xonecas.io
64875
Mc

mini-coder / commit 80aaec5

feat: auto-continue when max tool steps reached

When a turn hits the MAX_STEPS limit (50) with the last finish reason
being 'tool-calls', the session runner now automatically starts a new
turn instead of stopping and requiring user input. Context pruning
runs at the start of each new turn, keeping the context manageable.

Changes:
- Track lastFinishReason in TurnStepTracker
- Add hitMaxSteps to TurnCompleteEvent and TurnResult
- Wrap runTurn call in a while loop that continues on hitMaxSteps
- Propagate hitMaxSteps through both CLI and headless reporters

xonecas2026-03-18 80aaec5748d6157c8da12be75f63dce05e6e0058 parent 9d72202

8 files changed

src/agent/reporter.ts+1−0
@@ -7,6 +7,7 @@ export interface TurnResult {
77 contextTokens: number;
88 newMessages: CoreMessage[];
99 reasoningText: string;
10+ hitMaxSteps: boolean;
1011 }
1112
1213 export interface StatusBarData {
src/agent/session-runner.ts+43−33
@@ -218,45 +218,55 @@ export class SessionRunner {
218218 let lastAssistantText = "";
219219
220220 try {
221- this.reporter.startSpinner("thinking");
222- const events = runTurn({
223- model: llm,
224- modelString: this.currentModel,
225- messages: this.coreHistory,
226- tools: this.tools,
227- ...(systemPrompt ? { systemPrompt } : {}),
228- signal: abortController.signal,
229- pruningMode: this.pruningMode,
230- toolResultPayloadCapBytes: this.toolResultPayloadCapBytes,
231- promptCachingEnabled: this.promptCachingEnabled,
232- openaiPromptCacheRetention: this.openaiPromptCacheRetention,
233- googleCachedContent: this.googleCachedContent,
234- ...(this.currentThinkingEffort
235- ? { thinkingEffort: this.currentThinkingEffort }
236- : {}),
237- });
221+ let continueLoop = true;
222+ while (continueLoop) {
223+ this.reporter.startSpinner("thinking");
224+ const events = runTurn({
225+ model: llm,
226+ modelString: this.currentModel,
227+ messages: this.coreHistory,
228+ tools: this.tools,
229+ ...(systemPrompt ? { systemPrompt } : {}),
230+ signal: abortController.signal,
231+ pruningMode: this.pruningMode,
232+ toolResultPayloadCapBytes: this.toolResultPayloadCapBytes,
233+ promptCachingEnabled: this.promptCachingEnabled,
234+ openaiPromptCacheRetention: this.openaiPromptCacheRetention,
235+ googleCachedContent: this.googleCachedContent,
236+ ...(this.currentThinkingEffort
237+ ? { thinkingEffort: this.currentThinkingEffort }
238+ : {}),
239+ });
238240
239- const { inputTokens, outputTokens, contextTokens, newMessages } =
240- await this.reporter.renderTurn(events, {
241+ const {
242+ inputTokens,
243+ outputTokens,
244+ contextTokens,
245+ newMessages,
246+ hitMaxSteps,
247+ } = await this.reporter.renderTurn(events, {
241248 showReasoning: this.showReasoning,
242249 verboseOutput: this.verboseOutput,
243250 });
244- const historyMessages = sanitizeModelAuthoredMessages(
245- newMessages,
246- this.currentModel,
247- );
251+ const historyMessages = sanitizeModelAuthoredMessages(
252+ newMessages,
253+ this.currentModel,
254+ );
248255
249- if (historyMessages.length > 0) {
250- this.coreHistory.push(...historyMessages);
251- this.session.messages.push(...historyMessages);
252- saveMessages(this.session.id, historyMessages, thisTurn);
253- }
256+ if (historyMessages.length > 0) {
257+ this.coreHistory.push(...historyMessages);
258+ this.session.messages.push(...historyMessages);
259+ saveMessages(this.session.id, historyMessages, thisTurn);
260+ }
261+
262+ lastAssistantText = extractAssistantText(historyMessages);
263+ this.totalIn += inputTokens;
264+ this.totalOut += outputTokens;
265+ this.lastContextTokens = contextTokens;
266+ touchActiveSession(this.session);
254267
255- lastAssistantText = extractAssistantText(historyMessages);
256- this.totalIn += inputTokens;
257- this.totalOut += outputTokens;
258- this.lastContextTokens = contextTokens;
259- touchActiveSession(this.session);
268+ continueLoop = hitMaxSteps && !abortController.signal.aborted;
269+ }
260270 } catch (err) {
261271 const stubMsg = makeInterruptMessage("error");
262272 this.coreHistory.push(stubMsg);
src/cli/headless-reporter.ts+3−0
@@ -31,6 +31,7 @@ export class HeadlessReporter implements AgentReporter {
3131 let inputTokens = 0;
3232 let outputTokens = 0;
3333 let contextTokens = 0;
34+ let hitMaxSteps = false;
3435 let newMessages: CoreMessage[] = [];
3536 let accumulatedText = "";
3637 let reasoningText = "";
@@ -49,6 +50,7 @@ export class HeadlessReporter implements AgentReporter {
4950 inputTokens = event.inputTokens;
5051 outputTokens = event.outputTokens;
5152 contextTokens = event.contextTokens;
53+ hitMaxSteps = event.hitMaxSteps;
5254 newMessages = event.messages;
5355 break;
5456 case "turn-error": {
@@ -72,6 +74,7 @@ export class HeadlessReporter implements AgentReporter {
7274 contextTokens,
7375 newMessages,
7476 reasoningText: normalizeReasoningText(reasoningText),
77+ hitMaxSteps,
7578 };
7679 }
7780
src/cli/stream-render.test.ts+1−0
@@ -50,6 +50,7 @@ describe("renderTurn", () => {
5050 contextTokens: 3,
5151 newMessages: [],
5252 reasoningText: "",
53+ hitMaxSteps: false,
5354 });
5455 });
5556
src/cli/stream-render.ts+4−0
@@ -26,6 +26,7 @@ export async function renderTurn(
2626 contextTokens: number;
2727 newMessages: CoreMessage[];
2828 reasoningText: string;
29+ hitMaxSteps: boolean;
2930 }> {
3031 const showReasoning = opts?.showReasoning ?? true;
3132 const verboseOutput = opts?.verboseOutput ?? false;
@@ -35,6 +36,7 @@ export async function renderTurn(
3536 let inputTokens = 0;
3637 let outputTokens = 0;
3738 let contextTokens = 0;
39+ let hitMaxSteps = false;
3840 let newMessages: CoreMessage[] = [];
3941 const startedToolCalls = new Set<string>();
4042 const toolCallInfo = new Map<string, { toolName: string; label: string }>();
@@ -138,6 +140,7 @@ export async function renderTurn(
138140 inputTokens = event.inputTokens;
139141 outputTokens = event.outputTokens;
140142 contextTokens = event.contextTokens;
143+ hitMaxSteps = event.hitMaxSteps;
141144 newMessages = event.messages;
142145 break;
143146 }
@@ -166,5 +169,6 @@ export async function renderTurn(
166169 contextTokens,
167170 newMessages,
168171 reasoningText: getReasoningText(),
172+ hitMaxSteps,
169173 };
170174 }
src/cli/test-helpers.ts+1−0
@@ -128,6 +128,7 @@ export function turnDone(messages: CoreMessage[] = []): TurnEvent {
128128 inputTokens: 1,
129129 outputTokens: 2,
130130 contextTokens: 3,
131+ hitMaxSteps: false,
131132 messages,
132133 };
133134 }
src/llm-api/turn.ts+3−0
@@ -158,11 +158,14 @@ export async function* runTurn(options: {
158158 outputTokens: finalState.outputTokens,
159159 });
160160
161+ const hitMaxSteps = finalState.stepCount >= MAX_STEPS;
162+
161163 yield {
162164 type: "turn-complete",
163165 inputTokens: finalState.inputTokens,
164166 outputTokens: finalState.outputTokens,
165167 contextTokens: finalState.contextTokens,
168+ hitMaxSteps,
166169 messages: finalState.partialMessages,
167170 };
168171 } catch (err) {
src/llm-api/types.ts+2−0
@@ -44,6 +44,8 @@ export interface TurnCompleteEvent {
4444 outputTokens: number;
4545 /** Input tokens of the final step — equals the actual context window usage. */
4646 contextTokens: number;
47+ /** True when the turn ended because it hit the max steps limit. */
48+ hitMaxSteps: boolean;
4749 /**
4850 * Raw AI SDK ModelMessage objects generated this turn.
4951 * These are passed directly back to streamText on subsequent turns —