@tiens.nguyen/gonext-local-worker 1.0.246 → 1.0.247

package/gonext-local-worker.mjs

@@ -1879,6 +1879,9 @@ async function runAgentChatJob(job) {
       // REPL's "I can show a picker" signal (else python keeps hard-blocking, as on web).
       jobId,
       interactiveApproval: payload?.interactiveApproval === true,
+      // Session opt-in (task #80): auto-extend past the step budget without re-prompting
+      // (the REPL sets this once the user picks "yes, don't ask again").
+      alwaysExtendOnMaxStep: payload?.alwaysExtendOnMaxStep === true,
       // Auto-test mode (/test-auto): the agent verifies its work and fixes until it passes.
       autoTest: payload?.autoTest === true,
       // Deploy target chosen via the terminal /server picker (task #69). Host/user only.

package/gonext-repl.mjs

@@ -778,8 +778,25 @@ function onApprovalKey(key) {
 // Show the picker and resolve to the user's choice. Non-TTY (piped) can't pick → deny
 // (safe). Auto-denies after ~170s so an absent user never hangs the turn (and stays
 // under the python side's 180s wait); Enter is instant since Yes is preselected.
-function approvalPrompt(command) {
-  return new Promise((resolve) => {
+// The command carries a "__MAXSTEP__::<message>" sentinel when it's the step-budget
+// continue prompt (task #80) rather than a risky-command approval. That case shows a
+// 3-option picker (Yes / Yes-don't-ask-again / No); a normal command shows Yes/No.
+// Returns { allow, dontAsk } — dontAsk is only ever true for the max-step "don't ask
+// again this session" choice (the caller then stops sending the prompt).
+const MAXSTEP_PREFIX = "__MAXSTEP__::";
+async function approvalPrompt(command) {
+  const isMax = command.startsWith(MAXSTEP_PREFIX);
+  if (isMax) {
+    const label = command.slice(MAXSTEP_PREFIX.length);
+    if (!process.stdin.isTTY) return { allow: false, dontAsk: false };
+    process.stdout.write("\n" + yellow("⏭ " + label) + "\n");
+    const idx = await pickFromList(
+      ["Yes, keep going", "Yes, and don't ask again this session", "No, stop here"],
+      0
+    );
+    return { allow: idx === 0 || idx === 1, dontAsk: idx === 1 };
+  }
+  const allow = await new Promise((resolve) => {
     if (!process.stdin.isTTY) {
       process.stdout.write(dim(`\n(approval needed for: ${command} — no interactive terminal, skipping)\n`));
       resolve(false);
@@ -799,6 +816,7 @@ function approvalPrompt(command) {
       orig(v);
     };
   });
+  return { allow, dontAsk: false };
 }
 // Generic arrow-key list picker (↑/↓ + Enter) — used by /server so you navigate instead
 // of typing a number. Same mechanism as the Yes/No approval picker above, generalized to
@@ -860,6 +878,10 @@ let sessionCodingModel = "";
 // and fixes → re-tests until it passes before finishing. Default off; remembered per
 // folder in the session file; sent to /agent-ask as autoTest so python drives the loop.
 let sessionTestAuto = false;
+// Step-budget opt-out (task #80): set once the user picks "yes, don't ask again" at a
+// max-step prompt. Session-scoped (not persisted per folder); sent to /agent-ask each turn
+// so the agent auto-extends past the step budget without prompting again.
+let alwaysExtendOnMaxStep = false;
 // Deploy target chosen via /server (task #69): {name, host, user} or null. Session-scoped;
 // sent to /agent-ask as deployServer so the agent deploys to this host with KEY auth.
 let selectedServer = null;
@@ -965,6 +987,9 @@ async function runAgentTurn(history) {
       // The terminal can show a Yes/No picker → the agent PAUSES on a risky command and
       // asks instead of hard-blocking it (interactive command-approval gate).
       interactiveApproval: true,
+      // Step-budget opt-out (task #80): once the user chose "don't ask again", auto-extend
+      // past the step budget without prompting.
+      ...(alwaysExtendOnMaxStep ? { alwaysExtendOnMaxStep: true } : {}),
       // Auto-test mode (/test-auto): the agent verifies its work and fixes until it passes.
       ...(sessionTestAuto ? { autoTest: true } : {}),
       // Deploy target chosen via /server — host/user only (no secret), for key-auth deploys.
@@ -1409,7 +1434,13 @@ async function runAgentTurn(history) {
       ) {
         lastApprovalId = job.pendingApproval.id;
         clearStatus();
-        const allow = await approvalPrompt(job.pendingApproval.command || "");
+        const { allow, dontAsk } = await approvalPrompt(job.pendingApproval.command || "");
+        // "Yes, don't ask again this session" (max-step prompt only): remember it so every
+        // later turn sends alwaysExtendOnMaxStep and the agent auto-extends silently (#80).
+        if (dontAsk) {
+          alwaysExtendOnMaxStep = true;
+          process.stdout.write(dim("  (won't ask again about the step budget this session)\n"));
+        }
         try {
           await fetch(`${apiBase}/api/worker/jobs/${jobId}/approval`, {
             method: "POST",

package/gonext_agent_chat.py

@@ -2340,7 +2340,10 @@ def _ws_request_approval(api_base, worker_key, job_id, command, reason):
     except Exception as e:  # noqa: BLE001
         _log(f"approval-request failed: {e} — treating as deny")
         return False
-    _emit({"type": "step", "text": f"Awaiting your approval to run: {command[:70]}"})
+    if command.startswith("__MAXSTEP__::"):
+        _emit({"type": "step", "text": "Awaiting your choice: continue past the step budget?"})
+    else:
+        _emit({"type": "step", "text": f"Awaiting your approval to run: {command[:70]}"})
     deadline = _t.time() + 180  # 3 min; absent user → deny (never auto-run something risky)
     while _t.time() < deadline:
         _t.sleep(1.2)
@@ -5396,8 +5399,38 @@ def run_agent_chat(cfg):
                 _log(f"compact system prompt unavailable ({_e}); using smolagents default")
         agent = _ToolAgent(**agent_kwargs)
         _agent_ref["a"] = agent  # let step_callback trim this agent's step memory (#63)
+        # At the step budget the run ends WITHOUT final_answer (_maxsteps_partial hit). Offer
+        # to keep going instead of dead-ending (task #80): a terminal user says Yes → extend
+        # the budget to a large ceiling and CONTINUE with the existing memory (reset=False,
+        # so the model keeps all prior observations). smolagents captures max_steps as a
+        # LOCAL at run() time, so we can't bump it mid-loop — we re-enter run() with a raised
+        # agent.max_steps instead. The real bound is the 30-min worker job cap, so 1000 just
+        # means "run to completion". 'Yes, don't ask again' is handled REPL-side: it flips a
+        # session flag so the NEXT turn arrives with alwaysExtendOnMaxStep and we auto-extend
+        # without prompting.
+        _MAXSTEP_CEILING = 1000
+        _always_extend = bool(cfg.get("alwaysExtendOnMaxStep"))
         with contextlib.redirect_stdout(sys.stderr):
             result = agent.run(task_with_hint)
+            while _maxsteps_partial.get("hit"):
+                if _always_extend:
+                    _log(f"max-steps reached → auto-extending to {_MAXSTEP_CEILING} "
+                         "(don't-ask-again session flag)")
+                    _do_extend = True
+                elif _interactive_approval and _job_id and pdf_api_base and pdf_worker_key:
+                    _do_extend = _ws_request_approval(
+                        pdf_api_base, pdf_worker_key, _job_id,
+                        f"__MAXSTEP__::Reached the step budget ({agent.max_steps}). "
+                        "Keep going?", "max-steps")
+                else:
+                    _do_extend = False  # web / no interactive channel → current wrap-up
+                if not _do_extend:
+                    break
+                _maxsteps_partial["hit"] = False
+                agent.max_steps = _MAXSTEP_CEILING
+                _emit({"type": "step", "text": "Continuing — step budget extended…"})
+                _log(f"continuing run with raised budget {agent.max_steps}")
+                result = agent.run(task_with_hint, reset=False)
         # Final formatting — NO extra summarizer model call. In multi-step mode the
         # agent's own final_answer() (or the max-steps fallback above) already holds the
         # synthesized answer; we just strip the internal hint tags we appended to tool

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.246",
+  "version": "1.0.247",
   "description": "Polls GoNext cloud API for async local LLM jobs and runs them against Ollama/OpenAI-compatible servers on this Mac",
   "type": "module",
   "license": "MIT",