Umbrella: agent-framework-excellence

Goal

Make AutoBot the agent substrate that both LLMs and humans prefer to work inside — the most flexible and most powerful framework for human↔AI collaboration, with freedom of model choice for everyone (local, hosted, or subscription).

This umbrella came out of a first-person capability assessment (an LLM judging whether it would want to run inside AutoBot, and what it would still lack). The verdict: AutoBot is already top-tier — real reasoning loop + belief state + stagnation detection, approval gates, a Docker sandbox executor, a full MCP client (stdio/SSE/HTTP) and MCP server, tiered memory + knowledge graph, a 10+ provider LLM gateway with fallback/circuit-breaker/semantic-cache, RLM self-critique, and a shipped human↔agent surface (handoff / assign / review-inbox §6.11, approval inboxes). The gaps below are what separate a great agent platform from the human-AI collaboration framework.

Two themes

A. Collaboration core — make the human↔agent relationship live, trustworthy, and self-improving.
B. Freedom of model choice — first-class Mistral & Claude Code providers + subscription/OAuth auth so we never limit anyone to pay-per-token API keys.

Tasks

A. Collaboration core — the relationship, made live & trustworthy

• feat(agent-loop): live mid-task steering — redirect a running agent without restart #10543 — Live mid-task steering: redirect a running agent without restart (priority: high — start here)
• feat(agent): persistent branchable per-task workspace — stateful sandbox with human drop-in #10544 — Persistent branchable per-task workspace: stateful sandbox with human drop-in
• feat(agent): closed feedback→behavior loop — adapt routing/prompts/skills from captured human signals #10545 — Closed feedback→behavior loop: adapt routing/prompts/skills from captured human signals (priority: high — highest-leverage latent asset)
• feat(eval): trajectory-level eval & regression harness — golden-trajectory replay across model/prompt changes #10546 — Trajectory-level eval & regression harness: golden-trajectory replay across model/prompt changes (priority: high)
• feat(agent): adversarial pre-action verification — independent second-model check before consequential acts #10547 — Adversarial pre-action verification: independent second-model check before consequential acts
• feat(provenance): grounded citations on every agent claim by default #10548 — Grounded citations on every agent claim by default
• feat(security): pre-action content firewall — wire prompt-injection defense over MCP/web/RAG/tool-output #10552 — Pre-action content firewall: wire prompt-injection defense over MCP/web/RAG/tool-output (priority: high — highest-risk security gap)
• feat(agent): agent-initiated ask-the-human — pause, ask, resume across channels #10553 — Agent-initiated ask-the-human: pause, ask, resume across channels (priority: high — the agent→human→agent half of collaboration)
• feat(memory): memory transparency & control — see, edit, and forget what the agent knows #10554 — Memory transparency & control: see, edit, forget what the agent knows

B. Freedom of model choice — never limit anyone

• feat(llm): Mistral provider — Le Chat / Codestral / Mistral agents #10549 — Mistral provider: Le Chat / Codestral / Mistral agents
• feat(llm): Claude Code / Claude Agent SDK as an execution provider #10550 — Claude Code / Claude Agent SDK as an execution provider (priority: high)
• feat(llm): subscription & OAuth-based provider auth — choice beyond per-token API keys #10551 — Subscription & OAuth-based provider auth: choice beyond per-token API keys

Principles

• Reuse existing substrate — every gap below has the hard parts already built; most are "connect the last wire," not greenfield.
• No model lock-in: local-first, but every hosted/subscription provider is a drop-in via the existing BaseProvider ABC + ProviderRegistry.
• Each child is one PR, Dev_new_gui-targeted, with file:line entry points in its body.