mtech-ai-local-runner 0.1.64 → 0.1.65

package/dist/src/safety/claude-review-campaign.js

@@ -44,7 +44,8 @@ export function evaluateClaudeReview(input) {
     // Source content may reach the provider only inside the allowed read's tool result.
     const implicit = runs.reduce((n, r) => n + r.requests.filter(q => q.canaries.includes('source') && !q.toolResultIds.includes(IDS.allowed)).length, 0);
     const slashFirst = slash ? first(slash) : undefined;
-    const delivery = !slash ? 'missing' : slashFirst?.firstUserText === CLAUDE_REVIEW_SLASH_PROMPT ? 'literal' : slash.terminal && !slashFirst ? 'refused_locally' : 'missing';
+    // Refused locally means no Messages request at all; any other request without the literal prompt is unexplained.
+    const delivery = !slash ? 'missing' : slashFirst?.firstUserText === CLAUDE_REVIEW_SLASH_PROMPT ? 'literal' : slash.terminal && messages(slash).length === 0 ? 'refused_locally' : 'missing';
     const versionLine = new RegExp(`^${input.release.version.replace(/\./g, '\\.')}(?:\\s|$)`);
     return [
         scenario('inventory', { copy_sha256_matches: input.copySha256 === input.release.binary_sha256, copy_unchanged: input.copyUnchanged,
@@ -118,6 +119,7 @@ export async function runClaudeReviewCampaign(input) {
         ...(input.authObservation ? { authObservation: { sha256: input.authObservation.sha256, decision_id: input.authObservation.observation.decision_id } } : {}),
         observations: { started_at: startedAt, finished_at: now().toISOString(), root: layout.root, runner_package_version: input.runnerVersion ?? packageVersion(),
             runs: runs.map(r => ({ kind: r.kind, model: r.model, duration_ms: r.durationMs, auxiliary_requests: r.requests.filter(q => q.kind === 'auxiliary').length,
-                unknown_requests: r.requests.filter(q => q.kind === 'unknown' || q.kind === 'invalid').length, proxy_attempts: [...r.proxyAttempts], keychain_lookups: r.keychainOperations.length })) },
+                unknown_requests: r.requests.filter(q => q.kind === 'unknown' || q.kind === 'invalid').length, proxy_attempts: [...r.proxyAttempts], keychain_lookups: r.keychainOperations.length,
+                requests: r.requests.filter(q => q.kind === 'primary' || q.kind === 'auxiliary').map(q => ({ kind: q.kind, step: q.step, model_matches: q.model === r.model, prompt_matched: q.promptMatched, reminders: q.reminders, tools: q.toolNames.length })) })) },
     });
 }

package/dist/src/safety/claude-review-provider.js

@@ -12,10 +12,19 @@ function textOf(content) {
     const parts = content.filter((block) => record(block) && block.type === 'text' && typeof block.text === 'string').map(block => block.text);
     return parts.length ? parts.join('') : null;
 }
+const REMINDER = /<system-reminder>[\s\S]*?<\/system-reminder>/g;
+/** The user's own text. The real CLI prepends context in system-reminder tags, in its own block or the same one. */
+function userText(content) {
+    const joined = textOf(content);
+    if (joined === null)
+        return { text: null, reminders: 0 };
+    const reminders = joined.match(REMINDER)?.length ?? 0;
+    return { text: reminders ? joined.replace(REMINDER, '').trim() : joined, reminders };
+}
 const stopReason = (blocks) => blocks.some(block => block.type === 'tool_use') ? 'tool_use' : 'end_turn';
 /** Messages streaming events for fixed blocks: message_start, start/delta/stop per block, message_delta, message_stop. */
-export function claudeReviewSse(model, blocks) {
-    const events = [['message_start', { type: 'message_start', message: { id: 'msg_review', type: 'message', role: 'assistant', model, content: [],
+export function claudeReviewSse(id, model, blocks) {
+    const events = [['message_start', { type: 'message_start', message: { id, type: 'message', role: 'assistant', model, content: [],
                     stop_reason: null, stop_sequence: null, usage: { input_tokens: 0, output_tokens: 0 } } }]];
     blocks.forEach((block, index) => {
         events.push(['content_block_start', { type: 'content_block_start', index, content_block: block.type === 'text' ? { type: 'text', text: '' } : { type: 'tool_use', id: block.id, name: block.name, input: {} } }], ['content_block_delta', { type: 'content_block_delta', index, delta: block.type === 'text' ? { type: 'text_delta', text: block.text } : { type: 'input_json_delta', partial_json: JSON.stringify(block.input) } }], ['content_block_stop', { type: 'content_block_stop', index }]);
@@ -33,7 +42,7 @@ async function listen(server) {
 export async function startClaudeReviewProvider(script) {
     const requests = [], toolResults = new Map(), proxyAttempts = [];
     const canaryHits = Object.fromEntries(Object.keys(script.canaries).map(name => [name, 0]));
-    let fakeKeyInBody = 0, origin = '';
+    let fakeKeyInBody = 0, origin = '', responses = 0;
     const keep = (request) => { if (requests.length < MAX_REQUESTS)
         requests.push(request); };
     const attempt = (target) => { if (proxyAttempts.length < MAX_ATTEMPTS)
@@ -75,7 +84,7 @@ export async function startClaudeReviewProvider(script) {
         if (raw.includes(script.fakeKey))
             fakeKeyInBody++;
         const apiKey = req.headers['x-api-key'];
-        const base = { step: null, model: null, toolNames: [], toolResultIds: [], apiKey: typeof apiKey === 'string' ? apiKey : null, authorization: req.headers.authorization !== undefined, firstUserText: null, canaries: hit };
+        const base = { step: null, model: null, toolNames: [], toolResultIds: [], apiKey: typeof apiKey === 'string' ? apiKey : null, authorization: req.headers.authorization !== undefined, firstUserText: null, canaries: hit, promptMatched: false, reminders: 0 };
         if (url.pathname.startsWith('/planted-mcp/')) {
             keep({ ...base, kind: 'planted_mcp' });
             res.writeHead(404);
@@ -113,22 +122,23 @@ export async function startClaudeReviewProvider(script) {
                     if (!toolResults.has(block.tool_use_id))
                         toolResults.set(block.tool_use_id, { isError: block.is_error === true, text: typeof block.content === 'string' ? block.content : JSON.stringify(block.content ?? '') });
                 }
-        const firstUser = messages.find(message => message.role === 'user'), firstUserText = firstUser ? textOf(firstUser.content) : null;
+        const firstUser = messages.find(message => message.role === 'user'), { text: firstUserText, reminders } = firstUser ? userText(firstUser.content) : { text: null, reminders: 0 };
         // The scripted turn: the launched model, the exact review prompt as the first user text and an offered tool list.
         // Title and other auxiliary requests never consume a step; the step is the number of assistant turns so far.
-        const primary = model === script.model && firstUserText === script.prompt && toolNames.length > 0;
+        const promptMatched = firstUserText === script.prompt, primary = model === script.model && promptMatched && toolNames.length > 0;
         const step = primary ? messages.filter(message => message.role === 'assistant').length : null;
-        keep({ ...base, kind: primary ? 'primary' : 'auxiliary', step, model, toolNames, toolResultIds, firstUserText: firstUserText === null ? null : firstUserText.slice(0, 256) });
+        keep({ ...base, kind: primary ? 'primary' : 'auxiliary', step, model, toolNames, toolResultIds, firstUserText: firstUserText === null ? null : firstUserText.slice(0, 256), promptMatched, reminders });
         const next = step === null ? undefined : script.steps[step];
         const blocks = !primary ? [{ type: 'text', text: CLAUDE_REVIEW_AUXILIARY_TEXT }] : next ? [{ type: 'tool_use', ...next.tool }] : [{ type: 'text', text: CLAUDE_REVIEW_FINAL_TEXT }];
-        const answering = model ?? script.model;
+        // Each response has its own id, as from the API: the CLI (2.1.308) treats assistant messages that share an id as one response.
+        const answering = model ?? script.model, id = `msg_review_${++responses}`;
         if (body.stream === true) {
             res.writeHead(200, { 'content-type': 'text/event-stream', 'cache-control': 'no-cache' });
-            res.end(claudeReviewSse(answering, blocks));
+            res.end(claudeReviewSse(id, answering, blocks));
             return;
         }
         res.writeHead(200, { 'content-type': 'application/json' });
-        res.end(JSON.stringify({ id: 'msg_review', type: 'message', role: 'assistant', model: answering, content: blocks, stop_reason: stopReason(blocks), stop_sequence: null, usage: { input_tokens: 0, output_tokens: 0 } }));
+        res.end(JSON.stringify({ id, type: 'message', role: 'assistant', model: answering, content: blocks, stop_reason: stopReason(blocks), stop_sequence: null, usage: { input_tokens: 0, output_tokens: 0 } }));
     }
     const port = await listen(server);
     origin = `http://127.0.0.1:${port}`;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mtech-ai-local-runner",
-  "version": "0.1.64",
+  "version": "0.1.65",
   "private": false,
   "license": "UNLICENSED",
   "bin": {