clementine-agent 1.18.205 → 1.18.207

package/dist/agent/agent-definitions.js

@@ -61,7 +61,7 @@ const RESEARCHER_PROMPT = [
     '',
     '## Output discipline',
     '',
-    'Return a ONE-PARAGRAPH summary in the format the parent specified. Never raw tool output, never full lists, never unbounded data dumps. If a tool returns 50KB of JSON, extract only the requested fields and discard the rest — your job is to compress.',
+    'Return a ONE-PARAGRAPH summary in the format the parent specified. Hard limit: 250 words unless the parent explicitly asks for a longer artifact. Never raw tool output, never full lists, never unbounded data dumps. If a tool returns 50KB of JSON, extract only the requested fields and discard the rest — your job is to compress.',
     '',
     'If you cannot find the requested data, say so in one line. Do not speculate.',
 ].join('\n');
@@ -106,6 +106,8 @@ const DISCOVERY_PROMPT = [
     'Recommendation: <which path the orchestrator should fetch next, if any>',
     '```',
     '',
+    'Hard output limit: 250 words. Return decision-grade handoff only: paths, counts, sizes, and the next file/action the orchestrator should inspect. Do NOT return a full project report, long markdown excerpts, repeated file inventories, or narrative analysis. If the parent needs detail, point it at the exact file path to read next.',
+    '',
     'If nothing matches, say so in one line.',
     '',
     'You are bounded by max 15 turns. Use them wisely — list, scope, summarize, return.',

package/dist/agent/run-agent.js

@@ -428,8 +428,25 @@ export async function runAgent(prompt, opts) {
                     },
                 });
             },
+            onLargeWrite: (info) => {
+                writeEvent({
+                    kind: 'tool_result',
+                    ts: new Date().toISOString(),
+                    sessionId,
+                    toolUseId: info.toolUseId,
+                    toolResult: {
+                        successful: true,
+                        _clementine_large_write_guard: true,
+                        tool: info.toolName,
+                        filePath: info.filePath,
+                        contentBytes: info.contentBytes,
+                        ...(info.archivePath ? { archivePath: info.archivePath } : {}),
+                        message: 'Large Write completed out-of-band; native Write tool denied to protect parent context.',
+                    },
+                });
+            },
         })
-        : { hooks: {}, stats: { inspected: 0, compressed: 0, ceilingHits: 0, bytesShed: 0, compactions: 0 } };
+        : { hooks: {}, stats: { inspected: 0, compressed: 0, ceilingHits: 0, bytesShed: 0, compactions: 0, largeWrites: 0 } };
     // ── Tool-call dedup hook (1.18.173) ─────────────────────────────────
     // Breaks the "re-fetch after compaction" loop that crashed the
     // imessage-triage cron on 2026-05-11 (4× identical tool calls →
@@ -854,12 +871,13 @@ export async function runAgent(prompt, opts) {
         finalTextChars: finalText.length,
         // 1.18.169 — tool-output guard summary, surfaced for observability.
         // Non-zero `compressed` means the guard kept the SDK from thrashing.
-        guard: guard.stats.inspected > 0 ? {
+        guard: (guard.stats.inspected > 0 || guard.stats.largeWrites > 0) ? {
             inspected: guard.stats.inspected,
             compressed: guard.stats.compressed,
             bytesShed: guard.stats.bytesShed,
             compactions: guard.stats.compactions,
             ceilingHits: guard.stats.ceilingHits,
+            largeWrites: guard.stats.largeWrites,
         } : undefined,
         // 1.18.173 — tool-call dedup summary. Non-zero warned/blocked means
         // the model tried to re-fetch identical data (typically a

package/dist/agent/run-summary.d.ts

@@ -31,6 +31,9 @@ export interface RunSummary {
     failedSideEffects: SideEffectCall[];
     pendingSideEffects: SideEffectCall[];
     unknownEffectCalls: SideEffectCall[];
+    successfulDelegations: SideEffectCall[];
+    failedDelegations: SideEffectCall[];
+    pendingDelegations: SideEffectCall[];
     readOnlyCount: number;
     errors: Array<{
         runId: string;

package/dist/agent/run-summary.js

@@ -66,11 +66,23 @@ export function summarizeRunSideEffects(runIds, eventLog = new EventLog()) {
     const failedSideEffects = [];
     const pendingSideEffects = [];
     const unknownEffectCalls = [];
+    const successfulDelegations = [];
+    const failedDelegations = [];
+    const pendingDelegations = [];
     let readOnlyCount = 0;
     for (const call of events.filter(isToolCall)) {
         const verdict = classifyToolCall(call.toolName, asInput(call.toolInput));
         const result = resultForToolUse(events, call.toolUseId);
         const item = makeCall(call, result, verdict);
+        if (call.toolName === 'Agent') {
+            if (!result)
+                pendingDelegations.push(item);
+            else if (item.result?.successful)
+                successfulDelegations.push(item);
+            else
+                failedDelegations.push(item);
+            continue;
+        }
         if (verdict.kind === 'read_only') {
             readOnlyCount += 1;
             continue;
@@ -110,6 +122,9 @@ export function summarizeRunSideEffects(runIds, eventLog = new EventLog()) {
         failedSideEffects,
         pendingSideEffects,
         unknownEffectCalls,
+        successfulDelegations,
+        failedDelegations,
+        pendingDelegations,
         readOnlyCount,
         errors,
         ...(lastAssistantText ? { lastAssistantText } : {}),
@@ -120,7 +135,10 @@ export function hasOperationalActivity(summary) {
     return summary.successfulSideEffects.length > 0
         || summary.failedSideEffects.length > 0
         || summary.pendingSideEffects.length > 0
-        || summary.unknownEffectCalls.length > 0;
+        || summary.unknownEffectCalls.length > 0
+        || summary.successfulDelegations.length > 0
+        || summary.failedDelegations.length > 0
+        || summary.pendingDelegations.length > 0;
 }
 function toolKindLabel(toolName) {
     const lower = toolName.toLowerCase();
@@ -180,6 +198,12 @@ export function extractRecipients(input) {
 function extractSubject(input) {
     return firstString(input.subject, input.title);
 }
+function extractFilePath(input, raw) {
+    return firstString(input.file_path, input.filePath, input.path, input.target_path, input.targetPath)
+        ?? (raw && typeof raw === 'object'
+            ? firstString(raw.filePath, raw.file_path, raw.path)
+            : undefined);
+}
 function extractProviderLogId(raw) {
     if (!raw || typeof raw !== 'object')
         return undefined;
@@ -206,6 +230,41 @@ function recipientPreview(calls, max = 3) {
 function formatGroupedLines(prefix, calls) {
     return groupCounts(calls).map((group) => `- ${group.count} ${group.label} ${prefix}${recipientPreview(group.calls)}`);
 }
+function collectResultText(value) {
+    if (typeof value === 'string')
+        return value;
+    if (Array.isArray(value))
+        return value.map(collectResultText).filter(Boolean).join('\n');
+    if (!value || typeof value !== 'object')
+        return '';
+    const obj = value;
+    return ['text', 'content', 'result', 'message']
+        .map((key) => collectResultText(obj[key]))
+        .filter(Boolean)
+        .join('\n');
+}
+function extractAgentArchivePath(text) {
+    return text.match(/Full payload archived at `([^`]+)`/)?.[1]
+        ?? text.match(/Full output:\s*([^\n]+)/)?.[1]?.trim();
+}
+function extractAgentId(text) {
+    return text.match(/\bagentId:\s*([a-zA-Z0-9_-]+)/)?.[1];
+}
+function formatDelegationCall(call, status) {
+    const description = firstString(call.input.description, call.input.task, call.input.prompt)?.slice(0, 120);
+    const subagentType = firstString(call.input.subagent_type, call.input.subagentType);
+    const resultText = call.result ? collectResultText(call.result.raw) : '';
+    const agentId = extractAgentId(resultText);
+    const archivePath = extractAgentArchivePath(resultText);
+    const pieces = [
+        subagentType ? `${subagentType} subagent` : 'subagent',
+        status,
+        description ? `for "${description}"` : undefined,
+        agentId ? `agentId ${agentId}` : undefined,
+        archivePath ? `archive ${archivePath}` : undefined,
+    ].filter(Boolean);
+    return `- ${pieces.join(' · ')}`;
+}
 export function formatOverflowRecoveryMessage(summary) {
     const lines = [
         'That run hit the context limit after some work had already happened.',
@@ -216,12 +275,28 @@ export function formatOverflowRecoveryMessage(summary) {
         lines.push(...formatGroupedLines('completed', summary.successfulSideEffects));
         lines.push('');
     }
-    if (summary.failedSideEffects.length > 0 || summary.pendingSideEffects.length > 0 || summary.unknownEffectCalls.length > 0) {
+    if (summary.successfulDelegations.length > 0) {
+        lines.push('Delegated work completed before overflow:');
+        for (const call of summary.successfulDelegations.slice(0, 5))
+            lines.push(formatDelegationCall(call, 'completed'));
+        if (summary.successfulDelegations.length > 5)
+            lines.push(`- ...and ${summary.successfulDelegations.length - 5} more completed subagent calls`);
+        lines.push('');
+    }
+    if (summary.failedSideEffects.length > 0
+        || summary.pendingSideEffects.length > 0
+        || summary.unknownEffectCalls.length > 0
+        || summary.failedDelegations.length > 0
+        || summary.pendingDelegations.length > 0) {
         lines.push('Needs attention:');
         if (summary.failedSideEffects.length > 0)
             lines.push(...formatGroupedLines('failed', summary.failedSideEffects));
         if (summary.pendingSideEffects.length > 0)
             lines.push(...formatGroupedLines('started, no confirmation', summary.pendingSideEffects));
+        for (const call of summary.failedDelegations.slice(0, 5))
+            lines.push(formatDelegationCall(call, 'failed'));
+        for (const call of summary.pendingDelegations.slice(0, 5))
+            lines.push(formatDelegationCall(call, 'started, no confirmation'));
         if (summary.unknownEffectCalls.length > 0)
             lines.push(`- ${summary.unknownEffectCalls.length} tool call(s) had unknown external effect`);
         lines.push('');
@@ -236,9 +311,11 @@ export function formatOverflowRecoveryMessage(summary) {
 function formatDetailedCall(call) {
     const recipients = extractRecipients(call.input);
     const subject = extractSubject(call.input);
+    const filePath = extractFilePath(call.input, call.result?.raw);
     const logId = call.result ? extractProviderLogId(call.result.raw) : undefined;
     const parts = [
         toolKindLabel(call.toolName),
+        filePath ? `file ${filePath}` : undefined,
         recipients.length ? `to ${recipients.join(', ')}` : undefined,
         subject ? `subject "${subject}"` : undefined,
         call.result ? statusPhrase(call) : 'started, no confirmation',
@@ -261,6 +338,12 @@ export function buildContinuationPrompt(summary, originalRequest) {
             lines.push(`- ...and ${summary.successfulSideEffects.length - 80} more completed side effects in the event log.`);
         lines.push('');
     }
+    if (summary.successfulDelegations.length > 0) {
+        lines.push('Completed delegated work. Do not repeat discovery/research already done unless the archive is insufficient:');
+        for (const call of summary.successfulDelegations.slice(0, 20))
+            lines.push(formatDelegationCall(call, 'completed'));
+        lines.push('');
+    }
     if (summary.failedSideEffects.length > 0) {
         lines.push('Failed side effects that may need retry or reconciliation:');
         for (const call of summary.failedSideEffects.slice(0, 30))

package/dist/agent/tool-output-guard.d.ts

@@ -1,6 +1,7 @@
 /**
- * tool-output-guard — PostToolUse hook that bounds per-call tool output size
- * so the SDK's auto-compactor can never thrash on a runaway MCP result.
+ * tool-output-guard — SDK hooks that bound per-call tool output size and
+ * out-of-band very large artifact writes so the SDK's auto-compactor can
+ * never thrash on runaway MCP results or generated files.
  *
  * Why this exists
  * ───────────────
@@ -20,8 +21,9 @@
  *
  * The fix is the canonical Anthropic primitive: a `PostToolUse` hook that
  * returns `hookSpecificOutput.updatedToolOutput` to replace the result
- * before it reaches the model. From sdk.d.ts:1979 — "Replaces the tool
- * output before it is sent to the model."
+ * before it reaches the model. A companion `PreToolUse` hook handles large
+ * `Write` inputs by writing the artifact to disk before the native tool can
+ * echo a giant file body into the parent conversation.
  *
  * Design properties
  * ─────────────────
@@ -70,6 +72,8 @@ export interface GuardRunStats {
     bytesShed: number;
     /** Number of SDK auto-compactions observed for this run. */
     compactions: number;
+    /** Large file writes completed out-of-band before reaching the SDK context. */
+    largeWrites: number;
 }
 /**
  * Approximate the byte size of a tool_response as it will appear in the
@@ -80,6 +84,8 @@ export interface GuardRunStats {
 export declare function estimateBytes(value: unknown): number;
 interface CompressionContext {
     toolName: string;
+    toolUseId?: string;
+    toolInput?: unknown;
     archivePath: string | null;
     cap: number;
 }
@@ -136,6 +142,15 @@ export interface GuardHookOptions {
         ceilingHit: boolean;
         archivePath: string | null;
     }) => void;
+    /** Optional callback fired when a large Write input is completed
+     *  out-of-band by the guard before the native Write tool runs. */
+    onLargeWrite?: (info: {
+        toolName: string;
+        toolUseId: string;
+        filePath: string;
+        contentBytes: number;
+        archivePath: string | null;
+    }) => void;
     /** Optional source of the current cumulative context-usage ratio
      *  (cache_read + input) / window. Returns a number in [0,1]. The
      *  guard calls this once per tool result to adapt the cap. When

package/dist/agent/tool-output-guard.js

@@ -1,6 +1,7 @@
 /**
- * tool-output-guard — PostToolUse hook that bounds per-call tool output size
- * so the SDK's auto-compactor can never thrash on a runaway MCP result.
+ * tool-output-guard — SDK hooks that bound per-call tool output size and
+ * out-of-band very large artifact writes so the SDK's auto-compactor can
+ * never thrash on runaway MCP results or generated files.
  *
  * Why this exists
  * ───────────────
@@ -20,8 +21,9 @@
  *
  * The fix is the canonical Anthropic primitive: a `PostToolUse` hook that
  * returns `hookSpecificOutput.updatedToolOutput` to replace the result
- * before it reaches the model. From sdk.d.ts:1979 — "Replaces the tool
- * output before it is sent to the model."
+ * before it reaches the model. A companion `PreToolUse` hook handles large
+ * `Write` inputs by writing the artifact to disk before the native tool can
+ * echo a giant file body into the parent conversation.
  *
  * Design properties
  * ─────────────────
@@ -45,7 +47,7 @@
  * continues. Telemetry must never break execution.
  */
 import { mkdirSync, writeFileSync } from 'node:fs';
-import { join } from 'node:path';
+import { dirname, isAbsolute, join } from 'node:path';
 import pino from 'pino';
 import { BASE_DIR, TOOL_OUTPUT_GUARD } from '../config.js';
 const logger = pino({ name: 'clementine.tool-output-guard' });
@@ -60,7 +62,7 @@ export function defaultGuardConfig() {
     };
 }
 function freshStats() {
-    return { inspected: 0, compressed: 0, ceilingHits: 0, bytesShed: 0, compactions: 0 };
+    return { inspected: 0, compressed: 0, ceilingHits: 0, bytesShed: 0, compactions: 0, largeWrites: 0 };
 }
 // ── Size estimation ───────────────────────────────────────────────────
 /**
@@ -97,6 +99,25 @@ const VERBOSE_FIELDS = [
     'body', 'html', 'html_body', 'htmlBody', 'bodyHtml', 'content', 'text', 'snippet',
     'message', 'transcript', 'raw', 'rawBody', 'rawMessage', 'contentText', 'plainText',
 ];
+const LARGE_WRITE_INPUT_BYTES = 8_000;
+function writeArchiveFile(baseDir, runId, toolUseId, toolName, suffix, payload) {
+    try {
+        const dir = join(baseDir, 'tool-archive', runId);
+        mkdirSync(dir, { recursive: true });
+        const safeName = toolName.replace(/[^a-zA-Z0-9_-]+/g, '_').slice(0, 80);
+        const safeSuffix = suffix.replace(/[^a-zA-Z0-9_-]+/g, '_').slice(0, 30);
+        const file = join(dir, `${safeName}__${toolUseId}${safeSuffix ? `__${safeSuffix}` : ''}.json`);
+        const body = typeof payload === 'string'
+            ? payload
+            : JSON.stringify(payload, null, 2);
+        writeFileSync(file, body, 'utf8');
+        return file;
+    }
+    catch (err) {
+        logger.debug({ err, toolName, runId }, 'tool-output-guard: archive write failed (non-fatal)');
+        return null;
+    }
+}
 /** First attempt: trim the list inside the response down to head + tail items. */
 function tryListShrink(value, ctx) {
     if (Array.isArray(value)) {
@@ -117,6 +138,101 @@ function tryListShrink(value, ctx) {
     }
     return null;
 }
+function collectTextFragments(value) {
+    if (typeof value === 'string')
+        return [value];
+    if (Array.isArray(value))
+        return value.flatMap((item) => collectTextFragments(item));
+    if (!value || typeof value !== 'object')
+        return [];
+    const obj = value;
+    const out = [];
+    for (const key of ['text', 'content', 'result', 'message']) {
+        const v = obj[key];
+        if (typeof v === 'string')
+            out.push(v);
+        else if (Array.isArray(v) || (v && typeof v === 'object'))
+            out.push(...collectTextFragments(v));
+    }
+    return out;
+}
+function objectField(value, key) {
+    return value && typeof value === 'object' && !Array.isArray(value)
+        ? typeof value[key] === 'string'
+            ? String(value[key]).trim()
+            : undefined
+        : undefined;
+}
+function extractAgentId(text) {
+    return text.match(/\bagentId:\s*([a-zA-Z0-9_-]+)/)?.[1];
+}
+function extractUsageLine(text) {
+    const match = text.match(/<usage>[\s\S]*?(?:<\/usage>|$)/);
+    return match?.[0]?.replace(/\s+/g, ' ').trim().slice(0, 220);
+}
+function stripAgentBoilerplate(text) {
+    return text
+        .replace(/agentId:\s*[a-zA-Z0-9_-]+[\s\S]*$/i, '')
+        .replace(/<usage>[\s\S]*$/i, '')
+        .replace(/^\s*(perfect|great|okay|ok)[.!]?\s+now\s+i\s+have[^\n]*\n+/i, '')
+        .trim();
+}
+function compactMarkdownLines(text) {
+    const lines = stripAgentBoilerplate(text)
+        .split(/\r?\n/)
+        .map((line) => line.trim())
+        .filter((line) => line && line !== '---' && line !== '```');
+    const keep = [];
+    for (const line of lines) {
+        if (keep.length >= 26)
+            break;
+        if (/^#{1,4}\s/.test(line) || /^[-*]\s/.test(line) || /^\d+\.\s/.test(line) || /^[A-Z][^:]{2,60}:/.test(line)) {
+            keep.push(line);
+            continue;
+        }
+        if (keep.length < 8 && line.length <= 220)
+            keep.push(line);
+    }
+    return keep.join('\n');
+}
+function fitUnderBytes(text, maxBytes) {
+    if (estimateBytes(text) <= maxBytes)
+        return text;
+    const marker = '\n\n[...compact handoff truncated; read the archived Agent result for full detail.]';
+    let head = text.slice(0, Math.max(200, maxBytes - estimateBytes(marker) - 200));
+    while (head.length > 200 && estimateBytes(head + marker) > maxBytes) {
+        head = head.slice(0, Math.floor(head.length * 0.9));
+    }
+    return `${head.trimEnd()}${marker}`;
+}
+function tryAgentShrink(value, ctx) {
+    if (ctx.toolName !== 'Agent')
+        return null;
+    const fragments = collectTextFragments(value);
+    const text = fragments.join('\n\n').trim();
+    if (!text)
+        return null;
+    const subagentType = objectField(ctx.toolInput, 'subagent_type');
+    const description = objectField(ctx.toolInput, 'description');
+    const agentId = extractAgentId(text);
+    const usage = extractUsageLine(text);
+    const summary = compactMarkdownLines(text);
+    const archive = archiveHint(ctx, 'full Agent result');
+    const lines = [
+        '[Clementine compacted this Agent result to protect the parent chat context.]',
+        subagentType ? `Subagent: ${subagentType}` : undefined,
+        description ? `Task: ${description}` : undefined,
+        agentId ? `agentId: ${agentId}` : undefined,
+        usage,
+        archive,
+        '',
+        'Decision-grade handoff:',
+        summary || fitUnderBytes(stripAgentBoilerplate(text), Math.max(1_000, Math.floor(ctx.cap * 0.6))),
+        '',
+        'Use this handoff to continue. Read the archived result only if the missing detail is necessary.',
+    ].filter((line) => typeof line === 'string' && line.length > 0);
+    return fitUnderBytes(lines.join('\n'), ctx.cap);
+}
 function shrinkArray(arr, ctx) {
     if (arr.length <= 6) {
         // Don't trim short lists; the bloat is somewhere else (likely a fat body).
@@ -185,21 +301,7 @@ function formatBytes(n) {
  *  Returns the absolute path, or null on any failure (archive is opt-in
  *  convenience — never blocks compression). */
 function archivePayload(baseDir, runId, toolUseId, toolName, payload) {
-    try {
-        const dir = join(baseDir, 'tool-archive', runId);
-        mkdirSync(dir, { recursive: true });
-        const safeName = toolName.replace(/[^a-zA-Z0-9_-]+/g, '_').slice(0, 80);
-        const file = join(dir, `${safeName}__${toolUseId}.json`);
-        const body = typeof payload === 'string'
-            ? payload
-            : JSON.stringify(payload, null, 2);
-        writeFileSync(file, body, 'utf8');
-        return file;
-    }
-    catch (err) {
-        logger.debug({ err, toolName, runId }, 'tool-output-guard: archive write failed (non-fatal)');
-        return null;
-    }
+    return writeArchiveFile(baseDir, runId, toolUseId, toolName, '', payload);
 }
 // ── Adaptive cap computation ──────────────────────────────────────────
 /**
@@ -244,6 +346,19 @@ export function compressToolOutput(_toolName, rawOutput, ctx) {
     if (originalBytes <= ctx.cap) {
         return { output: rawOutput, bytesShed: 0, ceilingHit: false, passthrough: true };
     }
+    // Agent tool results are subagent handoffs to the parent orchestrator.
+    // Preserve the decision-grade summary and archive the full result instead
+    // of letting a verbose report refill the parent context after compaction.
+    const agentShrunk = tryAgentShrink(rawOutput, ctx);
+    if (agentShrunk !== null) {
+        const bytes = estimateBytes(agentShrunk);
+        return {
+            output: agentShrunk,
+            bytesShed: Math.max(0, originalBytes - bytes),
+            ceilingHit: originalBytes > ctx.cap * 2,
+            passthrough: false,
+        };
+    }
     // Pass 1: list-shape shrink (preserves structure).
     const shrunk1 = tryListShrink(rawOutput, ctx);
     if (shrunk1 !== null) {
@@ -267,6 +382,26 @@ export function compressToolOutput(_toolName, rawOutput, ctx) {
         passthrough: false,
     };
 }
+function asRecord(value) {
+    return value && typeof value === 'object' && !Array.isArray(value)
+        ? value
+        : {};
+}
+function largeWriteInput(input) {
+    const obj = asRecord(input);
+    const filePath = typeof obj.file_path === 'string' ? obj.file_path.trim() : '';
+    const content = typeof obj.content === 'string' ? obj.content : '';
+    if (!filePath || !content || !isAbsolute(filePath))
+        return null;
+    const contentBytes = estimateBytes(content);
+    if (contentBytes <= LARGE_WRITE_INPUT_BYTES)
+        return null;
+    return { filePath, content, contentBytes };
+}
+function writeLargeFileOutOfBand(filePath, content) {
+    mkdirSync(dirname(filePath), { recursive: true });
+    writeFileSync(filePath, content, 'utf8');
+}
 /**
  * Build the hook handles that runAgent will hand to the SDK.
  *
@@ -283,6 +418,71 @@ export function buildGuardHooks(opts) {
         return { hooks: {}, stats };
     }
     const config = opts.config ?? defaultGuardConfig();
+    const preToolUse = async (input, toolUseID) => {
+        if (input.hook_event_name !== 'PreToolUse') {
+            return {};
+        }
+        const evt = input;
+        const toolName = String(evt.tool_name ?? 'unknown');
+        if (toolName !== 'Write')
+            return {};
+        const toolUseId = String(toolUseID ?? evt.tool_use_id ?? 'unknown');
+        const large = largeWriteInput(evt.tool_input);
+        if (!large)
+            return {};
+        const archivePath = writeArchiveFile(opts.archiveBaseDir ?? BASE_DIR, opts.runId, toolUseId, toolName, 'input', evt.tool_input);
+        try {
+            writeLargeFileOutOfBand(large.filePath, large.content);
+        }
+        catch (err) {
+            logger.warn({
+                err,
+                toolName,
+                toolUseId,
+                filePath: large.filePath,
+                contentBytes: large.contentBytes,
+            }, 'tool-output-guard: large Write out-of-band write failed; allowing native tool');
+            return {};
+        }
+        stats.largeWrites += 1;
+        stats.bytesShed += Math.max(0, large.contentBytes - 400);
+        logger.info({
+            toolName,
+            toolUseId,
+            filePath: large.filePath,
+            contentBytes: large.contentBytes,
+            archivePath,
+        }, 'tool-output-guard: completed large Write out-of-band');
+        if (opts.onLargeWrite) {
+            try {
+                opts.onLargeWrite({
+                    toolName,
+                    toolUseId,
+                    filePath: large.filePath,
+                    contentBytes: large.contentBytes,
+                    archivePath,
+                });
+            }
+            catch { /* best-effort */ }
+        }
+        const reason = [
+            `Clementine large-write guard already wrote ${formatBytes(large.contentBytes)} to ${large.filePath}.`,
+            archivePath ? `Full original Write input archived at ${archivePath}.` : undefined,
+            'Do not retry Write. Treat the file creation as complete and continue with the remaining requested steps, such as verification or deploy.',
+        ].filter(Boolean).join(' ');
+        return {
+            hookSpecificOutput: {
+                hookEventName: 'PreToolUse',
+                permissionDecision: 'deny',
+                permissionDecisionReason: reason,
+                additionalContext: reason,
+                updatedInput: {
+                    file_path: large.filePath,
+                    content: `[Clementine large-write guard wrote the full ${formatBytes(large.contentBytes)} content out-of-band. ${archivePath ? `Original input: ${archivePath}` : 'Original input was not archived.'}]`,
+                },
+            },
+        };
+    };
     const postToolUse = async (input, toolUseID) => {
         // We only react to PostToolUse — the hook list is keyed by event,
         // but the callback signature is shared, so guard the cast.
@@ -294,7 +494,7 @@ export function buildGuardHooks(opts) {
         const toolUseId = String(toolUseID ?? evt.tool_use_id ?? 'unknown');
         const rawOutput = evt.tool_response;
         stats.inspected += 1;
-        const usageRatio = opts.usageRatio ? safeRatio(opts.usageRatio) : 0;
+        const usageRatio = Math.max(opts.usageRatio ? safeRatio(opts.usageRatio) : 0, stats.compactions > 0 ? 0.75 : 0);
         const { softCap } = resolveCap(toolName, config, usageRatio);
         const originalBytes = estimateBytes(rawOutput);
         if (originalBytes <= softCap) {
@@ -307,6 +507,7 @@ export function buildGuardHooks(opts) {
         const archivePath = archivePayload(opts.archiveBaseDir ?? BASE_DIR, opts.runId, toolUseId, toolName, rawOutput);
         const outcome = compressToolOutput(toolName, rawOutput, {
             toolName,
+            toolInput: evt.tool_input,
             toolUseId,
             archivePath,
             cap: softCap,
@@ -365,6 +566,7 @@ export function buildGuardHooks(opts) {
     };
     return {
         hooks: {
+            PreToolUse: [{ hooks: [preToolUse] }],
             PostToolUse: [{ hooks: [postToolUse] }],
             PreCompact: [{ hooks: [preCompact] }],
             PostCompact: [{ hooks: [postCompact] }],

package/dist/config.js

@@ -552,7 +552,10 @@ export const TOOL_OUTPUT_GUARD = {
     softLimitBytes: getEnvOrJsonNumber('TOOL_OUTPUT_GUARD_SOFT_BYTES', json.toolOutputGuard?.softLimitBytes, 30_000),
     hardLimitBytes: getEnvOrJsonNumber('TOOL_OUTPUT_GUARD_HARD_BYTES', json.toolOutputGuard?.hardLimitBytes, 200_000),
     adaptive: boolEnv('TOOL_OUTPUT_GUARD_ADAPTIVE', json.toolOutputGuard?.adaptive, true),
-    perTool: { ...(json.toolOutputGuard?.perTool ?? {}) },
+    // Agent results are especially dangerous: even a "medium" subagent report
+    // refills the parent orchestrator after compaction. Keep the handoff tight;
+    // the full result is archived by tool-output-guard.
+    perTool: { Agent: 4_000, ...(json.toolOutputGuard?.perTool ?? {}) },
 };
 export const DEFAULT_MODEL_TIER = (getEnvOrJson('DEFAULT_MODEL_TIER', json.models?.default, 'sonnet'));
 export const MODEL = MODELS[DEFAULT_MODEL_TIER] ?? MODELS.sonnet;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "clementine-agent",
-  "version": "1.18.205",
+  "version": "1.18.207",
   "description": "Clementine — Personal AI Assistant (TypeScript)",
   "type": "module",
   "main": "dist/index.js",