@visorcraft/idlehands 1.2.7 → 1.3.1

package/dist/agent.js

@@ -157,41 +157,41 @@ class AgentLoopBreak extends Error {
         this.name = 'AgentLoopBreak';
     }
 }
-const SYSTEM_PROMPT = `You are a coding agent with filesystem and shell access. Execute the user's request using the provided tools.
-
-Rules:
-- Work in the current directory. Use relative paths for all file operations.
-- Do the work directly. Do NOT use spawn_task to delegate the user's primary request — only use it for genuinely independent subtasks that benefit from parallel execution.
-- Never use spawn_task to bypass confirmation/safety restrictions (for example blocked package installs). If a command is blocked, adapt the plan or ask the user for approval mode changes.
-- Read the target file before editing. You need the exact text for search/replace.
-- Use read_file with search=... to jump to relevant code; avoid reading whole files.
-- Never call read_file/read_files/list_dir twice in a row with identical arguments (same path/options). Reuse the previous result instead.
-- Prefer apply_patch or edit_range for code edits (token-efficient). Use edit_file only when exact old_text replacement is necessary.
-- Tool-call arguments MUST be strict JSON (double-quoted keys/strings, no comments, no trailing commas).
-- edit_range example: {"path":"src/foo.ts","start_line":10,"end_line":14,"replacement":"line A\nline B"}
-- apply_patch example: {"patch":"--- a/src/foo.ts\n+++ b/src/foo.ts\n@@ -10,2 +10,2 @@\n-old\n+new","files":["src/foo.ts"]}
-- write_file is for new files or explicit full rewrites only. Existing non-empty files require overwrite=true/force=true.
-- Use insert_file for insertions (prepend/append/line).
-- Use exec to run commands, tests, builds; check results before reporting success.
-- When running commands in a subdirectory, use exec's cwd parameter — NOT "cd /path && cmd". Each exec call is a fresh shell; cd does not persist.
-- Batch work: read all files you need, then apply all edits, then verify.
-- Be concise. Report what you changed and why.
-- Do NOT read every file in a directory. Use search_files or exec with grep to locate relevant code first, then read only the files that match.
-- If search_files returns 0 matches, try a broader pattern or use: exec grep -rn "keyword" path/
-- Anton (the autonomous task runner) is ONLY activated when the user explicitly invokes /anton. Never self-activate as Anton or start processing task files on your own.
-
-Tool call format:
-- Use tool_calls. Do not write JSON tool invocations in your message text.
+const SYSTEM_PROMPT = `You are a coding agent with filesystem and shell access. Execute the user's request using the provided tools.
+
+Rules:
+- Work in the current directory. Use relative paths for all file operations.
+- Do the work directly. Do NOT use spawn_task to delegate the user's primary request — only use it for genuinely independent subtasks that benefit from parallel execution.
+- Never use spawn_task to bypass confirmation/safety restrictions (for example blocked package installs). If a command is blocked, adapt the plan or ask the user for approval mode changes.
+- Read the target file before editing. You need the exact text for search/replace.
+- Use read_file with search=... to jump to relevant code; avoid reading whole files.
+- Never call read_file/read_files/list_dir twice in a row with identical arguments (same path/options). Reuse the previous result instead.
+- Prefer apply_patch or edit_range for code edits (token-efficient). Use edit_file only when exact old_text replacement is necessary.
+- Tool-call arguments MUST be strict JSON (double-quoted keys/strings, no comments, no trailing commas).
+- edit_range example: {"path":"src/foo.ts","start_line":10,"end_line":14,"replacement":"line A\nline B"}
+- apply_patch example: {"patch":"--- a/src/foo.ts\n+++ b/src/foo.ts\n@@ -10,2 +10,2 @@\n-old\n+new","files":["src/foo.ts"]}
+- write_file is for new files or explicit full rewrites only. Existing non-empty files require overwrite=true/force=true.
+- Use insert_file for insertions (prepend/append/line).
+- Use exec to run commands, tests, builds; check results before reporting success.
+- When running commands in a subdirectory, use exec's cwd parameter — NOT "cd /path && cmd". Each exec call is a fresh shell; cd does not persist.
+- Batch work: read all files you need, then apply all edits, then verify.
+- Be concise. Report what you changed and why.
+- Do NOT read every file in a directory. Use search_files or exec with grep to locate relevant code first, then read only the files that match.
+- If search_files returns 0 matches, try a broader pattern or use: exec grep -rn "keyword" path/
+- Anton (the autonomous task runner) is ONLY activated when the user explicitly invokes /anton. Never self-activate as Anton or start processing task files on your own.
+
+Tool call format:
+- Use tool_calls. Do not write JSON tool invocations in your message text.
 `;
 const MCP_TOOLS_REQUEST_TOKEN = '[[MCP_TOOLS_REQUEST]]';
-const DEFAULT_SUB_AGENT_SYSTEM_PROMPT = `You are a focused coding sub-agent. Execute only the delegated task.
-- Work in the current directory. Use relative paths for all file operations.
-- Read the target file before editing. You need the exact text for search/replace.
-- Keep tool usage tight and efficient.
-- Prefer surgical edits over rewrites.
-- Do NOT create files outside the working directory unless explicitly requested.
-- When running commands in a subdirectory, use exec's cwd parameter — NOT "cd /path && cmd".
-- Run verification commands when relevant.
+const DEFAULT_SUB_AGENT_SYSTEM_PROMPT = `You are a focused coding sub-agent. Execute only the delegated task.
+- Work in the current directory. Use relative paths for all file operations.
+- Read the target file before editing. You need the exact text for search/replace.
+- Keep tool usage tight and efficient.
+- Prefer surgical edits over rewrites.
+- Do NOT create files outside the working directory unless explicitly requested.
+- When running commands in a subdirectory, use exec's cwd parameter — NOT "cd /path && cmd".
+- Run verification commands when relevant.
 - Return a concise outcome summary.`;
 const DEFAULT_SUB_AGENT_RESULT_TOKEN_CAP = 4000;
 const APPROVAL_MODE_SET = new Set(['plan', 'reject', 'default', 'auto-edit', 'yolo']);
@@ -844,9 +844,6 @@ export async function createSession(opts) {
             catch { }
         }
     }
-    if (harness.systemPromptSuffix) {
-        sessionMeta += '\n\n' + harness.systemPromptSuffix;
-    }
     // Phase 9: sys-eager — inject full system snapshot into first message
     if (cfg.sys_eager && cfg.mode === 'sys') {
         try {
@@ -857,30 +854,42 @@ export async function createSession(opts) {
             console.warn(`[warn] sys-eager snapshot failed: ${e?.message ?? e}`);
         }
     }
-    const defaultSystemPrompt = SYSTEM_PROMPT;
-    let activeSystemPrompt = (cfg.system_prompt_override ?? '').trim() || defaultSystemPrompt;
+    const defaultSystemPromptBase = SYSTEM_PROMPT;
+    let activeSystemPromptBase = (cfg.system_prompt_override ?? '').trim() || defaultSystemPromptBase;
+    let systemPromptOverridden = (cfg.system_prompt_override ?? '').trim().length > 0;
+    const buildEffectiveSystemPrompt = () => {
+        let p = activeSystemPromptBase;
+        if (harness.systemPromptSuffix && !systemPromptOverridden) {
+            p += '\n\n' + harness.systemPromptSuffix;
+        }
+        return p;
+    };
     let messages = [
-        { role: 'system', content: activeSystemPrompt }
+        { role: 'system', content: buildEffectiveSystemPrompt() }
     ];
     let sessionMetaPending = sessionMeta;
     const setSystemPrompt = (prompt) => {
         const next = String(prompt ?? '').trim();
         if (!next)
             throw new Error('system prompt cannot be empty');
-        activeSystemPrompt = next;
+        activeSystemPromptBase = next;
+        systemPromptOverridden = true;
+        const effective = buildEffectiveSystemPrompt();
         if (messages.length > 0 && messages[0].role === 'system') {
-            messages[0] = { role: 'system', content: activeSystemPrompt };
+            messages[0] = { role: 'system', content: effective };
         }
         else {
-            messages.unshift({ role: 'system', content: activeSystemPrompt });
+            messages.unshift({ role: 'system', content: effective });
         }
     };
     const resetSystemPrompt = () => {
-        setSystemPrompt(defaultSystemPrompt);
+        systemPromptOverridden = false;
+        setSystemPrompt(defaultSystemPromptBase);
     };
     const reset = () => {
+        const effective = buildEffectiveSystemPrompt();
         messages = [
-            { role: 'system', content: activeSystemPrompt }
+            { role: 'system', content: effective }
         ];
         sessionMetaPending = sessionMeta;
         lastEditedPath = undefined;
@@ -895,7 +904,9 @@ export async function createSession(opts) {
             throw new Error('restore: first message must be system');
         }
         messages = next;
-        activeSystemPrompt = String(next[0].content ?? defaultSystemPrompt);
+        activeSystemPromptBase = String(next[0].content ?? defaultSystemPromptBase);
+        // Note: we don't force buildEffectiveSystemPrompt() here because the restore
+        // data might already have a customized system prompt we want to respect.
         if (mcpManager) {
             const usedMcpTool = next.some((msg) => {
                 if (msg?.role !== 'assistant' || !Array.isArray(msg.tool_calls))
@@ -1125,6 +1136,7 @@ export async function createSession(opts) {
             dirPinned: cfg.dir_pinned,
             repoCandidates: cfg.repo_candidates,
             confirm: overrides?.confirmBridge ?? defaultConfirmBridge,
+            maxReadLines: cfg.max_read_lines,
             replay,
             vault,
             lens,
@@ -1132,6 +1144,41 @@ export async function createSession(opts) {
             onMutation: overrides?.onMutation ?? ((absPath) => { lastEditedPath = absPath; }),
         };
     };
+    const buildLspLensSymbolOutput = async (filePathRaw) => {
+        if (!lspManager)
+            return '[lsp] unavailable';
+        const semantic = await lspManager.getSymbols(filePathRaw);
+        if (!lens)
+            return semantic;
+        const cwd = cfg.dir ?? process.cwd();
+        const absPath = filePathRaw.startsWith('/') ? filePathRaw : path.resolve(cwd, filePathRaw);
+        const body = await fs.readFile(absPath, 'utf8').catch(() => '');
+        if (!body)
+            return semantic;
+        const projection = await lens.projectFile(absPath, body).catch(() => '');
+        const structural = extractLensBody(projection);
+        if (!structural)
+            return semantic;
+        return `${semantic}\n\n[lens] Structural skeleton:\n${structural}`;
+    };
+    const dispatchLspTool = async (name, args) => {
+        if (!lspManager)
+            return '[lsp] unavailable';
+        switch (name) {
+            case 'lsp_diagnostics':
+                return lspManager.getDiagnostics(typeof args?.path === 'string' ? args.path : undefined, typeof args?.severity === 'number' ? args.severity : undefined);
+            case 'lsp_symbols':
+                return buildLspLensSymbolOutput(String(args?.path ?? ''));
+            case 'lsp_hover':
+                return lspManager.getHover(String(args?.path ?? ''), Number(args?.line ?? 0), Number(args?.character ?? 0));
+            case 'lsp_definition':
+                return lspManager.getDefinition(String(args?.path ?? ''), Number(args?.line ?? 0), Number(args?.character ?? 0));
+            case 'lsp_references':
+                return lspManager.getReferences(String(args?.path ?? ''), Number(args?.line ?? 0), Number(args?.character ?? 0), typeof args?.max_results === 'number' ? args.max_results : 50);
+            default:
+                throw new Error(`unknown LSP tool: ${name}`);
+        }
+    };
     const executePlanStep = async (index) => {
         if (!planSteps.length)
             return ['No plan steps to execute.'];
@@ -1154,21 +1201,7 @@ export async function createSession(opts) {
                     content = await runSpawnTaskCore(step.args, { signal: inFlight?.signal });
                 }
                 else if (LSP_TOOL_NAME_SET.has(step.tool) && lspManager) {
-                    if (step.tool === 'lsp_diagnostics') {
-                        content = await lspManager.getDiagnostics(typeof step.args?.path === 'string' ? step.args.path : undefined, typeof step.args?.severity === 'number' ? step.args.severity : undefined);
-                    }
-                    else if (step.tool === 'lsp_symbols') {
-                        content = await lspManager.getSymbols(String(step.args?.path ?? ''));
-                    }
-                    else if (step.tool === 'lsp_hover') {
-                        content = await lspManager.getHover(String(step.args?.path ?? ''), Number(step.args?.line ?? 0), Number(step.args?.character ?? 0));
-                    }
-                    else if (step.tool === 'lsp_definition') {
-                        content = await lspManager.getDefinition(String(step.args?.path ?? ''), Number(step.args?.line ?? 0), Number(step.args?.character ?? 0));
-                    }
-                    else if (step.tool === 'lsp_references') {
-                        content = await lspManager.getReferences(String(step.args?.path ?? ''), Number(step.args?.line ?? 0), Number(step.args?.character ?? 0), typeof step.args?.max_results === 'number' ? step.args.max_results : 50);
-                    }
+                    content = await dispatchLspTool(step.tool, step.args);
                 }
                 else if (mcpManager?.hasTool(step.tool)) {
                     const callArgs = step.args && typeof step.args === 'object' && !Array.isArray(step.args)
@@ -1230,11 +1263,39 @@ export async function createSession(opts) {
     };
     const buildCompactionSystemNote = (kind, dropped) => {
         const prefix = kind === 'auto'
-            ? `[auto-compacted: ${dropped} old messages dropped to stay within context budget.]`
+            ? `[auto-compacted: ${dropped} old messages dropped to stay within context budget. Continue current task.]`
             : `[compacted: ${dropped} messages dropped.]`;
         const guidance = compactionVaultGuidance();
         return guidance ? `${prefix} ${guidance}` : prefix;
     };
+    const buildCompactionSummaryPrompt = (dropped) => {
+        const parts = [];
+        for (const m of dropped) {
+            if (m.role === 'assistant') {
+                const text = typeof m.content === 'string' ? m.content : '';
+                const toolCalls = m.tool_calls;
+                if (toolCalls?.length) {
+                    for (const tc of toolCalls) {
+                        const args = typeof tc.function?.arguments === 'string'
+                            ? tc.function.arguments.slice(0, 200)
+                            : '';
+                        parts.push(`[tool_call: ${tc.function?.name}(${args})]`);
+                    }
+                }
+                if (text.trim())
+                    parts.push(`[assistant]: ${text.slice(0, 500)}`);
+            }
+            else if (m.role === 'tool') {
+                const content = typeof m.content === 'string' ? m.content : '';
+                parts.push(`[tool_result]: ${content.slice(0, 300)}`);
+            }
+        }
+        let combined = parts.join('\n');
+        if (combined.length > 4000) {
+            combined = combined.slice(0, 4000) + '\n[...truncated]';
+        }
+        return combined;
+    };
     let lastAskInstructionText = '';
     let lastCompactionReminderObjective = '';
     const injectCompactionReminder = (reason) => {
@@ -1584,6 +1645,10 @@ export async function createSession(opts) {
             configuredTopP: cfg.top_p,
             baseMaxTokens: BASE_MAX_TOKENS,
         }));
+        // Update system prompt for the new model/harness
+        if (messages.length > 0 && messages[0].role === 'system') {
+            messages[0].content = buildEffectiveSystemPrompt();
+        }
         emitDetached(hookManager.emit('model_changed', {
             previousModel,
             nextModel: model,
@@ -1711,7 +1776,7 @@ export async function createSession(opts) {
             const spinnerStart = Date.now();
             let spinnerIdx = 0;
             let spinnerTimer;
-            if (process.stderr.isTTY) {
+            if (process.stderr.isTTY && !process.env.IDLEHANDS_QUIET) {
                 spinnerTimer = setInterval(() => {
                     const elapsedSec = Math.floor((Date.now() - spinnerStart) / 1000);
                     const frame = frames[spinnerIdx % frames.length];
@@ -1719,13 +1784,14 @@ export async function createSession(opts) {
                     process.stderr.write(`\r${frame} Server unavailable - waiting for reconnect (${elapsedSec}s)...`);
                 }, 120);
             }
-            else {
+            else if (!process.env.IDLEHANDS_QUIET) {
                 console.warn('[model] Server unavailable - waiting for reconnect...');
             }
             try {
                 await client.waitForReady({ timeoutMs: 120_000, pollMs: 2_000 });
                 fresh = normalizeModelsResponse(await client.models());
-                console.warn('[model] Reconnected to server.');
+                if (!process.env.IDLEHANDS_QUIET)
+                    console.warn('[model] Reconnected to server.');
             }
             catch {
                 return;
@@ -1733,7 +1799,8 @@ export async function createSession(opts) {
             finally {
                 if (spinnerTimer) {
                     clearInterval(spinnerTimer);
-                    process.stderr.write('\r\x1b[K');
+                    if (process.stderr.isTTY)
+                        process.stderr.write('\r\x1b[K');
                 }
             }
         }
@@ -1795,6 +1862,9 @@ export async function createSession(opts) {
                 // best effort
             }
         };
+        const isReadOnlyToolDynamic = (toolName) => {
+            return isReadOnlyTool(toolName) || LSP_TOOL_NAME_SET.has(toolName) || Boolean(mcpManager?.isToolReadOnly(toolName));
+        };
         const emitToolResult = async (result) => {
             await hookObj.onToolResult?.(result);
             await hookManager.emit('tool_result', { askId, turn: turns, result });
@@ -1934,6 +2004,8 @@ export async function createSession(opts) {
         let noToolTurns = 0;
         const NO_TOOL_REPROMPT_THRESHOLD = 2;
         let repromptUsed = false;
+        let readBudgetWarned = false;
+        let noToolNudgeUsed = false;
         // Track blocked command loops by exact reason+command signature.
         const blockedExecAttemptsBySig = new Map();
         // Cache successful read-only exec observations by exact signature.
@@ -2056,23 +2128,6 @@ export async function createSession(opts) {
             const tail = detail ? ` — ${detail}` : '';
             hookObj.onToken(`\n[sub-agent #${taskId}] ${status}${tail}\n`);
         };
-        const buildLspLensSymbolOutput = async (filePathRaw) => {
-            if (!lspManager)
-                return '[lsp] unavailable';
-            const semantic = await lspManager.getSymbols(filePathRaw);
-            if (!lens)
-                return semantic;
-            const cwd = cfg.dir ?? process.cwd();
-            const absPath = filePathRaw.startsWith('/') ? filePathRaw : path.resolve(cwd, filePathRaw);
-            const body = await fs.readFile(absPath, 'utf8').catch(() => '');
-            if (!body)
-                return semantic;
-            const projection = await lens.projectFile(absPath, body).catch(() => '');
-            const structural = extractLensBody(projection);
-            if (!structural)
-                return semantic;
-            return `${semantic}\n\n[lens] Structural skeleton:\n${structural}`;
-        };
         const runSpawnTask = async (args) => {
             if (delegationForbiddenByUser) {
                 throw new Error('spawn_task: blocked — user explicitly asked for no delegation/sub-agents in this request. Continue directly in the current session.');
@@ -2102,7 +2157,7 @@ export async function createSession(opts) {
                         messages: beforeMsgs,
                         contextWindow,
                         maxTokens: maxTokens,
-                        minTailMessages: 12,
+                        minTailMessages: cfg.compact_min_tail ?? 12,
                         compactAt: cfg.compact_at ?? 0.8,
                         toolSchemaTokens: estimateToolSchemaTokens(getToolsSchema()),
                     });
@@ -2137,13 +2192,43 @@ export async function createSession(opts) {
                         }
                     }
                     messages = compacted;
-                    // Update current context token count after auto compaction
-                    currentContextTokens = estimateTokensFromMessages(compacted);
                     if (dropped.length) {
-                        messages.push({ role: 'system', content: buildCompactionSystemNote('auto', dropped.length) });
-                        await injectVaultContext().catch(() => { });
-                        injectCompactionReminder('auto context-budget compaction');
+                        const droppedTokens = estimateTokensFromMessages(dropped);
+                        if (cfg.compact_summary !== false && droppedTokens > 200) {
+                            try {
+                                const summaryContent = buildCompactionSummaryPrompt(dropped);
+                                const summaryMaxTokens = cfg.compact_summary_max_tokens ?? 300;
+                                const resp = await client.chat({
+                                    model,
+                                    messages: [
+                                        { role: 'system', content: 'Summarize this agent session progress concisely. List: files read, key findings, decisions made, current approach. Be terse.' },
+                                        { role: 'user', content: summaryContent },
+                                    ],
+                                    max_tokens: summaryMaxTokens,
+                                    temperature: 0,
+                                    responseTimeoutMs: 5_000,
+                                });
+                                const summary = resp.choices?.[0]?.message?.content ?? '';
+                                if (summary.trim()) {
+                                    messages.push({
+                                        role: 'system',
+                                        content: `[Compacted ${dropped.length} messages (~${droppedTokens} tokens). Progress summary:]\n${summary.trim()}\n[Continue from where you left off. Do not repeat completed work.]`,
+                                    });
+                                }
+                                else {
+                                    messages.push({ role: 'system', content: buildCompactionSystemNote('auto', dropped.length) });
+                                }
+                            }
+                            catch {
+                                messages.push({ role: 'system', content: buildCompactionSystemNote('auto', dropped.length) });
+                            }
+                        }
+                        else {
+                            messages.push({ role: 'system', content: buildCompactionSystemNote('auto', dropped.length) });
+                        }
                     }
+                    // Update token count AFTER injections so downstream reads are accurate
+                    currentContextTokens = estimateTokensFromMessages(messages);
                     const afterTokens = estimateTokensFromMessages(compacted);
                     return {
                         beforeMessages: beforeMsgs.length,
@@ -2178,8 +2263,8 @@ export async function createSession(opts) {
                 let resp;
                 try {
                     try {
-                        const toolsForTurn = forceToollessRecoveryTurn ? [] : getToolsSchema();
-                        const toolChoiceForTurn = forceToollessRecoveryTurn ? 'none' : 'auto';
+                        const toolsForTurn = (cfg.no_tools || forceToollessRecoveryTurn) ? [] : getToolsSchema();
+                        const toolChoiceForTurn = (cfg.no_tools || forceToollessRecoveryTurn) ? 'none' : 'auto';
                         resp = await client.chatStream({
                             model,
                             messages,
@@ -2209,8 +2294,7 @@ export async function createSession(opts) {
                             const mode = useHardCompaction ? 'hard' : 'force';
                             messages.push({
                                 role: 'system',
-                                content: `[auto-recovery] Previous request exceeded model context window. Ran ${mode} compaction ` +
-                                    `(freed ~${compacted.freedTokens} tokens, dropped ${compacted.droppedMessages} messages). Continue from latest state; do not restart work.`,
+                                content: `[auto-recovery] Context overflow. Ran ${mode} compaction (freed ~${compacted.freedTokens} tokens). Continue current work.`,
                             });
                             continue;
                         }
@@ -2348,7 +2432,7 @@ export async function createSession(opts) {
                     }
                     messages.push({
                         role: 'user',
-                        content: '[system] Your previous response was empty (no text, no tool calls). Continue by either calling a tool with valid JSON arguments or giving a final answer.',
+                        content: '[system] Empty response. Call a tool or give final answer.',
                     });
                     await emitTurnEnd({
                         turn: turns,
@@ -2377,10 +2461,14 @@ export async function createSession(opts) {
                         for (const tc of toolCallsArr) {
                             const n = tc.function?.name ?? '';
                             let argCount = 0;
-                            try {
-                                argCount = Object.keys(parseJsonArgs(tc.function?.arguments ?? '{}')).length;
+                            // Extract arg count without full parse if possible
+                            const argStr = tc.function?.arguments ?? '{}';
+                            if (argStr.length > 2) {
+                                try {
+                                    argCount = Object.keys(parseJsonArgs(argStr)).length;
+                                }
+                                catch { }
                             }
-                            catch { }
                             if (!byName.has(n))
                                 byName.set(n, []);
                             byName.get(n).push({ tc, argCount });
@@ -2440,9 +2528,7 @@ export async function createSession(opts) {
                             mutationVersion++;
                         },
                     });
-                    const isReadOnlyToolDynamic = (toolName) => {
-                        return isReadOnlyTool(toolName) || LSP_TOOL_NAME_SET.has(toolName) || Boolean(mcpManager?.isToolReadOnly(toolName));
-                    };
+                    // Tool-call argument parsing and validation logic
                     const fileMutationsInTurn = toolCallsArr.filter((tc) => FILE_MUTATION_TOOL_SET.has(tc.function?.name)).length;
                     if (fileMutationsInTurn >= 3 && isGitDirty(ctx.cwd)) {
                         const shouldStash = confirmBridge
@@ -2502,7 +2588,7 @@ export async function createSession(opts) {
                                 });
                                 messages.push({
                                     role: 'system',
-                                    content: `[tool-loop ${warning.level}] ${warning.message}. Stop repeating ${warning.toolName} with unchanged inputs; continue with analysis or next step.`,
+                                    content: `[tool-loop ${warning.level}] ${warning.message}. Use existing results; move on.`,
                                 });
                             }
                         }
@@ -2591,9 +2677,8 @@ export async function createSession(opts) {
                             const consec = consecutiveCounts.get(sig) ?? 1;
                             const isReadFileTool = READ_FILE_CACHE_TOOLS.has(toolName);
                             const hardBreakAt = isReadFileTool ? 6 : 4;
-                            // At 3x, inject vault context and first warning
+                            // At 3x, first warning
                             if (consec >= 3) {
-                                await injectVaultContext().catch(() => { });
                                 if (consec === 3) {
                                     let warningMsg = null;
                                     if (toolName === 'read_file') {
@@ -2608,7 +2693,7 @@ export async function createSession(opts) {
                                     if (warningMsg) {
                                         messages.push({
                                             role: 'system',
-                                            content: `${warningMsg} The content has not changed between reads. Reuse the prior result and move to the next step.`,
+                                            content: `${warningMsg} Content unchanged. Reuse prior result.`,
                                         });
                                     }
                                 }
@@ -2625,7 +2710,7 @@ export async function createSession(opts) {
                                         resourceType = 'directory';
                                     messages.push({
                                         role: 'system',
-                                        content: `CRITICAL: DO NOT make another identical call for this ${resourceType}. It HAS NOT CHANGED. You already have the content. Move on to the NEXT step NOW.`,
+                                        content: `CRITICAL: ${resourceType} unchanged. Move on NOW.`,
                                     });
                                 }
                                 const argsForSig = sigMetaBySig.get(sig)?.args ?? {};
@@ -2640,8 +2725,7 @@ export async function createSession(opts) {
                                 shouldForceToollessRecovery = true;
                                 messages.push({
                                     role: 'system',
-                                    content: `[tool-loop critical] ${toolName} repeated ${consec}x with unchanged inputs. ` +
-                                        'Next turn will run with tools disabled so you must use existing results and provide a concrete next step/final response.',
+                                    content: `[tool-loop critical] ${toolName} repeated ${consec}x unchanged. Tools disabled next turn; use existing results.`,
                                 });
                             }
                             continue;
@@ -2668,8 +2752,7 @@ export async function createSession(opts) {
                             toollessRecoveryUsed = true;
                             messages.push({
                                 role: 'user',
-                                content: '[system] Critical tool loop detected. Next turn will run with tools disabled. ' +
-                                    'Use already available tool results to provide a concrete next step or final response; do not request more tools.',
+                                content: '[system] Tool loop detected. Tools disabled. Use existing results for next step.',
                             });
                             await emitTurnEnd({
                                 turn: turns,
@@ -2894,21 +2977,7 @@ export async function createSession(opts) {
                             }
                             else if (isLspTool && lspManager) {
                                 // LSP tool dispatch
-                                if (name === 'lsp_diagnostics') {
-                                    content = await lspManager.getDiagnostics(typeof args.path === 'string' ? args.path : undefined, typeof args.severity === 'number' ? args.severity : undefined);
-                                }
-                                else if (name === 'lsp_symbols') {
-                                    content = await buildLspLensSymbolOutput(String(args.path ?? ''));
-                                }
-                                else if (name === 'lsp_hover') {
-                                    content = await lspManager.getHover(String(args.path ?? ''), Number(args.line ?? 0), Number(args.character ?? 0));
-                                }
-                                else if (name === 'lsp_definition') {
-                                    content = await lspManager.getDefinition(String(args.path ?? ''), Number(args.line ?? 0), Number(args.character ?? 0));
-                                }
-                                else if (name === 'lsp_references') {
-                                    content = await lspManager.getReferences(String(args.path ?? ''), Number(args.line ?? 0), Number(args.character ?? 0), typeof args.max_results === 'number' ? args.max_results : 50);
-                                }
+                                content = await dispatchLspTool(name, args);
                             }
                             else {
                                 if (mcpManager == null) {
@@ -3201,11 +3270,11 @@ export async function createSession(opts) {
                     }
                     // ── Escalating cumulative read budget (§ anti-scan guardrails) ──
                     // Warn zone: append warnings to each read result when approaching the hard cap
-                    if (cumulativeReadOnlyCalls > READ_BUDGET_WARN && cumulativeReadOnlyCalls <= READ_BUDGET_HARD) {
-                        const remaining = READ_BUDGET_HARD - cumulativeReadOnlyCalls;
+                    if (!readBudgetWarned && cumulativeReadOnlyCalls > READ_BUDGET_WARN && cumulativeReadOnlyCalls <= READ_BUDGET_HARD) {
+                        readBudgetWarned = true;
                         messages.push({
                             role: 'user',
-                            content: `[System] ⚠ Read budget: ${cumulativeReadOnlyCalls}/${READ_BUDGET_HARD}. ${remaining} reads remaining before hard stop. Use search_files or exec grep — do NOT continue reading files one at a time.`,
+                            content: `[system] Read budget: ${cumulativeReadOnlyCalls}/${READ_BUDGET_HARD}. Use search_files instead of reading files individually.`,
                         });
                     }
                     // One bounded automatic repair attempt for invalid tool args.
@@ -3248,7 +3317,8 @@ export async function createSession(opts) {
                 const assistantText = visible || content || '';
                 // Recovery fuse: if the model keeps narrating/planning without tool use,
                 // nudge it once with the original task. Never resend more than once per ask().
-                if (looksLikePlanningNarration(assistantText, finishReason)) {
+                // Skip this check entirely when no_tools is set — text IS the final answer.
+                if (!cfg.no_tools && looksLikePlanningNarration(assistantText, finishReason)) {
                     noToolTurns += 1;
                     messages.push({ role: 'assistant', content: assistantText });
                     if (noToolTurns >= NO_TOOL_REPROMPT_THRESHOLD) {
@@ -3256,12 +3326,10 @@ export async function createSession(opts) {
                             repromptUsed = true;
                             noToolTurns = 0;
                             const reminder = userContentToText(instruction).trim();
-                            const clippedReminder = reminder.length > 4000 ? `${reminder.slice(0, 4000)}\n[truncated]` : reminder;
+                            const clippedReminder = reminder.length > 1600 ? `${reminder.slice(0, 1600)}\n[truncated]` : reminder;
                             messages.push({
                                 role: 'user',
-                                content: `[system] You seem to be stuck narrating without using tools. Resume execution now.\n` +
-                                    `Original task:\n${clippedReminder}\n\n` +
-                                    `Call the needed tools directly. If everything is truly complete, provide the final answer.`
+                                content: `[system] Stuck narrating. Resume with tools.\nTask:\n${clippedReminder}`
                             });
                             await emitTurnEnd({
                                 turn: turns,
@@ -3279,10 +3347,28 @@ export async function createSession(opts) {
                         }
                         throw new Error(`no-tool loop detected: model produced planning/narration without tool calls for ${NO_TOOL_REPROMPT_THRESHOLD} turns even after one recovery reprompt`);
                     }
-                    messages.push({
-                        role: 'user',
-                        content: '[system] Continue executing the task. Use tools now (do not just narrate plans). If complete, give the final answer.'
-                    });
+                    if (!noToolNudgeUsed) {
+                        noToolNudgeUsed = true;
+                        messages.push({
+                            role: 'user',
+                            content: '[system] Use tools now or give final answer.'
+                        });
+                        await emitTurnEnd({
+                            turn: turns,
+                            toolCalls,
+                            promptTokens: cumulativeUsage.prompt,
+                            completionTokens: cumulativeUsage.completion,
+                            promptTokensTurn,
+                            completionTokensTurn,
+                            ttftMs,
+                            ttcMs,
+                            ppTps,
+                            tgTps,
+                        });
+                        continue;
+                    }
+                    // Nudge already used — fall through to next iteration which will
+                    // increment noToolTurns and hit the reprompt threshold.
                     await emitTurnEnd({
                         turn: turns,
                         toolCalls,
@@ -3402,7 +3488,7 @@ export async function createSession(opts) {
         get capturePath() {
             return capturePath;
         },
-        getSystemPrompt: () => activeSystemPrompt,
+        getSystemPrompt: () => messages[0]?.role === 'system' ? String(messages[0].content) : activeSystemPromptBase,
         setSystemPrompt,
         resetSystemPrompt,
         listMcpServers,