fix: sanitize tool call IDs in agent loop for Mistral strict9 format (#23595) (#23698)

* fix: sanitize tool call IDs in agent loop for Mistral strict9 format (#23595)

Mistral requires tool call IDs to be exactly 9 alphanumeric characters
([a-zA-Z0-9]{9}). The existing sanitizeToolCallIdsForCloudCodeAssist
mechanism only ran on historical messages at attempt start via
sanitizeSessionHistory, but the pi-agent-core agent loop's internal
tool call → tool result cycles bypassed that path entirely.

Changes:
- Wrap streamFn (like dropThinkingBlocks) so every outbound request
  sees sanitized tool call IDs when the transcript policy requires it
- Replace call_${Date.now()} in pendingToolCalls with a 9-char hex ID
  generated from crypto.randomBytes
- Add Mistral tool call ID error pattern to ERROR_PATTERNS.format so
  the error is correctly classified for retry/rotation

* Changelog: document Mistral strict9 tool-call ID fix

---------

Co-authored-by: echoVic <AkiraVic@outlook.com>
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>

CHANGELOG.md

@@ -121,6 +121,7 @@ Docs: https://docs.openclaw.ai
 - Agents/Fallbacks: treat JSON payloads with `type: "api_error"` + `"Internal server error"` as transient failover errors so Anthropic 500-style failures trigger model fallback. (#23193) Thanks @jarvis-lane.
 - Agents/Google: sanitize non-base64 `thought_signature`/`thoughtSignature` values from assistant replay transcripts for native Google Gemini requests while preserving valid signatures and tool-call order. (#23457) Thanks @echoVic.
 - Agents/Transcripts: validate assistant tool-call names (syntax/length + registered tool allowlist) before transcript persistence and during replay sanitization so malformed failover tool names no longer poison sessions with repeated provider HTTP 400 errors. (#23324) Thanks @johnsantry.
+- Agents/Mistral: sanitize tool-call IDs in the embedded agent loop and generate strict provider-safe pending tool-call IDs, preventing Mistral strict9 `HTTP 400` failures on tool continuations. (#23698) Thanks @echoVic.
 - Agents/Compaction: strip stale assistant usage snapshots from pre-compaction turns when replaying history after a compaction summary so context-token estimation no longer reuses pre-compaction totals and immediately re-triggers destructive follow-up compactions. (#19127) Thanks @tedwatson.
 - Agents/Replies: emit a default completion acknowledgement (`✅ Done.`) when runs execute tools successfully but return no final assistant text, preventing silent no-reply turns after tool-only completions. (#22834) Thanks @Oldshue.
 - Agents/Subagents: honor `tools.subagents.tools.alsoAllow` and explicit subagent `allow` entries when resolving built-in subagent deny defaults, so explicitly granted tools (for example `sessions_send`) are no longer blocked unless re-denied in `tools.subagents.tools.deny`. (#23359) Thanks @goren-beehero.

src/agents/pi-embedded-helpers/errors.ts

@@ -614,6 +614,7 @@ const ERROR_PATTERNS = {
     "tool_use_id",
     "messages.1.content.1.tool_use.id",
     "invalid request format",
+    /tool call id was.*must be/i,
   ],
 } as const;
 

src/agents/pi-embedded-runner/run.ts

@@ -1,3 +1,4 @@
+import { randomBytes } from "node:crypto";
 import fs from "node:fs/promises";
 import type { ThinkLevel } from "../../auto-reply/thinking.js";
 import { generateSecureToken } from "../../infra/secure-random.js";
@@ -1122,7 +1123,7 @@ export async function runEmbeddedPiAgent(
               pendingToolCalls: attempt.clientToolCall
                 ? [
                     {
-                      id: `call_${Date.now()}`,
+                      id: randomBytes(5).toString("hex").slice(0, 9),
                       name: attempt.clientToolCall.name,
                       arguments: JSON.stringify(attempt.clientToolCall.params),
                     },

src/agents/pi-embedded-runner/run/attempt.ts

@@ -77,6 +77,7 @@ import {
 } from "../../skills.js";
 import { buildSystemPromptParams } from "../../system-prompt-params.js";
 import { buildSystemPromptReport } from "../../system-prompt-report.js";
+import { sanitizeToolCallIdsForCloudCodeAssist } from "../../tool-call-id.js";
 import { resolveTranscriptPolicy } from "../../transcript-policy.js";
 import { DEFAULT_BOOTSTRAP_FILENAME } from "../../workspace.js";
 import { isRunnerAbortError } from "../abort.js";
@@ -771,6 +772,32 @@ export async function runEmbeddedAttempt(
         };
       }
 
+      // Mistral (and other strict providers) reject tool call IDs that don't match their
+      // format requirements (e.g. [a-zA-Z0-9]{9}). sanitizeSessionHistory only processes
+      // historical messages at attempt start, but the agent loop's internal tool call →
+      // tool result cycles bypass that path. Wrap streamFn so every outbound request
+      // sees sanitized tool call IDs.
+      if (transcriptPolicy.sanitizeToolCallIds && transcriptPolicy.toolCallIdMode) {
+        const inner = activeSession.agent.streamFn;
+        const mode = transcriptPolicy.toolCallIdMode;
+        activeSession.agent.streamFn = (model, context, options) => {
+          const ctx = context as unknown as { messages?: unknown };
+          const messages = ctx?.messages;
+          if (!Array.isArray(messages)) {
+            return inner(model, context, options);
+          }
+          const sanitized = sanitizeToolCallIdsForCloudCodeAssist(messages as AgentMessage[], mode);
+          if (sanitized === messages) {
+            return inner(model, context, options);
+          }
+          const nextContext = {
+            ...(context as unknown as Record<string, unknown>),
+            messages: sanitized,
+          } as unknown;
+          return inner(model, nextContext as typeof context, options);
+        };
+      }
+
       if (anthropicPayloadLogger) {
         activeSession.agent.streamFn = anthropicPayloadLogger.wrapStreamFn(
           activeSession.agent.streamFn,