open-agents-ai 0.39.1 → 0.40.0

package/README.md

@@ -109,7 +109,7 @@ Long conversations consume context window tokens. Open Agents uses progressive c
 
 ### How It Works
 
-Compaction triggers automatically when estimated token usage reaches 75% of the model's context window. The system:
+Compaction triggers automatically when estimated token usage reaches a tier-proportional threshold of the model's context window. The system:
 
 1. **Preserves** the system prompt and initial user task (head messages)
 2. **Summarizes** middle messages (tool calls, results, exploration) into a structured digest
@@ -131,13 +131,36 @@ Six strategies are available via `/compact <strategy>`:
 
 ### Automatic Compaction
 
-Compaction thresholds scale dynamically with model size:
+Compaction thresholds scale **proportionally** with the model's actual context window size:
 
-| Model Tier | Threshold | Recent Messages Kept |
-|------------|-----------|---------------------|
-| Large (30B+) | 40,000 tokens (or 75% of context) | 12 messages |
-| Medium (8-29B) | 24,000 tokens (or 75% of context) | 8 messages |
-| Small (≤7B) | 12,000 tokens (or 75% of context) | 4-6 messages |
+| Model Tier | Normal Mode | Deep Context Mode | Recent Messages Kept |
+|------------|-------------|-------------------|---------------------|
+| Large (30B+) | 75% of context window | 85% of context window | 4-12 (normal) / 4-24 (deep) |
+| Medium (8-29B) | 70% of context window | 85% of context window | 4-12 (normal) / 4-24 (deep) |
+| Small (≤7B) | 65% of context window | 85% of context window | 4-12 (normal) / 4-24 (deep) |
+
+For example, a 128K-context large model compacts at ~96K tokens in normal mode (75%) or ~109K tokens in deep mode (85%) — instead of the previous fixed 40K threshold that wasted 69% of available context.
+
+### Deep Context Mode (`/deep`)
+
+Toggle with `/deep` — relaxes compaction so large models leverage more of their context window for complex multi-step reasoning.
+
+When deep context is active:
+- **Compaction fires at 85%** of context instead of 65-75% — the model retains much more working memory
+- **Double the recent messages** (up to 24 instead of 12) preserved after compaction
+- **Richer summaries** — compression budget increased from 20% to 30% of context
+- **Larger tool outputs** — cap raised from 8K to 16K chars per tool result
+- **Relaxed output folding** — more head/tail lines preserved (50/25 instead of 20/10 for large models)
+
+This mirrors how human cognition works during deep problem-solving: situationally-relevant memories are transiently activated to occupy a larger portion of working memory, with the most relevant details in high-attention positions while supporting context backs them up. LLM attention mechanisms work similarly — earlier relevant context still influences generation even at lower positional weight.
+
+Use deep context for:
+- Complex multi-file refactoring or debugging
+- Architecture analysis across many files
+- Long debugging sessions where error context from earlier is critical
+- Tasks where the agent needs to reason about patterns across many files
+
+The setting persists to `.oa/settings.json`. Deep context is particularly valuable for models with 64K+ context windows (Qwen3.5-122B, Llama 3.1 70B, etc.) where the default thresholds were leaving significant capacity unused.
 
 ### Status Bar Context Tracking (`Ctx:`)
 

package/dist/index.js

@@ -13180,6 +13180,7 @@ Rules:
           taskTimeoutMs: options?.taskTimeoutMs ?? 36e5,
           // 60 min
           compactionThreshold: options?.compactionThreshold ?? 4e4,
+          deepContext: options?.deepContext ?? false,
           dynamicContext: options?.dynamicContext ?? "",
           streamEnabled: options?.streamEnabled ?? false,
           bruteForce: options?.bruteForce ?? true,
@@ -13225,19 +13226,47 @@ Rules:
       /**
        * Compute all context-dependent limits from the current contextWindowSize
        * and modelTier. Returns sensible defaults when contextWindowSize is 0.
+       *
+       * Key insight: LLM attention is most effective when the most relevant content
+       * sits in high-attention positions (start and end of context). Rather than
+       * compacting aggressively and losing detail, we scale thresholds proportionally
+       * to the model's actual context window so larger models can leverage more of
+       * their capacity. The deepContext flag further relaxes limits for complex
+       * multi-step reasoning where earlier context carries critical signal.
+       *
+       * Context utilization targets:
+       * - Small models:  compact at ~65% (limited attention span, aggressive helps)
+       * - Medium models: compact at ~70% (balanced)
+       * - Large models:  compact at ~75% (plenty of room, let attention work)
+       * - Deep context:  compact at ~85% (trust the model's full attention window)
        */
       contextLimits() {
         const ctx = this.options.contextWindowSize;
         const tier = this.options.modelTier ?? "large";
-        const compactionThreshold = ctx > 0 ? Math.min(this.options.compactionThreshold, Math.floor(ctx * 0.75)) : this.options.compactionThreshold;
-        const keepRecentDivisor = tier === "small" ? 2e3 : tier === "medium" ? 3e3 : 4e3;
-        const keepRecent = ctx > 0 ? Math.max(4, Math.min(12, Math.floor(ctx / keepRecentDivisor))) : 12;
+        const deep = this.options.deepContext ?? false;
+        let compactionThreshold;
+        if (ctx > 0) {
+          const factor = deep ? 0.85 : tier === "small" ? 0.65 : tier === "medium" ? 0.7 : 0.75;
+          compactionThreshold = Math.floor(ctx * factor);
+        } else {
+          compactionThreshold = deep ? 8e4 : this.options.compactionThreshold;
+        }
+        const keepRecentMax = deep ? 24 : 12;
+        let keepRecent;
+        if (ctx > 0) {
+          const keepRecentDivisor = tier === "small" ? 2e3 : tier === "medium" ? 3e3 : 4e3;
+          keepRecent = Math.max(4, Math.min(keepRecentMax, Math.floor(ctx / keepRecentDivisor)));
+        } else {
+          keepRecent = deep ? 20 : 12;
+        }
         const maxOutputTokens = ctx > 0 ? Math.min(this.options.maxTokens, Math.max(2048, Math.floor(ctx * 0.25))) : this.options.maxTokens;
-        const toolOutputMaxChars = ctx > 0 ? Math.max(2e3, Math.min(8e3, Math.floor(ctx * 0.5))) : 8e3;
-        const foldLineThreshold = tier === "small" ? 30 : tier === "medium" ? 35 : 40;
-        const foldHeadLines = tier === "small" ? 12 : tier === "medium" ? 16 : 20;
-        const foldTailLines = tier === "small" ? 5 : tier === "medium" ? 8 : 10;
-        const maxSummaryChars = ctx > 0 ? Math.max(2e3, Math.min(8e3, Math.floor(ctx * 0.2))) : 4e3;
+        const toolOutputCeiling = deep ? 16e3 : 8e3;
+        const toolOutputMaxChars = ctx > 0 ? Math.max(2e3, Math.min(toolOutputCeiling, Math.floor(ctx * (deep ? 0.8 : 0.5)))) : toolOutputCeiling;
+        const foldLineThreshold = deep ? tier === "small" ? 50 : tier === "medium" ? 80 : 120 : tier === "small" ? 30 : tier === "medium" ? 35 : 40;
+        const foldHeadLines = deep ? tier === "small" ? 20 : tier === "medium" ? 30 : 50 : tier === "small" ? 12 : tier === "medium" ? 16 : 20;
+        const foldTailLines = deep ? tier === "small" ? 10 : tier === "medium" ? 15 : 25 : tier === "small" ? 5 : tier === "medium" ? 8 : 10;
+        const maxSummaryCeiling = deep ? 16e3 : 8e3;
+        const maxSummaryChars = ctx > 0 ? Math.max(2e3, Math.min(maxSummaryCeiling, Math.floor(ctx * (deep ? 0.3 : 0.2)))) : deep ? 8e3 : 4e3;
         const repetitionWindow = tier === "small" ? 6 : tier === "medium" ? 8 : 10;
         return {
           compactionThreshold,
@@ -16914,6 +16943,7 @@ function renderSlashHelp() {
     ["/compact", "Force context compaction now (default strategy)"],
     ["/compact <strategy>", "Compact with strategy: aggressive, decisions, errors, summary, structured"],
     ["/bruteforce", "Toggle brute-force mode (auto re-engage on turn limit)"],
+    ["/deep", "Toggle deep context \u2014 relaxes compaction so large models use 85% of context"],
     ["/tools", "List agent-created custom tools"],
     ["/skills", "List available AIWG skills"],
     ["/skills <keyword>", "Filter skills by name or trigger"],
@@ -19451,6 +19481,18 @@ async function handleSlashCommand(input, ctx) {
       renderInfo(`Brute-force mode: ${isOn ? "on" : "off"}${hasLocal ? " (project-local)" : ""}` + (isOn ? " \u2014 agent will auto re-engage when turn limit is hit, reassess and try creative strategies" : ""));
       return "handled";
     }
+    case "deep":
+    case "deep-context": {
+      if (!ctx.deepContextToggle) {
+        renderWarning("Deep context not available");
+        return "handled";
+      }
+      const isOn = ctx.deepContextToggle();
+      const save = hasLocal ? ctx.saveLocalSettings.bind(ctx) : ctx.saveSettings.bind(ctx);
+      save({ deepContext: isOn });
+      renderInfo(`Deep context mode: ${isOn ? "ON" : "OFF"}${hasLocal ? " (project-local)" : ""}` + (isOn ? " \u2014 compaction relaxed to 85% of context window. Large models will retain more working memory for complex reasoning." : " \u2014 compaction restored to default thresholds."));
+      return "handled";
+    }
     case "emojis":
     case "emoji": {
       const current = ctx.getEmojis?.() ?? true;
@@ -25552,7 +25594,7 @@ Use task_status("${taskId}") or task_output("${taskId}") to check progress.`
     }
   };
 }
-function startTask(task, config, repoRoot, voice, stream, taskStores, bruteForce, statusBar, sudoCallback, costTracker, onComplete, taskType, contextWindowSize, modelCaps, personality) {
+function startTask(task, config, repoRoot, voice, stream, taskStores, bruteForce, statusBar, sudoCallback, costTracker, onComplete, taskType, contextWindowSize, modelCaps, personality, deepContext) {
   const voiceStyleMap = {
     concise: 1,
     balanced: 3,
@@ -25576,6 +25618,7 @@ function startTask(task, config, repoRoot, voice, stream, taskStores, bruteForce
     taskTimeoutMs: 36e5,
     // 60 minutes — never give up prematurely
     compactionThreshold,
+    deepContext: deepContext ?? false,
     dynamicContext,
     modelTier,
     streamEnabled: stream?.enabled ?? false,
@@ -25924,6 +25967,7 @@ async function startInteractive(config, repoPath) {
   let streamEnabled = savedSettings.stream ?? false;
   let bruteForceEnabled = savedSettings.bruteforce ?? true;
   let currentStyle = PRESET_NAMES.includes(savedSettings.style) ? savedSettings.style : "balanced";
+  let deepContextEnabled = savedSettings.deepContext ?? false;
   if (savedSettings.emojis !== void 0)
     setEmojisEnabled(savedSettings.emojis);
   if (savedSettings.colors !== void 0)
@@ -26175,6 +26219,7 @@ Rationale: ${proposal.rationale}${provenanceNote}`;
     "/telegram",
     "/compact",
     "/gc",
+    "/deep",
     "/style",
     "/personality"
   ];
@@ -26336,6 +26381,10 @@ Rationale: ${proposal.rationale}${provenanceNote}`;
       bruteForceEnabled = !bruteForceEnabled;
       return bruteForceEnabled;
     },
+    deepContextToggle() {
+      deepContextEnabled = !deepContextEnabled;
+      return deepContextEnabled;
+    },
     setStyle(preset) {
       currentStyle = preset;
     },
@@ -26875,7 +26924,7 @@ Execute this skill now. Follow the behavioral guidance above.`;
             toolPatternStore: toolPatternStore ?? void 0
           }, bruteForceEnabled, statusBar, handleSudoRequest, costTracker, (summary) => {
             lastCompletedSummary = summary;
-          }, currentTaskType, resolvedContextWindowSize, resolvedCaps, currentStyle);
+          }, currentTaskType, resolvedContextWindowSize, resolvedCaps, currentStyle, deepContextEnabled);
           activeTask = task;
           showPrompt();
           await task.promise;
@@ -26987,7 +27036,7 @@ NEW TASK: ${fullInput}`;
         toolPatternStore: toolPatternStore ?? void 0
       }, bruteForceEnabled, statusBar, handleSudoRequest, costTracker, (summary) => {
         lastCompletedSummary = summary;
-      }, currentTaskType, resolvedContextWindowSize, resolvedCaps, currentStyle);
+      }, currentTaskType, resolvedContextWindowSize, resolvedCaps, currentStyle, deepContextEnabled);
       activeTask = task;
       showPrompt();
       await task.promise;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "open-agents-ai",
-  "version": "0.39.1",
+  "version": "0.40.0",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",