0.2.26040202 Alpha: login retry + Perplexity fix + platform-level fallback

shafqatevo · claude · shafqatevo · commit 6902e7e0cb6c · 2026-04-02T05:50:53.000+11:00
- Login retry: needs_login platforms are retried after a 90s sign-in window,
  not permanently skipped; other platforms' results already collected during wait
- Perplexity: guard against "Computer" feature in configure_mode (model picker
  and research toggle both explicitly skip any option containing 'computer');
  inject_prompt prefers textarea (new Perplexity UI) over contenteditable
- Platform-level browser-use fallback: full_platform_run() added to
  AgentFallbackManager; kicks in for STATUS_FAILED platforms when API key set;
  uses 25 agent steps (DEEP) / 15 (REGULAR) for complete lifecycle retry

Co-Authored-By: Claude Sonnet 4.6 &lt;noreply@anthropic.com&gt;
diff --git a/.claude-plugin/marketplace.json b/.claude-plugin/marketplace.json
@@ -14,7 +14,7 @@
         "url": "https://github.com/alo-exp/multai.git"
       },
       "description": "MultAI — Submit research prompts to 7 AI platforms simultaneously (Claude.ai, ChatGPT, Copilot, Perplexity, Grok, DeepSeek, Gemini), synthesize results, maintain capability comparison matrices, and produce landscape reports — all from Claude Code.",
-      "version": "0.2.26040201",
+      "version": "0.2.26040202",
       "author": {
         "name": "Ālo Labs",
         "url": "https://alolabs.dev"
diff --git a/.claude-plugin/plugin.json b/.claude-plugin/plugin.json
@@ -1,6 +1,6 @@
 {
   "name": "multai",
-  "version": "0.2.26040201",
+  "version": "0.2.26040202",
   "description": "MultAI — Submit research prompts to 7 AI platforms simultaneously (Claude.ai, ChatGPT, Copilot, Perplexity, Grok, DeepSeek, Gemini), synthesize results, maintain capability comparison matrices, and produce landscape reports — all from Claude Code.",
   "author": {
     "name": "Ālo Labs",
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -6,6 +6,48 @@ Versioning scheme: `Major.Minor.YYMMDDX Phase` — see [CI/CD Strategy](docs/CIC
 
 ---
 
+## 0.2.26040202 Alpha — Orchestrator: Login Retry, Perplexity Fix, Platform-Level Fallback
+
+**Date:** 2026-04-02
+
+### Fix: Login-needed platforms are now retried, not skipped
+
+Previously, if a platform returned `needs_login` (sign-in page detected), it was
+permanently skipped for that run. Now:
+- After all 7 platforms complete in parallel, the engine prints a clear sign-in prompt
+  for each `needs_login` platform (with URL) and waits 90 seconds
+- The user signs in to those platforms in Chrome during the countdown
+- The platforms are retried automatically after the countdown
+- All other platforms' results are already collected — only the login-needed ones wait
+
+### Fix: Perplexity — "Computer" feature no longer triggered
+
+`configure_mode` was inadvertently activating "Perplexity Computer" (a paid, credit-based
+computer-use feature) instead of a standard Sonar model or Research mode. Fixed:
+- All model picker options containing "computer" (case-insensitive) are explicitly skipped
+- Research toggle selection has the same guard
+- Model selection falls back gracefully if no safe Sonar option is found (uses page default)
+- `inject_prompt` updated to prefer textarea (new Perplexity UI) over contenteditable
+
+### Feature: Platform-level browser-use fallback
+
+When a platform returns `STATUS_FAILED` (all Playwright steps failed), and
+`ANTHROPIC_API_KEY` or `GOOGLE_API_KEY` is set, a full browser-use agent session
+now retries the entire platform interaction (navigate → type → send → wait → extract).
+Uses up to 25 agent steps in DEEP mode, 15 in REGULAR mode. Results are saved in
+the same format as normal platform output.
+
+This is additive — existing per-step fallbacks are unchanged.
+
+### Files changed
+
+- `skills/orchestrator/engine/orchestrator.py` — login retry loop + platform-level fallback call
+- `skills/orchestrator/engine/agent_fallback.py` — `full_platform_run()` method added
+- `skills/orchestrator/engine/platforms/perplexity.py` — `configure_mode` Computer guard, `inject_prompt` textarea-first
+- `skills/orchestrator/SKILL.md` — Phase 1 browser-use docs, Phase 3 login-retry docs
+
+---
+
 ## 0.2.26040201 Alpha — `/consolidator` Redesigned as Standalone Skill
 
 **Date:** 2026-04-02
diff --git a/CONTRIBUTOR-GUIDE.md b/CONTRIBUTOR-GUIDE.md
@@ -1,6 +1,6 @@
 # MultAI — Contributor Guide
 
-**Version:** 0.2.26040201 Alpha | **Date:** 2026-03-18
+**Version:** 0.2.26040202 Alpha | **Date:** 2026-03-18
 
 > **Looking for the end-user guide?** See [USER-GUIDE.md](USER-GUIDE.md).
 
diff --git a/USER-GUIDE.md b/USER-GUIDE.md
@@ -1,6 +1,6 @@
 # MultAI — User Guide
 
-**Version:** 0.2.26040201 Alpha | **Date:** 2026-03-18
+**Version:** 0.2.26040202 Alpha | **Date:** 2026-03-18
 
 ---
 
diff --git a/docs/Architecture-and-Design.md b/docs/Architecture-and-Design.md
@@ -1,7 +1,7 @@
 # Architecture and Design Document
 
 **Project:** MultAI
-**Version:** 0.2.26040201 Alpha
+**Version:** 0.2.26040202 Alpha
 **Date:** 2026-03-18
 
 | Version | Date | Summary |
diff --git a/docs/CICD-Strategy-and-Plan.md b/docs/CICD-Strategy-and-Plan.md
@@ -1,7 +1,7 @@
 # CI/CD Strategy and Plan
 
 **Project:** MultAI
-**Version:** 0.2.26040201 Alpha
+**Version:** 0.2.26040202 Alpha
 **Date:** 2026-03-18
 
 | Version | Date | Summary |
diff --git a/docs/SRS.md b/docs/SRS.md
@@ -1,7 +1,7 @@
 # Software Requirements Specification (SRS)
 
 **Project:** MultAI
-**Version:** 0.2.26040201 Alpha
+**Version:** 0.2.26040202 Alpha
 **Date:** 2026-03-18
 **Base Version:** multai/ (formerly multi-ai-skills/, solution-research-skill-260308A)
 
diff --git a/docs/Test-Strategy-and-Plan.md b/docs/Test-Strategy-and-Plan.md
@@ -1,7 +1,7 @@
 # Test Strategy and Plan
 
 **Project:** MultAI
-**Version:** 0.2.26040201 Alpha
+**Version:** 0.2.26040202 Alpha
 **Date:** 2026-03-18
 
 | Version | Date | Summary |
diff --git a/docs/index.html b/docs/index.html
@@ -871,7 +871,7 @@ <h3>CI/CD Strategy</h3>
       <a href="https://github.com/alo-exp/multai/blob/main/CHANGELOG.md" target="_blank" class="feature-card fade-in" style="text-decoration:none">
         <div style="font-size:1.5rem;margin-bottom:8px">&#128221;</div>
         <h3>Changelog</h3>
-        <p>Version history from v2.0 through v0.2.26040201 Alpha with migration notes.</p>
+        <p>Version history from v2.0 through v0.2.26040202 Alpha with migration notes.</p>
       </a>
     </div>
   </div>
diff --git a/pyproject.toml b/pyproject.toml
@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "multai"
-version = "0.2.26040201"
+version = "0.2.26040202"
 description = "MultAI — submit prompts to 7 AI platforms simultaneously, manage rate limits, operate XLSX comparison matrices, and synthesize results. Packaged as a Claude Cowork/Code plugin."
 readme = "USER-GUIDE.md"
 license = { text = "MIT" }
diff --git a/skills/orchestrator/SKILL.md b/skills/orchestrator/SKILL.md
@@ -170,6 +170,14 @@ ls skills/orchestrator/engine/.venv/bin/python 2>/dev/null && echo "venv OK" ||
 
 Do not proceed to Phase 2 until the venv exists.
 
+**Optional — enable browser-use fallback:**
+
+If `ANTHROPIC_API_KEY` (or `GOOGLE_API_KEY`) is set in the environment or `.env` file, two additional safety nets activate automatically:
+- **Per-step fallback:** when a Playwright selector fails at any individual step (configure, inject, send, extract), a browser-use agent takes over that step
+- **Platform-level fallback:** if a platform returns STATUS_FAILED after all Playwright steps failed, a full browser-use agent session retries the entire interaction from navigation to extraction
+
+Neither fallback is required — the engine runs without an API key.
+
 ---
 
 ## Phase 2 — Run the Engine (Code Tab)
@@ -362,7 +370,11 @@ After the engine completes, outputs are in `reports/<task-name>/`:
 | `{Platform}-raw-response.md` | Individual raw response per platform |
 | `{task-name} - Raw AI Responses.md` | **Auto-generated archive** — all responses collated |
 
-Platform statuses: `complete` / `partial` / `failed` / `timeout` / `rate_limited`
+Platform statuses: `complete` / `partial` / `failed` / `timeout` / `rate_limited` / `needs_login`
+
+**Login-needed platforms (`needs_login`):** the engine automatically waits 90 seconds after the parallel run completes, printing a sign-in prompt for each affected platform. After the countdown, those platforms are retried. Inform the user of this if any platform shows `needs_login` in the output — they should watch the terminal and sign in during the countdown.
+
+**Genuinely failed platforms (`failed`):** if `ANTHROPIC_API_KEY` is set, the engine will attempt a full browser-use agent run for each failed platform after the main run.
 
 ---
 
diff --git a/skills/orchestrator/engine/agent_fallback.py b/skills/orchestrator/engine/agent_fallback.py
@@ -22,6 +22,8 @@
 from pathlib import Path
 from typing import Optional
 
+from config import STATUS_COMPLETE, STATUS_NEEDS_LOGIN  # noqa: E402 (config has no external deps)
+
 log = logging.getLogger(__name__)
 
 
@@ -202,6 +204,126 @@ async def _run_agent(
             # Re-raise the ORIGINAL error (not the agent error)
             raise original_error from agent_exc
 
+    async def full_platform_run(
+        self,
+        platform_name: str,
+        platform_url: str,
+        display_name: str,
+        prompt: str,
+        mode: str,
+        output_dir: str,
+    ) -> dict | None:
+        """Full browser-use agent run for a platform where all Playwright steps failed.
+
+        Returns a result dict (same shape as run_single_platform) on success, or None.
+        Uses higher max_steps than per-step fallbacks since it owns the full lifecycle.
+        """
+        if not self._enabled or not platform_url:
+            return None
+
+        # Truncate prompt to keep agent task description manageable.
+        # The agent needs to type the full prompt, so we pass it inline; we truncate
+        # only the task description string, not what the agent actually types.
+        prompt_for_task = prompt[:3000] + ("\n...[prompt continues — type all of it]" if len(prompt) > 3000 else "")
+
+        task = (
+            f"Automate a browser to get an AI response from {display_name}. "
+            f"Step 1: Go to {platform_url}. "
+            f"Step 2: If you see a sign-in or login page, stop immediately and return the text 'NEEDS_LOGIN'. "
+            f"Step 3: Find the main text input (textarea or contenteditable area for typing messages). "
+            f"Step 4: Type the following prompt exactly into that input:\n\n"
+            f"{prompt_for_task}\n\n"
+            f"Step 5: Click the Send or Submit button. "
+            f"Step 6: Wait for the AI to finish generating (loading indicator gone, "
+            f"no stop/cancel button visible). This may take several minutes. "
+            f"Step 7: Extract and return the COMPLETE text of the AI's response."
+        )
+
+        max_steps = 25 if mode == "DEEP" else 15
+        t0 = time.monotonic()
+
+        async with self._lock:
+            try:
+                from browser_use import Agent, BrowserSession
+                from config import AGENT_MODEL_ANTHROPIC, AGENT_MODEL_GOOGLE
+
+                session = BrowserSession(cdp_url=self._cdp_url)
+
+                if self._llm_provider == "anthropic":
+                    from browser_use.llm.anthropic.chat import ChatAnthropic
+                    llm = ChatAnthropic(
+                        model=AGENT_MODEL_ANTHROPIC, timeout=120, max_tokens=8192,
+                    )
+                else:
+                    from browser_use.llm.google.chat import ChatGoogle
+                    llm = ChatGoogle(
+                        model=AGENT_MODEL_GOOGLE, api_key=os.environ.get("GOOGLE_API_KEY"),
+                    )
+
+                agent = Agent(
+                    task=task, llm=llm, browser_session=session, max_steps=max_steps,
+                )
+                history = await agent.run()
+                raw = history.final_result() if history else None
+                result_text = str(raw).strip() if raw is not None else ""
+                duration = round(time.monotonic() - t0, 1)
+
+                event = FallbackEvent(
+                    timestamp=datetime.now().isoformat(),
+                    platform=platform_name,
+                    step="full_platform_run",
+                    original_error="Playwright failed on all steps",
+                    agent_task=task[:500],
+                    agent_result=result_text[:500],
+                    agent_success=bool(result_text and len(result_text) > 200),
+                    duration_s=duration,
+                )
+                self._events.append(event)
+                self._save_log()
+
+                if result_text and "NEEDS_LOGIN" in result_text:
+                    log.warning(f"[{display_name}] Full agent fallback: NEEDS_LOGIN")
+                    return {
+                        "platform": platform_name,
+                        "display_name": display_name,
+                        "status": STATUS_NEEDS_LOGIN,
+                        "chars": 0,
+                        "file": "",
+                        "mode_used": f"{mode}-agent",
+                        "error": "Sign-in required (detected by agent fallback)",
+                        "duration_s": duration,
+                    }
+
+                if result_text and len(result_text) > 200:
+                    filename = f"{display_name.replace(' ', '-')}-raw-response.md"
+                    filepath = Path(output_dir) / filename
+                    filepath.parent.mkdir(parents=True, exist_ok=True)
+                    filepath.write_text(result_text, encoding="utf-8")
+                    log.info(
+                        f"[{display_name}] Full agent fallback SUCCEEDED: "
+                        f"{len(result_text)} chars in {duration}s"
+                    )
+                    return {
+                        "platform": platform_name,
+                        "display_name": display_name,
+                        "status": STATUS_COMPLETE,
+                        "chars": len(result_text),
+                        "file": str(filepath),
+                        "mode_used": f"{mode}-agent-fallback",
+                        "error": "",
+                        "duration_s": duration,
+                    }
+
+                log.warning(
+                    f"[{display_name}] Full agent fallback returned insufficient content "
+                    f"({len(result_text)} chars)"
+                )
+
+            except Exception as exc:
+                log.error(f"[{display_name}] Full agent fallback error: {exc}")
+
+        return None
+
     def _save_log(self) -> None:
         """Persist all fallback events to agent-fallback-log.json."""
         log_path = Path(self._output_dir) / "agent-fallback-log.json"
diff --git a/skills/orchestrator/engine/orchestrator.py b/skills/orchestrator/engine/orchestrator.py
@@ -160,10 +160,12 @@ def _ensure_dependencies() -> None:
     GLOBAL_TIMEOUT_REGULAR,
     PLATFORM_DISPLAY_NAMES,
     PLATFORM_URL_DOMAINS,
+    PLATFORM_URLS,
     REGULAR_MODE,
     STAGGER_DELAY,
     STATUS_FAILED,
     STATUS_ICONS,
+    STATUS_NEEDS_LOGIN,
     STATUS_RATE_LIMITED,
     STATUS_TIMEOUT,
     TIMEOUTS,
@@ -794,6 +796,71 @@ async def orchestrate(args: argparse.Namespace, effective_output_dir: str) -> li
             else:
                 final_results.append(r)
 
+        # ---------------------------------------------------------------
+        # Login retry — platforms that needed sign-in get a second chance.
+        # The other platforms have already completed; the user is notified
+        # and given 90s to sign in before the retry runs.
+        # ---------------------------------------------------------------
+        login_pending = [
+            (i, r) for i, r in enumerate(final_results)
+            if r.get("status") == STATUS_NEEDS_LOGIN
+        ]
+        if login_pending:
+            print("\n" + "=" * 72)
+            print("  🔑  SIGN-IN REQUIRED — the following platforms need you to log in:")
+            for _, r in login_pending:
+                url = PLATFORM_URLS.get(r["platform"], "")
+                print(f"     {r['display_name']:20s}  {url}")
+            print()
+            print("  All other platforms have already been collected.")
+            print("  Sign in to the platforms above in Chrome, then wait.")
+            print("  Retrying automatically in 90 seconds...")
+            print("=" * 72)
+            for remaining in range(90, 0, -10):
+                log.info(f"  Sign-in retry countdown: {remaining}s remaining...")
+                await asyncio.sleep(10)
+            log.info("Retrying sign-in-needed platforms...")
+            for idx, r in login_pending:
+                name = r["platform"]
+                log.info(f"[{PLATFORM_DISPLAY_NAMES.get(name, name)}] Retrying after sign-in wait...")
+                retry = await run_single_platform(
+                    name, context, full_prompt, condensed_prompt,
+                    prompt_sigs, args.mode, str(output_dir), agent_mgr,
+                    existing_page=None, followup=False,
+                )
+                limiter.record_usage(name, args.mode, retry["status"], retry.get("duration_s", 0))
+                final_results[idx] = retry
+
+        # ---------------------------------------------------------------
+        # Platform-level browser-use fallback — when all Playwright steps
+        # failed for a platform, attempt a complete agent-driven run.
+        # Only fires if ANTHROPIC_API_KEY (or GOOGLE_API_KEY) is set.
+        # ---------------------------------------------------------------
+        if agent_mgr.enabled:
+            failed_idxs = [
+                (i, r) for i, r in enumerate(final_results)
+                if r.get("status") == STATUS_FAILED
+            ]
+            if failed_idxs:
+                log.info(
+                    f"Attempting full browser-use fallback for "
+                    f"{len(failed_idxs)} failed platform(s)..."
+                )
+            for idx, r in failed_idxs:
+                url = PLATFORM_URLS.get(r["platform"], "")
+                if not url:
+                    continue
+                fallback = await agent_mgr.full_platform_run(
+                    platform_name=r["platform"],
+                    platform_url=url,
+                    display_name=r["display_name"],
+                    prompt=full_prompt,
+                    mode=args.mode,
+                    output_dir=str(output_dir),
+                )
+                if fallback:
+                    final_results[idx] = fallback
+
         # Disconnect Playwright from Chrome (does NOT close Chrome)
         try:
             async with asyncio.timeout(10):
diff --git a/skills/orchestrator/engine/platforms/perplexity.py b/skills/orchestrator/engine/platforms/perplexity.py

Original file line number	Diff line number	Diff line change
`@@ -1,6 +1,6 @@`
`1`	`1`	`{`
`2`	`2`	`"name": "multai",`
`3`		`- "version": "0.2.26040201",`
	`3`	`+ "version": "0.2.26040202",`
`4`	`4`	`"description": "MultAI — Submit research prompts to 7 AI platforms simultaneously (Claude.ai, ChatGPT, Copilot, Perplexity, Grok, DeepSeek, Gemini), synthesize results, maintain capability comparison matrices, and produce landscape reports — all from Claude Code.",`
`5`	`5`	`"author": {`
`6`	`6`	`"name": "Ālo Labs",`