@mindstudio-ai/remy 0.1.312 → 0.1.313

package/dist/headless.d.ts

@@ -122,8 +122,9 @@ declare class HeadlessSession {
     /** Apply queued tool block updates to state.messages. Safe to call any time. */
     private applyPendingBlockUpdates;
     /**
-     * Forced compaction gate. If lastContextSize exceeds the threshold, compact
-     * before letting the upcoming turn run. Coalesces with any in-flight
+     * Forced compaction gate. If lastContextSize exceeds the upcoming turn's
+     * model threshold (per-model — see ModelContextLimits in models/surfaces),
+     * compact before letting the turn run. Coalesces with any in-flight
      * compaction (e.g., one already started by /compact or a tool call). No
      * timeout — compaction takes as long as it takes.
      *

package/dist/headless.js

@@ -486,37 +486,59 @@ var MODEL_SURFACES = {
     userPickable: false
   }
 };
+var DEFAULT_SUGGEST_COMPACT_AT = 3e5;
+var TEXT_MODELS = {
+  // Anthropic 1M-context, flat-priced.
+  "claude-5-opus": { forceCompactAt: 85e4 },
+  "claude-4-8-opus": { forceCompactAt: 85e4 },
+  "claude-4-7-opus": { forceCompactAt: 85e4 },
+  // Anthropic 1M-context with 2x long-context pricing above 200K input.
+  "claude-4-6-opus": { forceCompactAt: 85e4, suggestCompactAt: 18e4 },
+  "claude-4-6-sonnet": { forceCompactAt: 85e4, suggestCompactAt: 18e4 },
+  "claude-fable-5": { forceCompactAt: 85e4 },
+  "claude-fable-5-1": { forceCompactAt: 85e4 },
+  "claude-5-sonnet": { forceCompactAt: 85e4 },
+  // OpenAI gpt-5.5/5.6: ~1M window, but the usable input ceiling under
+  // `truncation: 'auto'` is ~794K (output + reasoning reserve), and all
+  // rates double above 272K input.
+  "gpt-5.5": { forceCompactAt: 6e5, suggestCompactAt: 25e4 },
+  "gpt-5.6-sol": { forceCompactAt: 6e5, suggestCompactAt: 25e4 },
+  "gpt-5.6-terra": { forceCompactAt: 6e5, suggestCompactAt: 25e4 },
+  "gpt-5.6-luna": { forceCompactAt: 6e5, suggestCompactAt: 25e4 },
+  // Google ~1M-context; only 3.1-pro is tiered (higher rates above 200K).
+  "gemini-3-pro": { forceCompactAt: 85e4 },
+  "gemini-3.1-pro": { forceCompactAt: 85e4, suggestCompactAt: 18e4 },
+  "gemini-3-flash": { forceCompactAt: 85e4 },
+  "gemini-3.5-flash": { forceCompactAt: 85e4 },
+  "gemini-3.7-flash": { forceCompactAt: 85e4 },
+  // 256K window; its 200K pricing tier sits above the gate, so no nudge.
+  "grok-build-0.1": { forceCompactAt: 18e4 },
+  "grok-4.5": { forceCompactAt: 4e5 },
+  // 500K window
+  "grok-4.6": { forceCompactAt: 4e5 },
+  // 500K window
+  "glm-5.2": { forceCompactAt: 85e4 },
+  "muse-spark-1.1": { forceCompactAt: 85e4 },
+  "kimi-k2-7-code": { forceCompactAt: 2e5 },
+  // 262K window
+  "kimi-k3": { forceCompactAt: 85e4 },
+  "deepseek-v4-flash-0731": { forceCompactAt: 85e4 },
+  "qwen3.8-2.4t-a95b-deepinfra": { forceCompactAt: 2e5 },
+  // 262K window
+  "qwen3.8-27b-deepinfra": { forceCompactAt: 2e5 },
+  // 262K window
+  "minimax-m3": { forceCompactAt: 42e4 }
+  // 524K window
+};
+var DEFAULT_CONTEXT_LIMITS = { forceCompactAt: 85e4 };
+function getContextLimits(modelId) {
+  return TEXT_MODELS[modelId] ?? DEFAULT_CONTEXT_LIMITS;
+}
+function getSuggestCompactAt(modelId) {
+  return getContextLimits(modelId).suggestCompactAt ?? DEFAULT_SUGGEST_COMPACT_AT;
+}
 var ALLOWED_MODELS_BY_TYPE = {
-  text: [
-    "claude-5-opus",
-    "claude-4-8-opus",
-    "claude-4-7-opus",
-    "claude-4-6-opus",
-    "claude-4-6-sonnet",
-    "claude-fable-5",
-    "claude-fable-5-1",
-    "claude-5-sonnet",
-    "gpt-5.5",
-    "gpt-5.6-sol",
-    "gpt-5.6-terra",
-    "gpt-5.6-luna",
-    "gemini-3-pro",
-    "gemini-3.1-pro",
-    "gemini-3-flash",
-    "gemini-3.5-flash",
-    "gemini-3.7-flash",
-    "grok-build-0.1",
-    "grok-4.5",
-    "grok-4.6",
-    "glm-5.2",
-    "muse-spark-1.1",
-    "kimi-k2-7-code",
-    "kimi-k3",
-    "deepseek-v4-flash-0731",
-    "qwen3.8-2.4t-a95b-deepinfra",
-    "qwen3.8-27b-deepinfra",
-    "minimax-m3"
-  ]
+  text: Object.keys(TEXT_MODELS)
   // vision: undefined — unconstrained
   // image_generation: undefined — unconstrained
 };
@@ -559,6 +581,11 @@ function getEffectiveModelSurfaces() {
 function resolveModel(surfaceId, models, fallback) {
   return models?.[surfaceId] ?? fallback ?? orgDefaultModels[surfaceId] ?? MODEL_SURFACES[surfaceId].default;
 }
+function resolveParentModel(models, fallback, buildModel) {
+  const override = buildModel ? filterModelPicks({ parent: buildModel }).parent : void 0;
+  const baseline = resolveModel("parent", models, fallback);
+  return { baseline, effective: override ?? baseline };
+}
 
 // src/orgContext.ts
 var log3 = createLogger("orgContext");
@@ -8616,6 +8643,12 @@ function classify(line, matches) {
     tailClean = SEPARATORS_ONLY.test(gap);
   }
 }
+function chipCase(label) {
+  if (label.startsWith("`")) {
+    return label;
+  }
+  return label.charAt(0).toUpperCase() + label.slice(1);
+}
 function parseSuggestions(raw) {
   if (!raw.includes(SUGGEST_MARKER)) {
     return { text: raw, suggestions: [] };
@@ -8645,7 +8678,7 @@ function parseSuggestions(raw) {
       const key = `${m.label}::${m.message}`;
       if (!seen.has(key)) {
         seen.add(key);
-        suggestions.push({ label: m.label, message: m.message });
+        suggestions.push({ label: chipCase(m.label), message: m.message });
       }
     }
     let rebuilt = "";
@@ -8745,10 +8778,12 @@ async function runTurn(params) {
     onBackgroundComplete
   } = params;
   const tools2 = getToolDefinitions();
-  const buildModelOverride = buildModel ? filterModelPicks({ parent: buildModel }).parent : void 0;
-  const baseline = resolveModel("parent", state.models, model);
-  const parentModel = buildModelOverride ?? baseline;
-  const modelOverride = buildModelOverride && buildModelOverride !== baseline ? { from: baseline } : void 0;
+  const { baseline, effective: parentModel } = resolveParentModel(
+    state.models,
+    model,
+    buildModel
+  );
+  const modelOverride = parentModel !== baseline ? { from: baseline } : void 0;
   const totalAttachments = entries.reduce(
     (n, e) => n + (e.attachments?.length ?? 0),
     0
@@ -8756,7 +8791,7 @@ async function runTurn(params) {
   log15.info("Turn started", {
     requestId,
     model,
-    buildModel: buildModelOverride,
+    buildModel: modelOverride ? parentModel : void 0,
     toolCount: tools2.length,
     ...entries.length > 1 && { entryCount: entries.length },
     ...totalAttachments > 0 && { attachmentCount: totalAttachments }
@@ -9591,12 +9626,13 @@ function loadPassiveResults() {
   }
   return [];
 }
-function writeStats(stats, queue, passiveResults) {
+function writeStats(stats, queue, passiveResults, suggestCompactAt) {
   try {
     writeFileAtomicSync(
       STATS_FILE,
       JSON.stringify({
         ...stats,
+        suggestCompactAt,
         queue,
         passiveResults
       })
@@ -9798,7 +9834,6 @@ var USER_FACING_TOOLS = /* @__PURE__ */ new Set([
   "confirmDestructiveAction",
   "presentPublishPlan"
 ]);
-var FORCED_COMPACTION_THRESHOLD_TOKENS = 85e4;
 var HeadlessSession = class {
   // Configuration
   opts;
@@ -10071,7 +10106,15 @@ var HeadlessSession = class {
   /** Persist sessionStats + queue snapshot + passive pen to .remy-stats.json. */
   persistStats() {
     this.sessionStats.updatedAt = Date.now();
-    writeStats(this.sessionStats, this.queue.snapshot(), this.passivePen);
+    const suggestCompactAt = getSuggestCompactAt(
+      resolveParentModel(this.state.models, this.opts.model).effective
+    );
+    writeStats(
+      this.sessionStats,
+      this.queue.snapshot(),
+      this.passivePen,
+      suggestCompactAt
+    );
   }
   //////////////////////////////////////////////////////////////////////////////
   // Background completions (tool-block mutation; message delivery via queue)
@@ -10101,8 +10144,9 @@ var HeadlessSession = class {
     saveSession(this.state);
   }
   /**
-   * Forced compaction gate. If lastContextSize exceeds the threshold, compact
-   * before letting the upcoming turn run. Coalesces with any in-flight
+   * Forced compaction gate. If lastContextSize exceeds the upcoming turn's
+   * model threshold (per-model — see ModelContextLimits in models/surfaces),
+   * compact before letting the turn run. Coalesces with any in-flight
    * compaction (e.g., one already started by /compact or a tool call). No
    * timeout — compaction takes as long as it takes.
    *
@@ -10113,13 +10157,15 @@ var HeadlessSession = class {
    * On compaction failure we don't bail — the turn proceeds and surfaces any
    * downstream overflow through the existing "prompt is too long" path.
    */
-  async runForcedCompactionIfNeeded(requestId) {
-    if (this.sessionStats.lastContextSize <= FORCED_COMPACTION_THRESHOLD_TOKENS) {
+  async runForcedCompactionIfNeeded(requestId, parentModel) {
+    const threshold = getContextLimits(parentModel).forceCompactAt;
+    if (this.sessionStats.lastContextSize <= threshold) {
       return;
     }
     log17.info("Forced compaction gate triggered", {
       contextSize: this.sessionStats.lastContextSize,
-      threshold: FORCED_COMPACTION_THRESHOLD_TOKENS,
+      threshold,
+      model: parentModel,
       requestId
     });
     try {
@@ -10601,7 +10647,12 @@ var HeadlessSession = class {
       }
       return steered;
     };
-    await this.runForcedCompactionIfNeeded(requestId);
+    const parentModel = resolveParentModel(
+      this.state.models,
+      this.opts.model,
+      params.buildModel
+    ).effective;
+    await this.runForcedCompactionIfNeeded(requestId, parentModel);
     try {
       await runTurn({
         state: this.state,

package/dist/index.js

@@ -2111,6 +2111,12 @@ var init_compaction = __esm({
 });
 
 // src/models/surfaces.ts
+function getContextLimits(modelId) {
+  return TEXT_MODELS[modelId] ?? DEFAULT_CONTEXT_LIMITS;
+}
+function getSuggestCompactAt(modelId) {
+  return getContextLimits(modelId).suggestCompactAt ?? DEFAULT_SUGGEST_COMPACT_AT;
+}
 function setOrgDefaultModels(models) {
   orgDefaultModels = models;
 }
@@ -2149,7 +2155,12 @@ function getEffectiveModelSurfaces() {
 function resolveModel(surfaceId, models, fallback) {
   return models?.[surfaceId] ?? fallback ?? orgDefaultModels[surfaceId] ?? MODEL_SURFACES[surfaceId].default;
 }
-var MODEL_SURFACES, ALLOWED_MODELS_BY_TYPE, orgDefaultModels;
+function resolveParentModel(models, fallback, buildModel) {
+  const override = buildModel ? filterModelPicks({ parent: buildModel }).parent : void 0;
+  const baseline = resolveModel("parent", models, fallback);
+  return { baseline, effective: override ?? baseline };
+}
+var MODEL_SURFACES, DEFAULT_SUGGEST_COMPACT_AT, TEXT_MODELS, DEFAULT_CONTEXT_LIMITS, ALLOWED_MODELS_BY_TYPE, orgDefaultModels;
 var init_surfaces = __esm({
   "src/models/surfaces.ts"() {
     "use strict";
@@ -2241,37 +2252,53 @@ var init_surfaces = __esm({
         userPickable: false
       }
     };
+    DEFAULT_SUGGEST_COMPACT_AT = 3e5;
+    TEXT_MODELS = {
+      // Anthropic 1M-context, flat-priced.
+      "claude-5-opus": { forceCompactAt: 85e4 },
+      "claude-4-8-opus": { forceCompactAt: 85e4 },
+      "claude-4-7-opus": { forceCompactAt: 85e4 },
+      // Anthropic 1M-context with 2x long-context pricing above 200K input.
+      "claude-4-6-opus": { forceCompactAt: 85e4, suggestCompactAt: 18e4 },
+      "claude-4-6-sonnet": { forceCompactAt: 85e4, suggestCompactAt: 18e4 },
+      "claude-fable-5": { forceCompactAt: 85e4 },
+      "claude-fable-5-1": { forceCompactAt: 85e4 },
+      "claude-5-sonnet": { forceCompactAt: 85e4 },
+      // OpenAI gpt-5.5/5.6: ~1M window, but the usable input ceiling under
+      // `truncation: 'auto'` is ~794K (output + reasoning reserve), and all
+      // rates double above 272K input.
+      "gpt-5.5": { forceCompactAt: 6e5, suggestCompactAt: 25e4 },
+      "gpt-5.6-sol": { forceCompactAt: 6e5, suggestCompactAt: 25e4 },
+      "gpt-5.6-terra": { forceCompactAt: 6e5, suggestCompactAt: 25e4 },
+      "gpt-5.6-luna": { forceCompactAt: 6e5, suggestCompactAt: 25e4 },
+      // Google ~1M-context; only 3.1-pro is tiered (higher rates above 200K).
+      "gemini-3-pro": { forceCompactAt: 85e4 },
+      "gemini-3.1-pro": { forceCompactAt: 85e4, suggestCompactAt: 18e4 },
+      "gemini-3-flash": { forceCompactAt: 85e4 },
+      "gemini-3.5-flash": { forceCompactAt: 85e4 },
+      "gemini-3.7-flash": { forceCompactAt: 85e4 },
+      // 256K window; its 200K pricing tier sits above the gate, so no nudge.
+      "grok-build-0.1": { forceCompactAt: 18e4 },
+      "grok-4.5": { forceCompactAt: 4e5 },
+      // 500K window
+      "grok-4.6": { forceCompactAt: 4e5 },
+      // 500K window
+      "glm-5.2": { forceCompactAt: 85e4 },
+      "muse-spark-1.1": { forceCompactAt: 85e4 },
+      "kimi-k2-7-code": { forceCompactAt: 2e5 },
+      // 262K window
+      "kimi-k3": { forceCompactAt: 85e4 },
+      "deepseek-v4-flash-0731": { forceCompactAt: 85e4 },
+      "qwen3.8-2.4t-a95b-deepinfra": { forceCompactAt: 2e5 },
+      // 262K window
+      "qwen3.8-27b-deepinfra": { forceCompactAt: 2e5 },
+      // 262K window
+      "minimax-m3": { forceCompactAt: 42e4 }
+      // 524K window
+    };
+    DEFAULT_CONTEXT_LIMITS = { forceCompactAt: 85e4 };
     ALLOWED_MODELS_BY_TYPE = {
-      text: [
-        "claude-5-opus",
-        "claude-4-8-opus",
-        "claude-4-7-opus",
-        "claude-4-6-opus",
-        "claude-4-6-sonnet",
-        "claude-fable-5",
-        "claude-fable-5-1",
-        "claude-5-sonnet",
-        "gpt-5.5",
-        "gpt-5.6-sol",
-        "gpt-5.6-terra",
-        "gpt-5.6-luna",
-        "gemini-3-pro",
-        "gemini-3.1-pro",
-        "gemini-3-flash",
-        "gemini-3.5-flash",
-        "gemini-3.7-flash",
-        "grok-build-0.1",
-        "grok-4.5",
-        "grok-4.6",
-        "glm-5.2",
-        "muse-spark-1.1",
-        "kimi-k2-7-code",
-        "kimi-k3",
-        "deepseek-v4-flash-0731",
-        "qwen3.8-2.4t-a95b-deepinfra",
-        "qwen3.8-27b-deepinfra",
-        "minimax-m3"
-      ]
+      text: Object.keys(TEXT_MODELS)
       // vision: undefined — unconstrained
       // image_generation: undefined — unconstrained
     };
@@ -8921,6 +8948,12 @@ function classify(line, matches) {
     tailClean = SEPARATORS_ONLY.test(gap);
   }
 }
+function chipCase(label) {
+  if (label.startsWith("`")) {
+    return label;
+  }
+  return label.charAt(0).toUpperCase() + label.slice(1);
+}
 function parseSuggestions(raw) {
   if (!raw.includes(SUGGEST_MARKER)) {
     return { text: raw, suggestions: [] };
@@ -8950,7 +8983,7 @@ function parseSuggestions(raw) {
       const key = `${m.label}::${m.message}`;
       if (!seen.has(key)) {
         seen.add(key);
-        suggestions.push({ label: m.label, message: m.message });
+        suggestions.push({ label: chipCase(m.label), message: m.message });
       }
     }
     let rebuilt = "";
@@ -9401,10 +9434,12 @@ async function runTurn(params) {
     onBackgroundComplete
   } = params;
   const tools2 = getToolDefinitions();
-  const buildModelOverride = buildModel ? filterModelPicks({ parent: buildModel }).parent : void 0;
-  const baseline = resolveModel("parent", state.models, model);
-  const parentModel = buildModelOverride ?? baseline;
-  const modelOverride = buildModelOverride && buildModelOverride !== baseline ? { from: baseline } : void 0;
+  const { baseline, effective: parentModel } = resolveParentModel(
+    state.models,
+    model,
+    buildModel
+  );
+  const modelOverride = parentModel !== baseline ? { from: baseline } : void 0;
   const totalAttachments = entries.reduce(
     (n, e) => n + (e.attachments?.length ?? 0),
     0
@@ -9412,7 +9447,7 @@ async function runTurn(params) {
   log14.info("Turn started", {
     requestId,
     model,
-    buildModel: buildModelOverride,
+    buildModel: modelOverride ? parentModel : void 0,
     toolCount: tools2.length,
     ...entries.length > 1 && { entryCount: entries.length },
     ...totalAttachments > 0 && { attachmentCount: totalAttachments }
@@ -10551,12 +10586,13 @@ function loadPassiveResults() {
   }
   return [];
 }
-function writeStats(stats, queue, passiveResults) {
+function writeStats(stats, queue, passiveResults, suggestCompactAt) {
   try {
     writeFileAtomicSync(
       STATS_FILE,
       JSON.stringify({
         ...stats,
+        suggestCompactAt,
         queue,
         passiveResults
       })
@@ -10767,7 +10803,7 @@ var headless_exports = {};
 __export(headless_exports, {
   HeadlessSession: () => HeadlessSession
 });
-var log17, EXTERNAL_TOOL_TIMEOUT_MS, LONG_RUNNING_TOOLS, LONG_RUNNING_TOOL_TIMEOUT_MS, USER_FACING_TOOLS, FORCED_COMPACTION_THRESHOLD_TOKENS, HeadlessSession;
+var log17, EXTERNAL_TOOL_TIMEOUT_MS, LONG_RUNNING_TOOLS, LONG_RUNNING_TOOL_TIMEOUT_MS, USER_FACING_TOOLS, HeadlessSession;
 var init_headless = __esm({
   "src/headless/index.ts"() {
     "use strict";
@@ -10798,7 +10834,6 @@ var init_headless = __esm({
       "confirmDestructiveAction",
       "presentPublishPlan"
     ]);
-    FORCED_COMPACTION_THRESHOLD_TOKENS = 85e4;
     HeadlessSession = class {
       // Configuration
       opts;
@@ -11071,7 +11106,15 @@ var init_headless = __esm({
       /** Persist sessionStats + queue snapshot + passive pen to .remy-stats.json. */
       persistStats() {
         this.sessionStats.updatedAt = Date.now();
-        writeStats(this.sessionStats, this.queue.snapshot(), this.passivePen);
+        const suggestCompactAt = getSuggestCompactAt(
+          resolveParentModel(this.state.models, this.opts.model).effective
+        );
+        writeStats(
+          this.sessionStats,
+          this.queue.snapshot(),
+          this.passivePen,
+          suggestCompactAt
+        );
       }
       //////////////////////////////////////////////////////////////////////////////
       // Background completions (tool-block mutation; message delivery via queue)
@@ -11101,8 +11144,9 @@ var init_headless = __esm({
         saveSession(this.state);
       }
       /**
-       * Forced compaction gate. If lastContextSize exceeds the threshold, compact
-       * before letting the upcoming turn run. Coalesces with any in-flight
+       * Forced compaction gate. If lastContextSize exceeds the upcoming turn's
+       * model threshold (per-model — see ModelContextLimits in models/surfaces),
+       * compact before letting the turn run. Coalesces with any in-flight
        * compaction (e.g., one already started by /compact or a tool call). No
        * timeout — compaction takes as long as it takes.
        *
@@ -11113,13 +11157,15 @@ var init_headless = __esm({
        * On compaction failure we don't bail — the turn proceeds and surfaces any
        * downstream overflow through the existing "prompt is too long" path.
        */
-      async runForcedCompactionIfNeeded(requestId) {
-        if (this.sessionStats.lastContextSize <= FORCED_COMPACTION_THRESHOLD_TOKENS) {
+      async runForcedCompactionIfNeeded(requestId, parentModel) {
+        const threshold = getContextLimits(parentModel).forceCompactAt;
+        if (this.sessionStats.lastContextSize <= threshold) {
           return;
         }
         log17.info("Forced compaction gate triggered", {
           contextSize: this.sessionStats.lastContextSize,
-          threshold: FORCED_COMPACTION_THRESHOLD_TOKENS,
+          threshold,
+          model: parentModel,
           requestId
         });
         try {
@@ -11601,7 +11647,12 @@ var init_headless = __esm({
           }
           return steered;
         };
-        await this.runForcedCompactionIfNeeded(requestId);
+        const parentModel = resolveParentModel(
+          this.state.models,
+          this.opts.model,
+          params.buildModel
+        ).effective;
+        await this.runForcedCompactionIfNeeded(requestId, parentModel);
         try {
           await runTurn({
             state: this.state,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mindstudio-ai/remy",
-  "version": "0.1.312",
+  "version": "0.1.313",
   "description": "Remy coding agent",
   "repository": {
     "type": "git",