Garry's Opinionated OpenClaw/Hermes Agent Brain
TypeScript
fix(claude-cli): a tool loop turn reads the previous turn's prompt cache
garrytan54b9d2ada609
diff --git a/src/core/ai/providers/claude-cli-language-model.ts b/src/core/ai/providers/claude-cli-language-model.tsindex 2bf3f97587..8037f23428 100644--- a/src/core/ai/providers/claude-cli-language-model.ts+++ b/src/core/ai/providers/claude-cli-language-model.ts@@ -107,7 +107,7 @@ export function resolveHermeticConfigDir(return v;}-/** Parsed shape of `claude --print --output-format json`. */+/** Parsed shape of the result event `claude --print` emits. */interface ClaudeJsonResult {type: 'result';subtype: 'success' | string;@@ -153,18 +153,25 @@ type EnvelopeParse =| { ok: false; reason: 'no-result-event' };/**- * Unwrap `--output-format json` stdout into the result envelope and the event- * array it came in. `runClaude` passes `--verbose` (#6791), so the CLI emits- * [{type:"system",subtype:"init",...}, ..., {type:"result",...}]; the bare- * result object (events null) is still accepted. The array shape was verified on- * CLI 2.1.145; that it carries every assistant message, in the 2.1.296 bundle.+ * Unwrap the CLI's stdout into the result envelope and the event array it came+ * in. `runClaude` passes `--output-format stream-json --verbose` (#7352,+ * #6791), so the CLI prints one event per line: {type:"system",...}, the+ * assistant messages, then {type:"result",...}; verified on CLI 2.0.60 and+ * 2.1.319. A single JSON document (the `--output-format json` event array, or+ * a bare result object with events null) is still accepted.*/function parseResultEnvelope(stdout: string): EnvelopeParse {let parsed: unknown;try {parsed = JSON.parse(stdout);} catch (error) {- return { ok: false, reason: 'not-json', error };+ try {+ const lines = stdout.split('\n').filter(line => line.trim().length > 0);+ if (lines.length < 2) return { ok: false, reason: 'not-json', error };+ parsed = lines.map(line => JSON.parse(line));+ } catch {+ return { ok: false, reason: 'not-json', error };+ }}const events = Array.isArray(parsed) ? parsed : null;if (Array.isArray(parsed)) {@@ -279,13 +286,15 @@ function buildToolUseInstructions(}/**- * Render the ai-sdk message array into a single text prompt for `claude --print`- * stdin. System messages are extracted up-front and concatenated into the- * `--system-prompt` flag value. Tool calls and tool results are rendered as+ * Render the ai-sdk message array into one text block per conversation message+ * for `claude --print` stdin (#7352: the CLI's prompt cache matches at block+ * boundaries, so a tool loop's next turn reads this turn's cache only when this+ * turn's blocks are an unchanged prefix of it). System messages are extracted+ * up-front and concatenated into the `--system-prompt` flag value. Tool calls and tool results are rendered as* placeholders so the model sees the conversation in a coherent shape even* though the adapter does not natively round-trip tool calls through claude-cli.*/-function renderPrompt(prompt: LanguageModelV2Prompt): { systemText: string; userPrompt: string } {+function renderPrompt(prompt: LanguageModelV2Prompt): { systemText: string; userBlocks: string[] } {const systemParts: string[] = [];const convo: string[] = [];@@ -339,7 +348,7 @@ function renderPrompt(prompt: LanguageModelV2Prompt): { systemText: string; user}}- return { systemText: systemParts.join('\n'), userPrompt: convo.join('\n\n') };+ return { systemText: systemParts.join('\n'), userBlocks: convo };}/** Oldest claude CLI that accepts every flag `runClaude` passes (probed:@@ -376,12 +385,12 @@ export function thinkingDisabledBelowFloor(maxOutputTokens: number | undefined):/*** Spawn `claude --print` with the contamination-suppression flags and return- * the parsed `--output-format json` envelope. Aborts propagate to SIGTERM on+ * the parsed result envelope. Aborts propagate to SIGTERM on* the child.*/function runClaude(systemPrompt: string,- userPrompt: string,+ userBlocks: string[],model: string,signal?: AbortSignal,maxOutputTokens?: number,@@ -398,10 +407,13 @@ function runClaude(: '{"disableAllHooks":true}';const args = ['--print',- // #6791: --verbose turns the json output into the session's event- // array, the only record of the segments before an automatic- // continuation (the envelope's result holds the last one).- '--output-format', 'json', '--verbose',+ // #7352: one user message whose content blocks are the conversation's+ // messages, read as stream-json; the CLI then answers in stream-json.+ // #6791: --verbose keeps the session's events in that output, the only+ // record of the segments before an automatic continuation (the+ // envelope's result holds the last one).+ '--input-format', 'stream-json',+ '--output-format', 'stream-json', '--verbose','--model', model,'--disable-slash-commands',// Agent isolation: this subprocess must behave like a raw LLM, not a@@ -542,7 +554,7 @@ function runClaude(// error itself is safe to swallow.child.stdin.on('error', () => { /* surfaced via child 'error'/'close' */ });try {- child.stdin.write(userPrompt);+ child.stdin.write(JSON.stringify({ type: 'user', message: { role: 'user', content: userBlocks.map(text => ({ type: 'text', text: text.toWellFormed() })) } }) + '\n');child.stdin.end();} catch (e) {if (signal) signal.removeEventListener('abort', onAbort);@@ -728,11 +740,11 @@ export class ClaudeCliLanguageModel implements LanguageModelV2 {providerMetadata?: SharedV2ProviderMetadata;warnings: never[];}> {- const { systemText, userPrompt } = renderPrompt(options.prompt);+ const { systemText, userBlocks } = renderPrompt(options.prompt);const toolInstructions = buildToolUseInstructions(options.tools);const systemPrompt = [systemText, toolInstructions].filter(s => s.length > 0).join('\n');- const result = await runClaude(systemPrompt, userPrompt, this.modelId, options.abortSignal, options.maxOutputTokens);+ const result = await runClaude(systemPrompt, userBlocks, this.modelId, options.abortSignal, options.maxOutputTokens);// The CLI reports Anthropic's separate buckets; LanguageModelV2Usage wants// the TOTAL input with the cache read as a subset (what every SDK provider// reports), so the gateway prices cache tokens once. Cache creation has nodiff --git a/test/claude-cli-prompt-cache-7352.test.ts b/test/claude-cli-prompt-cache-7352.test.tsnew file mode 100644index 0000000000..b25ecafa7b--- /dev/null+++ b/test/claude-cli-prompt-cache-7352.test.ts@@ -0,0 +1,114 @@+/**+ * #7352 — a claude-cli tool loop must let turn N+1 read turn N's prompt+ * cache. The CLI caches at content-block boundaries, so a prompt sent as one+ * growing text block never shares a boundary with the previous turn and every+ * turn re-writes the whole transcript to the cache. The adapter sends one+ * block per conversation message (`--input-format stream-json`), so turn N's+ * blocks are an unchanged prefix of turn N+1's, and reads the CLI's+ * stream-json (NDJSON) output. Hermetic: a stub binary records stdin and argv.+ */+import { describe, test, expect, beforeAll, afterAll } from 'bun:test';+import { writeFileSync, chmodSync, mkdirSync, rmSync, readFileSync } from 'node:fs';+import { join } from 'node:path';+import { tmpdir } from 'node:os';+import type { LanguageModelV2CallOptions } from '@ai-sdk/provider';+import { withEnv } from './helpers/with-env.ts';++const stubDir = join(tmpdir(), `claude-cli-cache-stub-${process.pid}`);+const stubBin = join(stubDir, 'claude');+const stdinLog = join(stubDir, 'stdin.log');+const argvLog = join(stubDir, 'argv.log');+const responsePath = join(stubDir, 'response.txt');++beforeAll(() => {+ mkdirSync(stubDir, { recursive: true });+ writeFileSync(stubBin, ['#!/bin/sh', `cat > "${stdinLog}"`, `printf "%s\\n" "$@" > "${argvLog}"`, `cat "${responsePath}"`].join('\n'));+ chmodSync(stubBin, 0o755);+});++afterAll(() => rmSync(stubDir, { recursive: true, force: true }));++const result = (text: string) => ({+ type: 'result', subtype: 'success', is_error: false, result: text, stop_reason: 'end_turn',+ session_id: 's', num_turns: 1, usage: { input_tokens: 10, output_tokens: 2, cache_read_input_tokens: 900, cache_creation_input_tokens: 20 },+});++/** What the real CLI prints under --output-format stream-json --verbose: one event per line. */+function streamJson(text: string): string {+ return [+ { type: 'system', subtype: 'init', session_id: 's' },+ { type: 'assistant', message: { id: 'msg_1', type: 'message', role: 'assistant', model: 'claude-sonnet-5-5', content: [{ type: 'text', text }], stop_reason: null }, parent_tool_use_id: null, session_id: 's' },+ result(text),+ ].map(e => JSON.stringify(e)).join('\n') + '\n';+}++async function generate(prompt: LanguageModelV2CallOptions['prompt']) {+ const { ClaudeCliLanguageModel } = await import('../src/core/ai/providers/claude-cli-language-model.ts');+ return new ClaudeCliLanguageModel('claude-sonnet-5-5').doGenerate({ prompt } as LanguageModelV2CallOptions);+}++function sentBlocks(): string[] {+ const lines = readFileSync(stdinLog, 'utf8').split('\n').filter(Boolean);+ expect(lines).toHaveLength(1);+ const msg = JSON.parse(lines[0]!);+ expect(msg).toMatchObject({ type: 'user', message: { role: 'user' } });+ return msg.message.content.map((b: { type: string; text: string }) => { expect(b.type).toBe('text'); return b.text; });+}++const turnN: LanguageModelV2CallOptions['prompt'] = [+ { role: 'system', content: 'You are a test agent.' },+ { role: 'user', content: [{ type: 'text', text: 'Find the page.' }] },+ { role: 'assistant', content: [{ type: 'tool-call', toolCallId: 't1', toolName: 'search', input: { query: 'acme-example' } }] },+ { role: 'tool', content: [{ type: 'tool-result', toolCallId: 't1', toolName: 'search', output: { type: 'text', value: 'companies/acme-example' } }] },+] as LanguageModelV2CallOptions['prompt'];+const turnN1 = [...turnN,+ { role: 'assistant', content: [{ type: 'tool-call', toolCallId: 't2', toolName: 'get_page', input: { slug: 'companies/acme-example' } }] },+ { role: 'tool', content: [{ type: 'tool-result', toolCallId: 't2', toolName: 'get_page', output: { type: 'text', value: 'Acme Example page' } }] },+] as LanguageModelV2CallOptions['prompt'];++describe('claude-cli tool loops reuse the previous turn prompt cache (#7352)', () => {+ test('one block per message; turn N is an unchanged prefix of turn N+1', async () => {+ await withEnv({ GBRAIN_CLAUDE_CLI_BIN: stubBin }, async () => {+ writeFileSync(responsePath, streamJson('ok'));+ await generate(turnN);+ const first = sentBlocks();+ await generate(turnN1);+ const second = sentBlocks();+ expect(first.length).toBe(3);+ expect(second.length).toBe(5);+ expect(second.slice(0, first.length)).toEqual(first);+ expect(first[0]).toBe('User: Find the page.');+ const argv = readFileSync(argvLog, 'utf8').split('\n');+ expect(argv[argv.indexOf('--input-format') + 1]).toBe('stream-json');+ expect(argv[argv.indexOf('--output-format') + 1]).toBe('stream-json');+ expect(argv).toContain('--verbose');+ });+ });++ test('the stream-json result event is read: text and cache usage', async () => {+ await withEnv({ GBRAIN_CLAUDE_CLI_BIN: stubBin }, async () => {+ writeFileSync(responsePath, streamJson('the answer'));+ const res = await generate(turnN);+ expect(res.content).toEqual([{ type: 'text', text: 'the answer' }]);+ expect(res.usage.cachedInputTokens).toBe(900);+ expect(res.usage.inputTokens).toBe(930);+ });+ });++ test('a lone surrogate (a page cut mid-emoji) reaches the CLI as U+FFFD, as the raw stdin write did', async () => {+ await withEnv({ GBRAIN_CLAUDE_CLI_BIN: stubBin }, async () => {+ writeFileSync(responsePath, streamJson('ok'));+ await generate([{ role: 'user', content: [{ type: 'text', text: 'ship it \ud83d' }] }] as LanguageModelV2CallOptions['prompt']);+ expect(sentBlocks()).toEqual(['User: ship it \ufffd']);+ expect(readFileSync(stdinLog, 'utf8')).not.toContain('\\ud83d');+ });+ });++ test('an error result in the stream still surfaces as a typed error', async () => {+ await withEnv({ GBRAIN_CLAUDE_CLI_BIN: stubBin }, async () => {+ writeFileSync(responsePath, JSON.stringify({ type: 'system', subtype: 'init' }) + '\n'+ + JSON.stringify({ ...result('API Error: rate limited'), is_error: true, subtype: 'error', api_error_status: 429 }) + '\n');+ await expect(generate(turnN)).rejects.toMatchObject({ name: 'ClaudeCliProcessError', apiErrorStatus: 429 });+ });+ });+});diff --git a/test/claude-cli-recipe.test.ts b/test/claude-cli-recipe.test.tsindex 0c92235616..10a6380ef4 100644--- a/test/claude-cli-recipe.test.ts+++ b/test/claude-cli-recipe.test.ts@@ -766,8 +766,9 @@ describe('claude-cli LanguageModel — context isolation', () => {const cwd = fs.readFileSync(cwdLog, 'utf8').trim();expect(argv).toContain('--print');- expect(argv).toContain('--output-format');- expect(argv).toContain('json');+ expect(argv[argv.indexOf('--input-format') + 1]).toBe('stream-json');+ expect(argv[argv.indexOf('--output-format') + 1]).toBe('stream-json');+ expect(argv).toContain('--verbose');expect(argv).toContain('--disable-slash-commands');// Agent-isolation hardening: no built-in tools, no inherited MCP servers.expect(argv).toContain('--tools');diff --git a/test/claude-cli-stop-reason.test.ts b/test/claude-cli-stop-reason.test.tsindex 0d5520b2a3..b91ec5ec49 100644--- a/test/claude-cli-stop-reason.test.ts+++ b/test/claude-cli-stop-reason.test.ts@@ -72,7 +72,7 @@ describe('claude-cli replays earlier tool calls with their arguments (#6236)', (],} as LanguageModelV2CallOptions);});- const sent = readFileSync(stdinPath, 'utf8');+ const sent = (JSON.parse(readFileSync(stdinPath, 'utf8')).message.content as Array<{ text: string }>).map(b => b.text).join('\n');expect(sent).toContain('[tool_use brain_get_page({"slug":"wiki/personal/patterns/x"})]');expect(sent).not.toContain('[object Object]');});