Harden agent loop, device control, and integrations

[neooriginal]

Summary

• Make agent runs durable and controllable across startup, model calls, tool calls, pause/resume, stop, shutdown, and restart recovery.
• Prevent false completion and silent exits by requiring completion evidence, rejecting non-terminal promises such as "I'm working on it", and producing a truthful fallback when a safety ceiling is reached.
• Keep long-running work flexible: the iteration ceiling is a configurable runaway guard, genuine new evidence resets the no-progress counter, and only unchanged/repeated work is cut off.
• Harden browser, Chrome-extension, Android, and desktop control with end-to-end cancellation, bounded output, protocol cleanup, crash/disconnect recovery, safer screenshots, URL/network validation, and clearer errors.
• Improve provider and official-integration reliability with bounded model calls, transient retry/fallback behavior, dynamic tool access metadata, correct read-only parallelism, OAuth race protection, and retry-safe HTTP handling.
• Make messaging, memory ingestion, runtime, voice, social, and wearable paths cancellation-aware and resilient to partial failures.
• Update vulnerable/outdated dependencies and raise the supported Node runtime to 20.19+.

Why

Several reliability failures shared the same underlying patterns:

• cancellation was not consistently propagated to the real subprocess, socket, request, browser page, or device command;
• terminal state could race with pause/stop/shutdown, and startup failures could disappear before being recorded;
• successful MCP and official-integration reads were invisible to the loop's progress guard;
• a wrapper argument mismatch effectively disabled parallel batching for safe read-only tools;
• provider/network responses and retries were not uniformly bounded;
• the memory embedding-index migration recreated an index on every startup and could race in multi-process startup;
• messaging and recovery paths could lose work or leave orphaned execution.

This patch centralizes those lifecycle guarantees and adds regression coverage around the failure boundaries.

Agent-loop behavior audited

• Default execution allows up to 250 iterations (bounded at 400) as a last-resort runaway guard, not a normal stopping condition.
• Loop budgets are persistently configurable within safe bounds and preserve context-specific defaults when unset.
• New tool evidence counts as progress; repeated unchanged calls are rejected without blocking legitimate research.
• Completion remains model-assisted, but a deterministic evidence/goal contract prevents status promises from being accepted as completed work.
• Pause/resume preserves the same run, stop/abort interrupts active work, shutdown drains owned background work, and terminal transitions are first-writer-wins.
• Model timeouts, transient retries, fallback selection, blank responses, compaction, tool failures, and restart recovery all have explicit terminal or recovery behavior.

User impact

• Fewer unexplained stalls, false "done" messages, and abandoned runs.
• Prompt stop/pause behavior even during external calls or device operations.
• More reliable long-horizon research and integration workflows without hard-coded premature termination.
• Better browser/device error feedback and recovery after disconnects or crashes.
• Durable inbound messaging and safer retries that do not replay writes.

Validation

• npm run test:backend — 536 passed: 496 unit, 19 integration, 11 security, 4 contract, 2 E2E, and 4 WebSocket tests.
• npm run flutter:test — 13 passed.
• flutter analyze — no issues.
• All changed JavaScript/MJS files passed syntax checks.
• git diff --check passed.
• npm audit --audit-level=low and npm run audit:ci — 0 vulnerabilities.
• npm ls --all passed.
• npm run docs:build passed.
• Playwright Chromium launch/navigation smoke test passed with the installed browser.

Platform-validation notes

• A Flutter web build was intentionally not run, per repository instructions.
• Windows companion C++ was reviewed and covered through protocol/unit tests but was not compiled on macOS.
• Physical Android and Windows hardware validation remains appropriate during beta rollout.