+    "gpt-5.1-codex-max": { inputPer1K: 0.00125, cachedInputPer1K: 0.000125, outputPer1K: 0.010 },
     "gpt-5.2-codex": { inputPer1K: 0.00175, cachedInputPer1K: 0.000175, outputPer1K: 0.014 },
-    "codex-mini-latest": { inputPer1K: 0.0015, cachedInputPer1K: 0.000375, outputPer1K: 0.006 }
+    "codex-mini-latest": { inputPer1K: 0.0015, cachedInputPer1K: 0.000375, outputPer1K: 0.006 },
+    // Google Gemini current + common supported models.
+    "gemini-3.8-flash": { inputPer1K: 0.00075, cachedInputPer1K: 0.000075, outputPer1K: 0.00375, pricingVersion: PRICING_SNAPSHOT_2026_10_04 },
+    "gemini-3.7-flash": { inputPer1K: 0.00075, cachedInputPer1K: 0.000075, outputPer1K: 0.00375, pricingVersion: PRICING_SNAPSHOT_2026_10_04 },
+    "gemini-3.5-flash-lite": { inputPer1K: 0.0003, cachedInputPer1K: 0.00003, outputPer1K: 0.0025, pricingVersion: PRICING_SNAPSHOT_2026_10_04 },
+    "gemini-3.1-flash-lite": { inputPer1K: 0.00025, cachedInputPer1K: 0.000025, outputPer1K: 0.0015, pricingVersion: PRICING_SNAPSHOT_2026_10_04 },
+    "gemini-3.1-pro-preview": {
+        inputPer1K: 0.002, cachedInputPer1K: 0.0002, outputPer1K: 0.012, pricingVersion: PRICING_SNAPSHOT_2026_10_04,
+        longContext: { thresholdInputTokens: 200_000, inputPer1K: 0.004, cachedInputPer1K: 0.0004, outputPer1K: 0.018 }
+    },
+    "gemini-3-flash-preview": { inputPer1K: 0.0005, cachedInputPer1K: 0.00005, outputPer1K: 0.003, pricingVersion: PRICING_SNAPSHOT_2026_10_04 },
+    "gemini-2.5-pro": {
+        inputPer1K: 0.00125, cachedInputPer1K: 0.000125, outputPer1K: 0.010, pricingVersion: PRICING_SNAPSHOT_2026_10_04,
+        longContext: { thresholdInputTokens: 200_000, inputPer1K: 0.0025, cachedInputPer1K: 0.00025, outputPer1K: 0.015 }
+    },
+    "gemini-2.5-flash": { inputPer1K: 0.0003, cachedInputPer1K: 0.00003, outputPer1K: 0.0025, pricingVersion: PRICING_SNAPSHOT_2026_10_04 },
+    "gemini-2.5-flash-lite": { inputPer1K: 0.0001, cachedInputPer1K: 0.00001, outputPer1K: 0.0004, pricingVersion: PRICING_SNAPSHOT_2026_10_04 }
 };
 const CLAUDE_MODEL_ALIASES = [
-    { canonicalModelId: "claude-opus-4-6", aliases: [/^claude-opus-4-6-\d{8}$/u, /^claude-opus$/u] },
-    { canonicalModelId: "claude-sonnet-4-6", aliases: [/^claude-sonnet-4-6-\d{8}$/u, /^claude-sonnet$/u] },
-    { canonicalModelId: "claude-haiku-4-5", aliases: [/^claude-haiku-4-5-\d{8}$/u, /^claude-haiku$/u] }
+    { canonicalModelId: "claude-fable-5-1", aliases: [/^claude-fable-5-1-\d{8}$/u, /^claude-fable$/u, /^fable$/u] },
+    { canonicalModelId: "claude-opus-5-5", aliases: [/^claude-opus-5-5-\d{8}$/u, /^claude-opus$/u, /^opus$/u] },
+    { canonicalModelId: "claude-sonnet-5-5", aliases: [/^claude-sonnet-5-5-\d{8}$/u, /^claude-sonnet$/u, /^sonnet$/u] },
+    { canonicalModelId: "claude-opus-4-6", aliases: [/^claude-opus-4-6-\d{8}$/u] },
+    { canonicalModelId: "claude-sonnet-4-6", aliases: [/^claude-sonnet-4-6-\d{8}$/u] },
+    { canonicalModelId: "claude-haiku-4-5", aliases: [/^claude-haiku-4-5-\d{8}$/u, /^claude-haiku$/u, /^haiku$/u] }
 ];
 function resolveModelPricing(modelLabel) {
     const normalized = modelLabel?.trim().toLowerCase();
@@ -63,13 +115,28 @@ function resolveModelPricing(modelLabel) {
     }
     return { status: "unknown" };
 }
-function calculateUsageCost(usage, pricing) {
-    return ((usage.inputTokens / 1000) * pricing.inputPer1K +
-        (usage.cachedInputTokens / 1000) * (pricing.cachedInputPer1K ?? pricing.inputPer1K) +
-        (usage.cacheCreationInputTokens / 1000) * (pricing.cacheCreationInputPer1K ?? pricing.inputPer1K) +
-        (usage.outputTokens / 1000) * pricing.outputPer1K);
+function effectivePricingForInput(pricing, inputTokens) {
+    const long = pricing.longContext;
+    if (!long || inputTokens <= long.thresholdInputTokens) {
+        return pricing;
+    }
+    return {
+        inputPer1K: long.inputPer1K,
+        cachedInputPer1K: long.cachedInputPer1K,
+        cacheCreationInputPer1K: long.cacheCreationInputPer1K,
+        outputPer1K: long.outputPer1K,
+        pricingVersion: pricing.pricingVersion
+    };
+}
+function calculateUsageCost(usage, pricing, contextInputTokens = usage.inputTokens + usage.cachedInputTokens + usage.cacheCreationInputTokens) {
+    const effective = effectivePricingForInput(pricing, contextInputTokens);
+    return ((usage.inputTokens / 1000) * effective.inputPer1K +
+        (usage.cachedInputTokens / 1000) * (effective.cachedInputPer1K ?? effective.inputPer1K) +
+        (usage.cacheCreationInputTokens / 1000) * (effective.cacheCreationInputPer1K ?? effective.inputPer1K) +
+        (usage.outputTokens / 1000) * effective.outputPer1K);
 }
 const EMBEDDED_PRICING_VERSION = "embedded-v1";
+// Source snapshot: https://developers.openai.com/api/docs/models/gpt-6.1-sol
 function extractClaudeObservedModel(stdout) {
     for (const line of stdout.split(/\r?\n/u)) {
         try {
@@ -135,7 +202,9 @@ function extractUsage(parsed, modelLabel, modelSource) {
                 : pricingResolution.pricing
                     ? "static_catalog"
                     : "none",
-            ...(pricingResolution.pricing ? { pricingVersion: EMBEDDED_PRICING_VERSION } : {}),
+            ...(pricingResolution.pricing
+                ? { pricingVersion: pricingResolution.pricing.pricingVersion ?? EMBEDDED_PRICING_VERSION }
+                : {}),
             rawUsageAvailable: true,
             settledAt: new Date().toISOString()
         }
@@ -192,14 +261,20 @@ function extractCodexJsonlResult(stdout, modelLabel) {
     const cachedInputTokens = latestTurnCompleted.usage.cached_input_tokens ?? 0;
     const outputTokens = latestTurnCompleted.usage.output_tokens ?? 0;
     const reasoningOutputTokens = latestTurnCompleted.usage.reasoning_output_tokens ?? 0;
-    const tokensIn = promptTokens + cachedInputTokens;
-    const tokensOut = outputTokens + reasoningOutputTokens;
-    const exactPricing = modelLabel ? MODEL_PRICING[modelLabel] : undefined;
+    // Codex exposes cached/reasoning token detail fields as subsets of the
+    // corresponding input/output totals. Do not add those details twice.
+    const tokensIn = promptTokens;
+    const tokensOut = outputTokens;
+    const pricingResolution = resolveModelPricing(modelLabel);
+    const exactPricing = pricingResolution.pricing;
     const pricing = exactPricing ?? MODEL_PRICING["codex"] ??
         { inputPer1K: BLENDED_INPUT_COST_PER_1K, outputPer1K: BLENDED_OUTPUT_COST_PER_1K };
-    const actualUsd = (promptTokens / 1000) * pricing.inputPer1K +
-        (cachedInputTokens / 1000) * (pricing.cachedInputPer1K ?? pricing.inputPer1K) +
-        (tokensOut / 1000) * pricing.outputPer1K;
+    const actualUsd = calculateUsageCost({
+        inputTokens: Math.max(promptTokens - cachedInputTokens, 0),
+        cachedInputTokens,
+        cacheCreationInputTokens: 0,
+        outputTokens
+    }, pricing, promptTokens);
     return {
         summary,
         usage: normalizeUsage({
@@ -221,7 +296,7 @@ function extractCodexJsonlResult(stdout, modelLabel) {
                 billingMode: "unknown",
                 modelSource: modelLabel ? "explicit_override" : "agent_default",
                 pricingSource: exactPricing ? "static_catalog" : "blended_fallback",
-                pricingVersion: EMBEDDED_PRICING_VERSION,
+                pricingVersion: exactPricing?.pricingVersion ?? EMBEDDED_PRICING_VERSION,
                 rawUsageAvailable: true,
                 settledAt: new Date().toISOString()
             }
@@ -270,11 +345,16 @@ function extractGeminiJsonResult(stdout, modelLabel) {
     }
     const tokensIn = promptTokens + cachedInputTokens;
     const tokensOut = outputTokens + reasoningOutputTokens;
-    const pricing = (modelLabel ? MODEL_PRICING[modelLabel] : undefined) ??
+    const pricingResolution = resolveModelPricing(modelLabel);
+    const exactPricing = pricingResolution.pricing;
+    const pricing = exactPricing ??
         { inputPer1K: BLENDED_INPUT_COST_PER_1K, outputPer1K: BLENDED_OUTPUT_COST_PER_1K };
-    const actualUsd = (promptTokens / 1000) * pricing.inputPer1K +
-        (cachedInputTokens / 1000) * (pricing.cachedInputPer1K ?? pricing.inputPer1K) +
-        (tokensOut / 1000) * pricing.outputPer1K;
+    const actualUsd = calculateUsageCost({
+        inputTokens: promptTokens,
+        cachedInputTokens,
+        cacheCreationInputTokens: 0,
+        outputTokens: tokensOut
+    }, pricing, promptTokens + cachedInputTokens);
     return {
         summary,
         usage: normalizeUsage({
@@ -295,15 +375,15 @@ function extractGeminiJsonResult(stdout, modelLabel) {
                 reasoningOutputTokens,
                 billingMode: "unknown",
                 modelSource: modelLabel ? "explicit_override" : "agent_default",
-                pricingSource: "blended_fallback",
-                pricingVersion: EMBEDDED_PRICING_VERSION,
+                pricingSource: exactPricing ? "static_catalog" : "blended_fallback",
+                pricingVersion: exactPricing?.pricingVersion ?? EMBEDDED_PRICING_VERSION,
                 rawUsageAvailable: true,
                 settledAt: new Date().toISOString()
             }
         })
     };
 }
-function createStreamingUsageInspector(capUsd, modelLabel, promptTokenEstimate) {
+function createStreamingUsageInspector(capUsd, modelLabel, promptTokenEstimate, capTokens, detailTokensIncludedInTotals = false, observedUsageGovernor) {
     let pricingResolution = resolveModelPricing(modelLabel);
     // Safety margin: terminate at 80% of cap to bound one-turn overshoot.
     // Without this, a single expensive turn can blow past the cap before the
@@ -333,54 +413,101 @@ function createStreamingUsageInspector(capUsd, modelLabel, promptTokenEstimate)
     let usageEventSeen = false;
     let firstChunkAt;
     let finalResult;
-    const checkBudgetExceeded = (terminate) => {
+    let finalObservationPublished = false;
+    const observedEventIdentities = new Set();
+    const checkBudgetExceeded = (terminate, final = false) => {
+        const cumulativeTokens = tokensIn + tokensOut;
+        if (observedUsageGovernor) {
+            if (final && finalObservationPublished)
+                return;
+            if (final)
+                finalObservationPublished = true;
+            try {
+                const decision = observedUsageGovernor({ cumulativeUsd, cumulativeTokens, turns, final });
+                if (decision.action === "terminate") {
+                    terminate(`Parent observed-usage governor terminated the subprocess: ${decision.reason}.`);
+                }
+            }
+            catch (error) {
+                terminate(`Parent observed-usage governor failed closed: ${error instanceof Error ? error.message : String(error)}.`);
+            }
+            return;
+        }
+        if (capTokens !== undefined && capTokens > 0 && cumulativeTokens > capTokens) {
+            terminate(`Streaming token lease exceeded after ${String(turns)} turn(s): observed ${String(cumulativeTokens)} tokens ` +
+                `surpassed the per-attempt token cap ${String(capTokens)}. Subprocess terminated to prevent additional token usage.`);
+            return;
+        }
         if (capUsd > 0 && cumulativeUsd > effectiveCapUsd) {
             terminate(`Streaming usage cap exceeded after ${String(turns)} turn(s): cumulative cost ~$${cumulativeUsd.toFixed(4)} ` +
                 `surpassed the per-attempt cap $${capUsd.toFixed(4)} (${String(Math.round(effectiveCapRatio * 100))}% threshold: $${effectiveCapUsd.toFixed(4)}). ` +
                 `Subprocess terminated to bound runaway overspend.`);
         }
     };
-    const extractUsageFromEvent = (event, terminate) => {
+    const extractUsageFromEvent = (event, terminate, final = false) => {
+        const nestedMessage = event.message && typeof event.message === "object"
+            ? event.message
+            : undefined;
+        const providerEventId = typeof event.id === "string"
+            ? event.id
+            : typeof nestedMessage?.id === "string" ? nestedMessage.id : undefined;
+        if (providerEventId) {
+            const identity = `${String(event.type ?? "event")}:${providerEventId}`;
+            if (observedEventIdentities.has(identity))
+                return;
+            observedEventIdentities.add(identity);
+        }
         if (event.type === "system" && event.subtype === "init" && typeof event.model === "string") {
             pricingResolution = resolveModelPricing(event.model);
         }
         // Check for authoritative total_cost_usd on ANY event — if Claude reports
         // cost exceeding cap, terminate immediately regardless of event type.
-        if (typeof event.total_cost_usd === "number" && event.total_cost_usd > 0) {
-            cumulativeUsd = event.total_cost_usd;
-            checkBudgetExceeded(terminate);
-            return;
-        }
+        const authoritativeCost = typeof event.total_cost_usd === "number" && event.total_cost_usd >= 0;
+        if (authoritativeCost)
+            cumulativeUsd = Math.max(cumulativeUsd, event.total_cost_usd);
         // Extract usage from any event shape that carries it:
         //   - { type: "assistant", message: { usage: { ... } } }  (original format)
         //   - { usage: { input_tokens, output_tokens, ... } }      (top-level usage)
         //   - { message: { usage: { ... } } }                      (nested without type check)
-        const usage = (event.message && typeof event.message === "object" && "usage" in event.message
-            ? event.message.usage
+        const usage = (nestedMessage && "usage" in nestedMessage
+            ? nestedMessage.usage
             : undefined) ??
             (event.usage && typeof event.usage === "object" ? event.usage : undefined);
         if (!usage || typeof usage !== "object") {
+            if (authoritativeCost || final)
+                checkBudgetExceeded(terminate, final);
             return;
         }
         const usageRecord = usage;
         const turnInputTokens = usageRecord.input_tokens ?? usageRecord.inputTokens ?? 0;
-        const turnCachedInputTokens = usageRecord.cache_read_input_tokens ?? usageRecord.cacheReadInputTokens ?? 0;
+        const turnCachedInputTokens = usageRecord.cached_input_tokens ?? usageRecord.cachedInputTokens ??
+            usageRecord.cache_read_input_tokens ?? usageRecord.cacheReadInputTokens ?? 0;
         const turnCacheCreationInputTokens = usageRecord.cache_creation_input_tokens ?? usageRecord.cacheCreationInputTokens ?? 0;
-        const turnTokensIn = turnInputTokens + turnCachedInputTokens + turnCacheCreationInputTokens;
-        const turnTokensOut = usageRecord.output_tokens ?? usageRecord.outputTokens ?? 0;
+        const turnTokensIn = detailTokensIncludedInTotals
+            ? turnInputTokens
+            : turnInputTokens + turnCachedInputTokens + turnCacheCreationInputTokens;
+        const turnOutputTokens = usageRecord.output_tokens ?? usageRecord.outputTokens ?? 0;
+        const turnReasoningOutputTokens = usageRecord.reasoning_output_tokens ?? usageRecord.reasoningOutputTokens ?? 0;
+        const turnTokensOut = detailTokensIncludedInTotals
+            ? turnOutputTokens
+            : turnOutputTokens + turnReasoningOutputTokens;
         if (turnTokensIn === 0 && turnTokensOut === 0) {
+            if (authoritativeCost || final)
+                checkBudgetExceeded(terminate, final);
             return;
         }
         const turnUsd = pricingResolution.pricing
             ? calculateUsageCost({
-                inputTokens: turnInputTokens,
+                inputTokens: detailTokensIncludedInTotals
+                    ? Math.max(turnInputTokens - turnCachedInputTokens - turnCacheCreationInputTokens, 0)
+                    : turnInputTokens,
                 cachedInputTokens: turnCachedInputTokens,
                 cacheCreationInputTokens: turnCacheCreationInputTokens,
                 outputTokens: turnTokensOut
             }, pricingResolution.pricing)
             : undefined;
         const remainingBudgetBeforeTurn = Math.max(capUsd - cumulativeUsd, 0);
-        if (turnUsd !== undefined &&
+        if (!observedUsageGovernor && !final && !authoritativeCost && turnUsd !== undefined &&
             capUsd > 0 &&
             remainingBudgetBeforeTurn > 0 &&
             turnUsd > remainingBudgetBeforeTurn * 0.5) {
@@ -397,10 +524,10 @@ function createStreamingUsageInspector(capUsd, modelLabel, promptTokenEstimate)
         tokensOut += turnTokensOut;
         turns += 1;
         usageEventSeen = true;
-        if (turnUsd !== undefined) {
+        if (!authoritativeCost && turnUsd !== undefined) {
             cumulativeUsd += turnUsd;
         }
-        checkBudgetExceeded(terminate);
+        checkBudgetExceeded(terminate, final);
     };
     const ingestLine = (line, terminate) => {
         const trimmed = line.trim();
@@ -414,10 +541,23 @@ function createStreamingUsageInspector(capUsd, modelLabel, promptTokenEstimate)
         catch {
             return;
         }
-        extractUsageFromEvent(event, terminate);
         if (event.type === "result") {
             finalResult = event;
+            // Claude's final result usage is an aggregate repeat of usage already
+            // emitted on assistant events. Count it only when no incremental usage
+            // was observed, otherwise token leases are falsely doubled.
+            if (!usageEventSeen) {
+                extractUsageFromEvent(event, terminate, true);
+            }
+            else {
+                if (typeof event.total_cost_usd === "number" && event.total_cost_usd >= 0) {
+                    cumulativeUsd = Math.max(cumulativeUsd, event.total_cost_usd);
+                }
+                checkBudgetExceeded(terminate, true);
+            }
+            return;
         }
+        extractUsageFromEvent(event, terminate);
     };
     return {
         onChunk: (chunk, terminate) => {
@@ -582,10 +722,12 @@ export function createAgentCliAdapter(options) {
             // watch per-turn usage events as they arrive and kill the subprocess the
             // instant projected spend crosses what remains — bounding the worst case
             // to roughly one turn's overshoot rather than the entire runaway session.
-            const streamingUsage = options.streamingUsageCap && request.context.remainingBudgetUsd > 0
-                ? createStreamingUsageInspector(request.context.remainingBudgetUsd, options.model ?? options.command, estimatedUsage.tokensIn)
+            const streamingUsage = options.streamingUsageCap &&
+                (request.context.remainingBudgetUsd > 0 ||
+                    (options.streamingTokenCap && (request.context.remainingTokens ?? 0) > 0))
+                ? createStreamingUsageInspector(request.context.remainingBudgetUsd, options.model ?? options.command, estimatedUsage.tokensIn, options.streamingTokenCap ? request.context.remainingTokens : undefined, options.streamingUsageDetailsIncludedInTotals, request.observedUsageGovernor)
                 : undefined;
-            const agentResult = await runSubprocess(options.command, args, {
+            const agentResult = await runSubprocess(options.executionCommand ?? options.command, args, {
                 cwd: workingDirectory,
                 timeoutMs: executionTimeoutMs,
                 spawnImpl: options.spawnImpl,
@@ -747,8 +889,12 @@ export function createAgentCliAdapter(options) {
             const verification = await runVerification(request.context.verificationPlan, workingDirectory, verifyTimeoutMs, verificationStack, options.spawnImpl, {
                 runId: request.loopId,
                 workspaceId: request.workspaceId,
+                attemptId: request.attemptId,
                 cwd: workingDirectory,
-            });
+                ...(request.context.runsRoot ? { runsRoot: request.context.runsRoot } : {}),
+                ...(request.context.executionProfile ? { executionProfile: request.context.executionProfile } : {}),
+                ...(request.context.allowedNetworkDomains?.length ? { allowedNetworkDomains: [...request.context.allowedNetworkDomains] } : {}),
+            }, request.signal);
             const verificationConfigured = verification.binding.commands.length > 0;
             // Check for zero-diff (agent ran but made no file changes)
             const postRunChangedFiles = gitRepoRoot
@@ -759,11 +905,12 @@ export function createAgentCliAdapter(options) {
             // Extract structured errors from stderr/stdout for better failure context
             const structuredErrors = normalizeStructuredErrors(extractStructuredErrors(agentResult.stderr, agentResult.stdout));
             const rawExecutionArtifacts = gitRepoRoot
-                ? await readGitExecutionArtifacts(gitRepoRoot, 5000, options.spawnImpl)
+                ? await readGitExecutionArtifacts(gitRepoRoot, 5000, options.spawnImpl, agentChangedFiles)
                 : undefined;
             const executionArtifacts = rawExecutionArtifacts
                 ? {
                     ...(agentChangedFiles.length > 0 ? { changedFiles: agentChangedFiles } : {}),
+                    ...(rawExecutionArtifacts.patch ? { patch: rawExecutionArtifacts.patch } : {}),
                     ...(baselineChangedFiles.size === 0 && rawExecutionArtifacts.diffStats
                         ? { diffStats: rawExecutionArtifacts.diffStats }
                         : {})
@@ -821,8 +968,11 @@ export function createAgentCliAdapter(options) {
             const scopeBlock = scopeViolations.length > 0
                 ? `\n  Scope violations: ${scopeViolations.join(", ")}`
                 : "";
-            // Write PROGRESS.md to help the next attempt re-anchor on the original objective
-            if (repoRoot) {
+            // PROGRESS.md is legacy adapter-local retry state. Never create it in a
+            // scope-constrained workspace: doing so manufactures an undeclared
+            // product change and causes the Core leash to hide the provider's real
+            // failure behind surface_write_not_allowed.
+            if (repoRoot && (request.context.allowedPaths?.length ?? 0) === 0) {
                 try {
                     const { writeFile, readFile, appendFile: appendFs } = await import("node:fs/promises");
                     const progressPath = `${repoRoot}/PROGRESS.md`;
@@ -882,6 +1032,33 @@ export function createAgentCliAdapter(options) {
 // ---------------------------------------------------------------------------
 // Pre-configured: Claude CLI
 // ---------------------------------------------------------------------------
+const CLAUDE_PERMISSION_CONTROL_ARGS = new Set([
+    "--allow-dangerously-skip-permissions",
+    "--allowed-tools",
+    "--allowedTools",
+    "--dangerously-skip-permissions",
+    "--permission-mode"
+]);
+function assertClaudePermissionControlsAreGoverned(extraArgs) {
+    const overridesPermissionControl = extraArgs.some((arg) => {
+        const [flag] = arg.split("=", 1);
+        return flag !== undefined && CLAUDE_PERMISSION_CONTROL_ARGS.has(flag);
+    });
+    if (overridesPermissionControl) {
+        throw new Error("Claude permission controls cannot be overridden via extraArgs.");
+    }
+}
+function hostOwnsVerification(request) {
+    return request.context.verificationExecutionOwner === "host_only";
+}
+function buildClaudeVerifierAllowedTools(request) {
+    if (hostOwnsVerification(request))
+        return [];
+    const commands = [...new Set(request.context.verificationPlan.filter((command) => command.trim().length > 0))];
+    return commands.length > 0
+        ? ["--allowedTools", ...commands.map((command) => `Bash(${command})`)]
+        : [];
+}
 /**
  * Spawns `claude --output-format stream-json --verbose --print "<prompt>" [extraArgs]`.
  *
@@ -900,6 +1077,7 @@ export function createAgentCliAdapter(options) {
 export function createClaudeCliAdapter(options = {}) {
     const modelArgs = options.model ? ["--model", options.model] : [];
     const extraArgs = options.extraArgs ?? [];
+    assertClaudePermissionControlsAreGoverned(extraArgs);
     return createAgentCliAdapter({
         command: "claude",
         adapterIdSuffix: "claude",
@@ -912,20 +1090,26 @@ export function createClaudeCliAdapter(options = {}) {
         verifyTimeoutMs: options.verifyTimeoutMs,
         supportsJsonOutput: true,
         streamingUsageCap: true,
+        streamingTokenCap: true,
         spawnImpl: options.spawnImpl,
         argsBuilder: (_prompt, request) => [
             "--output-format",
             "stream-json",
             "--verbose",
             "--print",
-            "--dangerously-skip-permissions",
+            ...(options.readOnly
+                ? ["--permission-mode", "plan"]
+                : ["--permission-mode", "acceptEdits"]),
+            ...buildClaudeVerifierAllowedTools(request),
             // Subprocess isolation strategy:
             // --bare: skips hooks (prevents SessionEnd/hook failures causing non-zero exits),
             //   MCP server loading, LSP, CLAUDE.md discovery, and background prefetches.
             //   Requires ANTHROPIC_API_KEY (OAuth/keychain auth not available in bare mode).
             // --strict-mcp-config: fallback when ANTHROPIC_API_KEY is not set — still
             //   prevents parent MCP servers from being inherited by the subprocess.
-            ...(process.env["ANTHROPIC_API_KEY"] ? ["--bare"] : ["--strict-mcp-config"]),
+            ...(process.env["ANTHROPIC_API_KEY"]
+                ? ["--bare"]
+                : ["--strict-mcp-config", "--setting-sources", "project", "--disable-slash-commands"]),
             // NOTE: --max-tokens does not exist in the claude CLI. Token cap enforcement
             // is handled at the MartinLoop layer via streamingUsageCap, not via subprocess flags.
             ...modelArgs,
@@ -988,10 +1172,13 @@ export function createCodexCliAdapter(options = {}) {
  *   npm install -g @google/gemini-cli
  */
 export function createGeminiCliAdapter(options = {}) {
-    const approvalMode = options.approvalMode ?? "yolo";
-    if (approvalMode !== "yolo") {
+    const approvalMode = options.readOnly ? "plan" : options.approvalMode ?? "yolo";
+    if (!options.readOnly && approvalMode !== "yolo") {
         throw new Error("governed-autonomous Gemini execution requires yolo approval mode; interactive downgrade rejected.");
     }
+    if (options.readOnly && options.approvalMode !== undefined && options.approvalMode !== "plan") {
+        throw new Error("read-only Gemini execution requires plan approval mode.");
+    }
     const extraArgs = options.extraArgs ?? [];
     if (extraArgs.some((arg) => arg === "--approval-mode" || arg.startsWith("--approval-mode="))) {
         throw new Error("Gemini approval mode cannot be overridden through extraArgs.");
@@ -1034,15 +1221,32 @@ export function createGeminiCliAdapter(options = {}) {
 function buildPrompt(request) {
     const lines = [];
     const mutationMode = request.context.mutationMode ?? "edit";
+    const verificationIsHostOwned = hostOwnsVerification(request);
     lines.push("You are running in autonomous agentic mode.");
-    lines.push("MAKE ALL REQUIRED FILE EDITS NOW. Do not ask for confirmation. Do not ask clarifying questions.");
-    lines.push("Do not explain what you found without also making the changes. Edit the files and complete the task.");
-    lines.push("");
-    lines.push("If PROGRESS.md exists in your working directory, read it first for context from prior attempts.");
-    lines.push("If it does not exist, proceed with the objective below.");
+    if (mutationMode === "read_only") {
+        lines.push("READ-ONLY EXECUTION. Do not create, edit, delete, rename, or move files.");
+        lines.push("Inspect and verify only; report evidence without changing the workspace.");
+    }
+    else {
+        lines.push("MAKE ALL REQUIRED FILE EDITS NOW. Do not ask for confirmation. Do not ask clarifying questions.");
+        lines.push("Do not explain what you found without also making the changes. Edit the files and complete the task.");
+    }
     lines.push("");
-    lines.push("Complete the following coding task. Make all necessary file changes.");
-    lines.push("When you are done, the verification commands listed below must pass.");
+    if ((request.context.allowedPaths?.length ?? 0) === 0) {
+        lines.push("If PROGRESS.md exists in your working directory, read it first for context from prior attempts.");
+        lines.push("If it does not exist, proceed with the objective below.");
+        lines.push("");
+    }
+    lines.push(mutationMode === "read_only"
+        ? "Complete the following inspection task without making file changes."
+        : "Complete the following coding task. Make all necessary file changes.");
+    if (verificationIsHostOwned) {
+        lines.push("MartinLoop owns verification and runs it after provider completion.");
+        lines.push("Do not execute these verifier commands yourself.");
+    }
+    else {
+        lines.push("When you are done, the verification commands listed below must pass.");
+    }
     lines.push("");
     lines.push("OBJECTIVE:");
     lines.push(sanitizeForPrompt(request.context.objective));
@@ -1068,7 +1272,9 @@ function buildPrompt(request) {
         lines.push("");
     }
     if (request.context.verificationPlan.length > 0) {
-        lines.push("VERIFICATION (all commands must exit with code 0):");
+        lines.push(verificationIsHostOwned
+            ? "HOST-OWNED VERIFICATION (success conditions only; do not execute):"
+            : "VERIFICATION (all commands must exit with code 0):");
         for (const cmd of request.context.verificationPlan) {
             lines.push(`  ${cmd}`);
         }
@@ -1116,6 +1322,9 @@ function buildPrompt(request) {
         lines.push("");
     }
     lines.push(`FOCUS: ${sanitizeForPrompt(request.context.focus)}`);
+    if (verificationIsHostOwned && request.context.verificationPlan.length > 0) {
+        lines.push("FINAL BOUNDARY: stop after the scoped work and report it; MartinLoop will execute verification.");
+    }
     return lines.join("\n");
 }
 // ---------------------------------------------------------------------------
@@ -1159,7 +1368,12 @@ function estimatePromptCost(promptText, model, providerCommand) {
     if (!pricing) {
         return undefined;
     }
-    return (inputTokens / 1000) * pricing.inputPer1K + (outputTokens / 1000) * pricing.outputPer1K;
+    return calculateUsageCost({
+        inputTokens,
+        cachedInputTokens: 0,
+        cacheCreationInputTokens: 0,
+        outputTokens
+    }, pricing, inputTokens);
 }
 function estimateUsage(promptText, model, providerCommand) {
     const inputTokens = Math.ceil(promptText.length / 3.5);