Tool Inventory · Two-Layer Dispatch × Language Intelligence

Route: read → workspace_read_file (Tauri command) → JSON-RPC fs/readTextFile → codepapr-server → codepapr-core::workspace_fs. If symbol is passed, AST locates the exact line range; output is truncated safely (>60k chars folded in middle, >100k offloaded to disk with 2k preview, 150k hard cap).

Route: write → workspace_write_file (Tauri command) → JSON-RPC fs/writeTextFile → codepapr-server → codepapr-core::workspace_fs. Pre-check compares before/after syntax errors, rejecting regressions; triggers lsp_open_document to refresh diagnostics, then sends a JSON-RPC notification (workspace-files-changed) back through the Tauri event bus to invalidate stale graph caches.

Route: edit → workspace_apply_patch. Byte-exact String.indexOf matching. Ambiguities resolve symbols via AST to produce actionable disambiguation errors; returns fresh LSP diagnostics for the modified file (max 20MB).

Route: patch → workspace_apply_diff. Pre-checks ambiguity against live buffer content → applies sequential diffs → runs per-file AST verification & writes → emits batch mutation updates.

Route: Defaults to workspace_search_text; semantic: true routes to workspace_workspace_symbol. Falls back to regex with degraded: true when LSP is absent.

Route: glob → globToRegex → workspace_search_files.

Route: list → workspace_list_files (Tauri command) → JSON-RPC fs/listFiles → codepapr-server → codepapr-core directory scan + concurrent AST top-level symbol extraction into symbolsByFile.

Route: read_image → workspace_read_image (Tauri command) → JSON-RPC fs/readFile → codepapr-server → codepapr-core native I/O. Stripped from toolsets unless the current slot has vision enabled (explicit profile setting wins; unconfigured known text-only models default off; a vision-capable fast slot is used as fallback).

Route: Dispatches to corresponding LSP handlers; on timeout (30s) or uninitialized servers, falls back to AST symbol queries with source: "ast", confidence: "medium".

Route: Calculates WorkspaceEdits via LSP textDocument/rename before committing, ensuring atomic multi-file consistency.

Route: Single-file queries use JSON-RPC lsp/diagnostics (codepapr-server → codepapr-core::lsp; short-circuits on fresh diagnostics, 2s timeout guard); project-wide mode aggregates issues across active language providers.

Constructs unified project graphs from AST and LSP edges; includes a 60-second build cache with in-flight deduplication, cleared automatically on file mutations. Also acts as the fallback backend for lsp queries.

Route: Foreground execution routes to workspace_run_shell_command; background jobs route to workspace_start_shell_background_command with process supervision and automatic teardown.

Dispatches to fine-grained workspace_git_* handlers. Reset actions enforce target parameters and backup branch prefixes with undo snapshot rollback protection.

webfetch routes to web_fetch_url or web_download_file; websearch queries the backend search aggregator directly.

Built on Tauri embedded Webviews, supporting selector-based clicks, form fills, and visual page verification for automated web debugging.

The Agent has no memory_* tools and direct file writes are rejected; every turn reads the file into session bootstrap (effective the next turn after saving). The panel (Context inspector → Memory) is that file's editor: budget bar + save + curator status line.

app_render is strictly disabled for subagents to prevent recursive rendering loops; pass appId only. Permissions, agents, and backend processes live in manifest.json. app_list is App-mode only: duplicate checks and running state. The coding Agent discovers publish targets from session context, not from list.

How the Agent knows what to push: do not call app_list. That tool is App-mode only (duplicate checks / backend lifecycle). The coding Agent reads session context "## Enabled plugins", which lists enabled overlays that declared inbox plus fullscreen apps that declared inbox. Follow that appId / channel / example. Self-refreshing widgets without inbox (tickers, clocks) stay out of context — do not publish to them.

Once inbox is declared, only those channels are accepted (unknown channels are rejected with the valid list). inbox:* keys are written only by app_publish — the app side is read-only. Payload capped at 256KB; last 200 events kept per channel. Examples are truncated before they enter context so the catalog stays short.

skill routes to skill_load; question returns structured UI selection cards; availability across modes (Ask/Plan/Agent) is governed by strict whitelists.

graph is soft-hidden from the LLM prompt (registered internally for UI panels and Explore). Ask mode enforces write guards via MUTATING_TOOL_NAMES.

Registered with exposeGraphToLlm: true to unmask the graph tool; read_image is dynamically stripped when multimodal support is disabled.

Downloads external resources safely into .CodePapr/downloads/ via webfetch save: true for subsequent consumption by the main agent.

Configured with an empty whitelist {}; if information is insufficient, it instructs the main agent to gather required facts.

Driven by GoalRunner with a dedicated budget: max 6 tool turns, 3-minute timeout, independent token limits, and thinking mode disabled.

Isolated from skills/memory injections with cancellation signal handling; automatically falls back to rule-based compaction if fast models are unavailable.