Merge remote-tracking branch 'origin/main' into light-mode

# Conflicts:
#	Cargo.lock
#	Cargo.toml
#	crates/browser-use-tui/Cargo.toml

Author: laithrw

Cargo.lock

@@ -1,8 +1,9 @@
 [workspace]
 members = [
+  "crates/browser-use-agent",
   "crates/browser-use-browser",
   "crates/browser-use-cli",
-  "crates/browser-use-core",
+  "crates/browser-use-llm",
   "crates/browser-use-providers",
   "crates/browser-use-python-worker",
   "crates/browser-use-protocol",
@@ -19,12 +20,16 @@ version = "0.1.2"
 [workspace.dependencies]
 anyhow = "1"
 arboard = "3.6.1"
+async-trait = "0.1"
 base64 = "0.22"
 bm25 = "2.3.2"
+bytes = "1"
 chrono = { version = "0.4", default-features = false, features = ["clock"] }
 clap = { version = "4", features = ["derive"] }
 crossterm = "0.29"
+futures-util = "0.3"
 ignore = "0.4"
+rand = "0.9"
 iana-time-zone = "0.1"
 image = { version = "0.25", default-features = false, features = ["png", "jpeg", "gif", "webp"] }
 ratatui = { version = "0.30", default-features = false, features = ["crossterm_0_29"] }
@@ -39,6 +44,9 @@ sha2 = "0.10"
 similar = "2.7.0"
 tempfile = "3"
 terminal-colorsaurus = "0.4"
+thiserror = "1"
+tokio = { version = "1", features = ["rt-multi-thread", "macros"] }
+tokio-util = "0.7"
 toml = "0.8"
 tree-sitter = "0.25.10"
 tree-sitter-bash = "0.25"

Cargo.toml

@@ -0,0 +1,26 @@
+# DECISIONS.md — settled choices & sanctioned divergences
+
+Decision log for the rearchitecture (see `REARCHITECTURE.md`, `IMPLEMENTATION_PLAN.md`).
+Append new entries as they're made.
+
+## Sanctioned divergences from codex (intentional — preserve, justify, test)
+- **D-DIV-1 Multi-provider** instead of OpenAI-only (the product needs OpenAI/Anthropic/local).
+- **D-DIV-2 SQLite as a write-only sink, not a hot loader.** Dump events/state to SQLite for durability + debuggability + resume, but keep runtime state **in-memory**; **never read/poll SQLite on the hot path**; read it **only on resume**. (Reading it hot is "slow as hell".)
+- **D-DIV-3 Sync/serial `view_image`** — blocking read, never parallel with browser actions, so screenshots are observed in order.
+- **D-DIV-4 Browser + Python tool surface** (our product; browser layer = `browser-harness-js`, treated as a black box).
+- **D-DIV-5 Drop the codex/ChatGPT backend** entirely (server-side, can't use) — its headers/OAuth/identifiers go with it.
+
+Everything else = **EXTREME PARITY** with codex (mechanism, heuristics, thresholds, outputs), proven by tests.
+
+## Settled open questions
+- **D-1 Context messages:** align with codex's typed `reference_context` mechanism for parity; our extra kinds (permissions/personality/goal/collaboration/hook/mention/generated-image/browser) become additional typed items. *(User: "sure whatever" → pick the parity-aligned option.)*
+- **D-2 Subagent wait:** **event-notify** (in-memory mailbox notification), not the 50ms SQLite poll. Messages still dumped to SQLite for the record. *(User: "I prefer event-notify"; consistent with D-DIV-2.)*
+- **D-3 Crate granularity:** engine = one `browser-use-core` crate with submodules; separate crates for `browser-use-llm`, `browser-use-mcp`, `browser-use-sandbox`, `browser-use-guardian`. *(User: "whatever you think is prettier".)*
+- **D-4 Async migration order:** top-down (runtime → loop → tools) behind the frozen Phase-0 interfaces, optimized for **correctness + testability, not speed**. *(User: "whatever will make you implement BETTER … focus is NOT speed … everything gets done, run and tested".)*
+- **D-5 v1 providers:** OpenAI (Responses), Anthropic (Messages + Claude-Code OAuth), Ollama, **DeepSeek, OpenRouter, Fireworks** (via the `openai-chat` protocol where compatible). The protocol × provider design makes additional providers config-only — added freely later.
+
+## Process decisions
+- **D-P-1 No human in the loop.** The orchestrating agent does all coordination, review, gating, and the serial carve.
+- **D-P-2 Tests first, commit first.** Each WP commits parity/e2e tests before implementing; frequent working-increment commits; Phase 0 commits per extraction step.
+- **D-P-3 Run-and-test, not compile-only.** Acceptance = behavior runs + tests pass against a **live model**. E2E uses the **existing codex auth** as the live-model vehicle during the build (revisit once OpenAI/Anthropic direct auth is wired — see `IMPLEMENTATION_PLAN.md` §5).
+- **D-P-4 Goal = completeness over speed.** Parallelism is for isolation/depth, not finishing fast.

DECISIONS.md

@@ -0,0 +1,46 @@
+[package]
+name = "browser-use-agent"
+edition.workspace = true
+license.workspace = true
+version.workspace = true
+
+# The new async agent engine (rearchitecture Milestone 3, strategy B: parallel
+# rewrite). Built alongside the legacy sync `browser-use-core`; ported
+# subsystem-by-subsystem on top of the async `browser-use-llm`, with the TUI/CLI
+# cut over to it only at the end. WP-A0 lands the frozen compiling scaffold.
+
+[dependencies]
+browser-use-browser = { path = "../browser-use-browser" }
+browser-use-llm = { path = "../browser-use-llm" }
+browser-use-providers = { path = "../browser-use-providers" }
+browser-use-protocol = { path = "../browser-use-protocol" }
+browser-use-python-worker = { path = "../browser-use-python-worker" }
+browser-use-store = { path = "../browser-use-store" }
+anyhow.workspace = true
+async-trait.workspace = true
+base64.workspace = true
+bm25.workspace = true
+rand.workspace = true
+regex.workspace = true
+reqwest = { workspace = true, features = ["stream"] }
+rustls.workspace = true
+serde.workspace = true
+serde_json.workspace = true
+sha2.workspace = true
+thiserror.workspace = true
+toml.workspace = true
+tokio = { workspace = true, features = ["rt-multi-thread", "macros", "process", "io-util", "sync", "time", "net"] }
+tokio-util.workspace = true
+futures-util.workspace = true
+portable-pty.workspace = true
+
+[dev-dependencies]
+tokio = { workspace = true, features = ["rt-multi-thread", "macros", "sync", "time"] }
+tempfile.workspace = true
+
+[features]
+# Opt-in flag gating the live multi-provider smoke test so it is never compiled
+# into the default `cargo test` binary. The live call needs a real provider key
+# in the environment (OPENAI_API_KEY / ANTHROPIC_API_KEY / LLM_BROWSER_*). Run:
+#   cargo test -p browser-use-agent --features live -- --ignored live_model_smoke
+live = []