Agent core in Rust after Pi: pi-ai, pi-agent-core, pi-coding-agent

.agents/skills/develop-maple-agent/SKILL.md

@@ -94,7 +94,7 @@ request. Keep fixture launch and environment setup shared with the Codex tests.
 
 Apply `$review-maple-security`'s trust-boundary and evidence methodology to the
 actual GPUI source; its Tauri-specific file list is for Research. Validate
-account isolation, tool approval, MCP/ACP inputs, persistence, and process
+account isolation, tool execution, MCP/ACP inputs, persistence, and process
 ownership at the layer implementing the effect. Never treat a passing source
 import or a native login screen as authenticated chat or containment proof.
 

apps/maple-agent/CLAUDE.md

@@ -5,8 +5,9 @@ Read the repository-root `AGENTS.md` and `$develop-maple-agent` as well.
 Commands below run from this component directory through its pinned Nix shell.
 
 GPUI desktop app for Maple. Workspace crates: `app` (package `maple-agent-app`, binary `maple-agent`),
-`crates/maple-agent`, `crates/maple-billing`. See `README.md` for the
-layout, prerequisites, and command line modes.
+`crates/maple-agent`, `crates/maple-billing`, and the agent core `crates/pi-ai`,
+`crates/pi-agent-core`, `crates/pi-coding-agent`, which must not depend on Maple
+crates. See `README.md` for the layout, prerequisites, and command line modes.
 
 Packaged Dev and Prod OAuth use their compiled Auth origins and a one-attempt
 loopback listener in `app/src/hosted_oauth.rs`; unpackaged builds retain the

apps/maple-agent/Cargo.toml

@@ -1,6 +1,6 @@
 [workspace]
 resolver = "2"
-members = ["crates/maple-agent", "crates/maple-billing", "app"]
+members = ["crates/maple-agent", "crates/maple-billing", "crates/pi-ai", "crates/pi-agent-core", "crates/pi-coding-agent", "app"]
 
 [workspace.package]
 edition = "2024"

apps/maple-agent/README.md

@@ -16,9 +16,19 @@ app/                  The maple-agent binary. Owns the window, login, chat,
 crates/maple-agent/   Maple's transport-neutral agent runtime, extracted from
                       the Tauri app with Tauri removed. Owns embedded Goose,
                       the Maple provider over the Maple Rust SDK, developer
-                      tools, permission policy, account-scoped session
-                      storage, and the ACP server.
+                      tools, account-scoped session storage, and the ACP
+                      server.
 crates/maple-billing/ HTTP client for the Maple billing API.
+crates/pi-ai/         Model messages, streaming events and an OpenAI-compatible
+                      provider, after Pi's pi-ai package.
+crates/pi-agent-core/ The agent loop with tools, hooks and message queues, after
+                      Pi's pi-agent-core package.
+crates/pi-coding-agent/
+                      Sessions as an append-only tree, compaction, skills and
+                      prompt templates, the extension API and the agent
+                      session, after Pi's coding-agent core. The pi-* crates
+                      have no Maple dependencies; the runtime does not use
+                      them yet.
 docs/                 Theme spec measured from the Tauri app.
 scripts/              One maintainer helper: screenshot.py takes a desktop
                       screenshot through the xdg portal on GNOME Wayland.
@@ -46,9 +56,8 @@ remove Tauri:
 - public visibility opened on the service surface the app consumes
 
 Goose is pinned to a fork revision based on upstream v1.53.0 in this
-component’s Cargo manifests and lockfile. The fork forwards subagent approvals
-to the parent and supports ephemeral native clients for embedded CUA. Research
-has an independent dependency graph.
+component’s Cargo manifests and lockfile. The fork supports ephemeral native
+clients for embedded CUA. Research has an independent dependency graph.
 
 ## Features
 
@@ -61,31 +70,26 @@ has an independent dependency graph.
   the `acp` mode skip sign-in. The window opens while the saved session is
   checked, and a check that cannot reach the server keeps the credentials
   for the next launch; only a refusal from the server signs the user out.
-- Agent chat with streaming Markdown, tool calls, permission prompts,
-  agent questions, image attachments (picker, paste, or drag and drop),
-  a per-message Copy button, and a context-window indicator.
+- Agent chat with streaming Markdown, tool calls, agent questions, image
+  attachments (picker, paste, or drag and drop), a per-message Copy button,
+  and a context-window indicator. Every tool call runs without asking for
+  approval.
 - Slash commands in the composer: `/btw` asks a side question the task
   never sees, plus `/compact`, `/new`, `/pin`, `/web`, `/model`, and
   `/help`. The account's skills appear in the same list.
 - The task's latest todo list stays pinned above the composer.
-- Subagents: the task can give a piece of work to a subagent with the
-  `delegate` tool, which runs it in its own context. The Goose fork makes
-  the subagent inherit the task's permission mode and forwards its approval
-  requests to the parent session. Maple also asks before each handoff in
-  Read only mode. Subagents show above the composer with the tool each one
-  runs and how long it has worked. A subagent that runs in the background
-  keeps its row after the turn ends, and Maple tells the task when it
-  finishes, with a bounded result in the running turn or a new turn Maple
-  starts automatically. The task can use `load` to retrieve any truncated output.
 - External agents: a task can hand work to Codex or Claude Code installed on
   this computer with the `agent_start`, `agent_send`, `agent_status`,
   `agent_cancel`, and `list_agent_providers` tools, once the provider is enabled
   under Settings > Integrations. Each agent runs in the project with its own
-  account, context, and sandbox settings; whatever it asks approval for
-  comes to you through Maple's permission card, and Allow all grants it. Its progress streams
-  into the tool call's row and its row above the composer has a Stop
-  button. Three skills, `/handoff`, `/committee`, and `/advisor`, teach the
-  task when and how to delegate. See
+  account, context, and sandbox settings; Maple accepts its approval
+  requests, and its questions still come to you through Maple's question
+  card. Its progress streams into the tool call's row, and its row above
+  the composer shows how long it has worked, with a Stop button. A
+  background agent keeps its row after the turn ends, and Maple tells the
+  task when it finishes, with a bounded result in the running turn or a new
+  turn Maple starts automatically. Three skills, `/handoff`, `/committee`,
+  and `/advisor`, teach the task when and how to delegate. See
   [`docs/external-agents.md`](docs/external-agents.md).
 - Voice: dictate a message with the microphone button, and read any
   message aloud. Both use Maple's speech models; the voice and speed
@@ -111,15 +115,15 @@ has an independent dependency graph.
   projects that provide skills ask once for a trust decision before
   their guidance loads.
 - Sessions grouped by project, with rename, archive, and restore.
-- Settings: General (default permission mode, web tools, appearance,
-  tool call details, desktop notifications, tool call summaries, composer Vim,
-  application Vim, and the speech voice and speed), Keyboard Shortcuts, System
-  prompt, Integrations (detected built-ins and custom MCP servers), Usage (plan
-  meter from the billing API plus totals from the Goose ledger), and About.
+- Settings: General (web tools, appearance, tool call details, desktop
+  notifications, tool call summaries, composer Vim, application Vim, and the
+  speech voice and speed), Keyboard Shortcuts, System prompt, Integrations
+  (detected built-ins and custom MCP servers), Usage (the plan meter from
+  the billing API), and About.
 - Dark and light themes; the default follows the system.
 - Billing status from the Maple billing API.
-- Desktop notifications when a task finishes, asks a question, or needs
-  permission while the window is not focused.
+- Desktop notifications when a task finishes or asks a question while the
+  window is not focused.
 - Release check on launch: a banner links to a newer GitHub release.
   Nothing is downloaded or installed by the app.
 - Window size and maximized state persist between launches.
@@ -151,20 +155,19 @@ the embedded transport.
 Enabling an integration sets a device-local default for new tasks. Existing
 tasks keep their frozen integration choice and expose CUA as an independent
 per-task switch in the composer. A task that never chose a backend adopts the
-device default only when it can actually run it. A detected standalone
-CuaDriver, which Maple looks for on macOS only, remains a legacy-compatible
-backend until the user explicitly sets up the built-in one;
-Maple does not install or update it, start or stop its daemon, or alter another
-client's configuration. Custom STDIO and Streamable HTTP MCP servers remain
-account configuration that may roam between devices.
+device default only when it can actually run it. Maple no longer looks for a
+separately installed CuaDriver application: a task saved with its stdio entry
+loses that entry on its next run, and a device setting that selected it reads
+as not set up. Custom STDIO and Streamable HTTP MCP servers remain account
+configuration that may roam between devices.
 
 The embedded design, migration rules, privacy boundary, and preview limits are
 documented in [`docs/embedded-cua.md`](docs/embedded-cua.md).
 
 #### Claude Code
 
 Settings > Integrations lists Claude Code (`claude`) alongside Codex, with the
-same per-task selection, streamed activity, permission cards, and Stop control.
+same per-task selection, streamed activity, question cards, and Stop control.
 Install the Claude Code CLI on the app's PATH and sign in using
 `claude auth login`. Maple uses a Rust transport adapted from Goose's Claude
 Code provider. The CLI is the only external runtime dependency. The integration
@@ -443,7 +446,6 @@ ignored. See [build profiles](docs/release-profiles.md).
 | `MAPLE_BILLING_API_URL` | Maple billing API. | `https://billing.opensecret.cloud` |
 | `MAPLE_CLIENT_ID` | OpenSecret client id (UUID). | Maple's id |
 | `MAPLE_MODEL` | Model to select at start. | Runtime default |
-| `MAPLE_PERMISSION_MODE` | `smart_approve` or `auto`. Overrides the saved setting. | Saved setting |
 | `MAPLE_CONTEXT_LIMIT` | Context window size in tokens, when the model catalog does not report one. | Catalog value |
 | `GOOSE_SHELL` | Shell for the agent's shell tool. | `bash` (Windows: `cmd`) |
 | `MAPLE_UPDATE_REPO` | GitHub `owner/repo` containing stable `maple-agent-vX.Y.Z` releases. | `MaplePrivacyLabs/Maple` |
@@ -467,11 +469,10 @@ shows unpackaged builds; packaged Dev and Prod append `maple-agent-dev` and
 | --- | --- |
 | `<config>/settings.json` | App settings. |
 | `<config>/agent/accounts/<scope>/config.json` | Per-account agent configuration (default root, model, custom MCP servers, project trust). May roam between machines. |
-| `<config>/agent/accounts/<scope>/goose/config/` | Goose permission file for the account. |
 | `<config>/agent/accounts/<scope>/goose/config/skills/` | Skills the account's tasks can load, including the delegation skills Maple installs while any external agent is enabled. |
 | `<config>/agent/goose-runtime/` | Goose process configuration. |
 | `<local data>/auth.json` | Sign-in credentials (mode 0600). Device-local; never in a roaming profile. |
-| `<local data>/agent/accounts/<scope>/integrations.json` | Per-account defaults and validated launch details for integrations detected on this device. |
+| `<local data>/agent/accounts/<scope>/integrations.json` | Per-account defaults for the integrations on this device. |
 | `<local data>/agent/accounts/<scope>/goose/data/sessions/sessions.db` | Goose session history and usage ledger (SQLite, WAL). |
 | `<local data>/agent/accounts/<scope>/tool_summaries.db` | Model-written one-line summaries of tool calls (SQLite, WAL). |
 | `<local data>/agent/accounts/<scope>/attachments/` | Image attachments. |
@@ -504,6 +505,12 @@ a release build. `just ci` is the full local format, lint, build and test gate;
 `just release` separately validates the optimized binary. PR jobs have no
 signing or publishing credentials. See the root agent guide for shared checks.
 
+Two by-hand checks complement the automated ones: the
+[scenario checklist](docs/scenario-checklist.md), one scenario per feature
+with the results of each run, and the
+[performance check](docs/performance-check.md), a quick release-build sanity
+check.
+
 ## Update and release boundary
 
 Prod and unpackaged builds accept stable `maple-agent-vX.Y.Z` releases. Dev

apps/maple-agent/app/src/backend.rs

@@ -45,18 +45,6 @@ pub struct PendingQuestion {
     pub questions: Vec<maple_agent::agent::AgentQuestion>,
 }
 
-#[derive(Debug, Clone)]
-pub struct PendingPermission {
-    pub session_id: String,
-    pub run_id: String,
-    pub request_id: String,
-    pub tool_name: String,
-    pub prompt: Option<String>,
-    /// Pretty-printed tool arguments, formatted once when the request
-    /// arrives instead of on every frame.
-    pub arguments: Arc<str>,
-}
-
 // Re-exported for the settings screens; a headless build has no reader.
 #[cfg_attr(not(feature = "desktop"), allow(unused_imports))]
 pub use maple_agent::maple_api::{
@@ -1562,13 +1550,11 @@ impl AgentBackend {
             Some(AgentStartRequest {
                 project_root: None,
                 model,
-                mode,
             }) => {
                 let config = handle.load_config().await?;
                 Some(AgentStartRequest {
                     project_root: gui_start_root(&config),
                     model,
-                    mode,
                 })
             }
             other => other,
@@ -2093,24 +2079,6 @@ impl AgentBackend {
             .await
     }
 
-    /// Set the permission policy for a session: "smart_approve" asks for
-    /// each gated tool, "auto" approves everything (bypass).
-    pub async fn set_permission_mode(
-        &self,
-        user_id: &str,
-        session_id: &str,
-        mode: &str,
-    ) -> Result<(), String> {
-        self.service
-            .handle_for_user(user_id)
-            .await?
-            .set_permission_mode(maple_agent::agent::AgentPermissionModeRequest {
-                session_id: session_id.to_string(),
-                mode: mode.to_string(),
-            })
-            .await
-    }
-
     /// Compact a session's history now; reload the session afterwards.
     pub async fn compact_session(&self, user_id: &str, session_id: &str) -> Result<(), String> {
         self.service
@@ -2356,35 +2324,12 @@ impl AgentBackend {
             .await
     }
 
-    pub async fn permission_respond(
-        &self,
-        user_id: &str,
-        session_id: &str,
-        request_id: &str,
-        allow: bool,
-    ) -> Result<(), String> {
-        self.service
-            .handle_for_user(user_id)
-            .await?
-            .permission_respond(maple_agent::agent::AgentPermissionResponse {
-                session_id: session_id.to_string(),
-                request_id: request_id.to_string(),
-                decision: if allow {
-                    "allow_once".to_string()
-                } else {
-                    "deny_once".to_string()
-                },
-            })
-            .await
-    }
-
     /// Standard start request for this app: the saved project root (see
-    /// `start_runtime`) with the configured model and the SmartApprove policy.
+    /// `start_runtime`) with the configured model.
     pub fn default_start_request(&self) -> AgentStartRequest {
         AgentStartRequest {
             project_root: None,
             model: std::env::var("MAPLE_MODEL").ok(),
-            mode: None,
         }
     }