pi-agent: add agent-browser

flake.nix

@@ -4,6 +4,7 @@
     inputs@{ flake-parts, import-tree, ... }:
     flake-parts.lib.mkFlake { inherit inputs; } ((import-tree ./nix) // { flake.paths.root = ./.; });
   inputs = {
+    agent-browser.url = "github:rrvsh/agent-browser-nix";
     flake-parts.inputs.nixpkgs-lib.follows = "nixpkgs";
     flake-parts.url = "github:hercules-ci/flake-parts";
     home-manager.inputs.nixpkgs.follows = "nixpkgs";

nix/modules/pi-agent.nix

@@ -20,35 +20,71 @@ in
     nixos.pi-agent = osModule;
     homeManager.pi-agent =
       { pkgs, config, ... }:
+      let
+        system = pkgs.stdenv.hostPlatform.system;
+        agentBrowser = inputs.agent-browser.packages.${system}.agent-browser;
+        # NixOS cannot run agent-browser's downloaded generic Chrome, so point the
+        # extension at Nix-provided Chromium for fresh browser launches by default.
+        agentBrowserConfig = {
+          version = 1;
+        }
+        // lib.optionalAttrs pkgs.stdenv.isLinux {
+          browser.executablePath = "${pkgs.chromium}/bin/chromium";
+        };
+        agentBrowserStatePath = "${config.home.homeDirectory}/.agent-browser/state/agent.json";
+      in
       {
+        home.packages = [ agentBrowser ] ++ lib.optionals pkgs.stdenv.isLinux [ pkgs.chromium ];
+        home.file.".pi/config/pi-agent-browser-native/config.json".text =
+          builtins.toJSON agentBrowserConfig;
         programs.pi-coding-agent = {
           enable = true;
-          package = inputs.pi.packages.${pkgs.stdenv.hostPlatform.system}.pi-coding-agent;
-          extraPackages = [ pkgs.nodejs_22 ];
+          package = inputs.pi.packages.${system}.pi-coding-agent;
+          extraPackages = [
+            pkgs.nodejs_22
+            agentBrowser
+          ];
           settings.lastChangelogVersion = lib.getVersion config.programs.pi-coding-agent.package;
           settings.packages = [
+            # Keep this pinned with the agent-browser flake input: pi-agent-browser-native
+            # tracks specific agent-browser CLI versions in its command surface and result parsing.
+            "npm:pi-agent-browser-native@0.2.64"
             "npm:pi-mcp-adapter"
             "npm:pi-subagents"
             "npm:pi-web-access"
           ];
-          context = ''
-            # Global pi instructions
-
-            - If a command is missing, use comma: `, <command> [args...]`
-            - For Python, run via comma: `, python3 ...`
-            - Load and follow the global skill: `comma` (`~/.pi/agent/skills/comma/SKILL.md`)
-            - For quick web search, prefer `ddgr` first (example: `ddgr -n 5 -x --np <query>`).
-            - If `ddgr` is missing, use comma: `, ddgr -n 5 -x --np <query>`.
-            - For scripted/API search, use Python via comma and prefer stdlib first (`urllib`, `json`, `re`, `xml.etree`).
-            - Avoid assuming third-party Python deps (e.g., `requests`, `bs4`) are installed.
-            - DuckDuckGo HTML may return bot challenges; if blocked, fall back to Bing RSS or site APIs.
-            - Useful fallbacks:
-              - Bing RSS: `https://www.bing.com/search?format=rss&q=...`
-              - GitHub repo search API: `https://api.github.com/search/repositories?q=...`
-            - To inspect flake inputs locally:
-              - `nix flake metadata --json | jq '.locks.nodes | keys'`
-              - `nix eval --impure --raw --expr '(builtins.getFlake (toString ./.)).inputs.<name>.outPath'`
-          '';
+          context = lib.concatStringsSep "\n" (
+            [
+              "# Global pi instructions"
+              ""
+              "- If a command is missing, use comma: `, <command> [args...]`"
+              "- For Python, run via comma: `, python3 ...`"
+              "- Load and follow the global skill: `comma` (`~/.pi/agent/skills/comma/SKILL.md`)"
+              "- Avoid assuming third-party Python deps (e.g., `requests`, `bs4`) are installed."
+              "- To inspect flake inputs locally:"
+              "  - `nix flake metadata --json | jq '.locks.nodes | keys'`"
+              "  - `nix eval --impure --raw --expr '(builtins.getFlake (toString ./.)).inputs.<name>.outPath'`"
+              "- Agent browser:"
+              "  - `agent-browser` is installed in the user shell and on Pi's wrapped runtime `PATH`."
+              "  - Use the native `agent_browser` tool for browser automation; do not drive `agent-browser` through bash from Pi."
+              "  - Auth-walled and account-specific sites, especially LinkedIn, must use `agent_browser` with the shared auth state before declaring an auth wall or login problem. Do not use `ddgr`, `web_search`, `fetch_content`, `curl`, or bash to read LinkedIn page content."
+              "  - Use one shared auth state file for all sites: `${agentBrowserStatePath}`. Do not create per-service restore buckets such as `--restore=linkedin` unless explicitly asked."
+              "  - Normal authenticated browsing is headless and state-backed. Launch fresh with the shared state file, e.g. `args`: `[\"--state\", \"${agentBrowserStatePath}\", \"open\", \"https://www.linkedin.com/feed/\"]`, `sessionMode`: `\"fresh\"`. Use the same pattern for LinkedIn profiles, posts, search, and jobs by replacing the URL."
+              "  - After a state-backed open, inspect with `snapshot -i`, `get`, or `eval --stdin`. If the page still shows a login form or password field, ask the user whether to refresh the shared login; do not stop at \"auth wall\"."
+              "  - Do not use `--headed` for normal browsing. Use `--headed` only for initial auth setup or refreshing an expired login when the user needs to sign in manually."
+              "  - For auth setup/refresh, launch a headed session and keep it alive with a long wait while the user signs in. Example LinkedIn flow:"
+              "    ```json"
+              ''{"args":["--headed","batch"],"stdin":"[[\"open\",\"https://www.linkedin.com/feed/\"],[\"wait\",\"--fn\",\"location.pathname.includes('/feed') && !document.querySelector('input[type=password]')\",\"--timeout\",\"600000\"]]","sessionMode":"fresh","timeoutMs":650000}''
+              "    ```"
+              "  - After the user is logged in, save or refresh the shared state with `args`: `[\"state\", \"save\", \"${agentBrowserStatePath}\"]`."
+            ]
+            ++ lib.optionals pkgs.stdenv.isLinux [
+              "  - On Linux/NixOS, the agent_browser extension config points fresh launches at Nix-provided Chromium; the generic browser downloaded by `agent-browser install` is not NixOS-compatible."
+            ]
+            ++ lib.optionals pkgs.stdenv.isDarwin [
+              "  - On Darwin, run `agent-browser install` if the agent browser executable is not present."
+            ]
+          );
         };
       };
   };