@dreb/coding-agent 2.30.0 → 2.31.0

package/README.md

@@ -173,7 +173,7 @@ Type `/` in the editor to trigger commands. [Extensions](#extensions) can regist
 | `/tree` | Jump to any point in the session and continue from there |
 | `/fork` | Create a new session from the current branch |
 | `/compact [prompt]` | Manually compact context, optional custom instructions |
-| `/copy` | Open multi-select message picker to copy any messages to clipboard |
+| `/copy` | Open multi-select message picker to copy any messages to clipboard. Assistant reasoning is excluded by default and offered as a separate, selectable `Thinking` row. |
 | `/dream` | Consolidate and prune memories — backs up, merges duplicates, scans sessions for patterns |
 | `/export [file]` | Export session to HTML file |
 | `/buddy` | Terminal companion — hatch, pet, reroll, set model, or hide. See [docs/buddy.md](docs/buddy.md) |
@@ -288,6 +288,17 @@ dreb loads `AGENTS.md` (or `CLAUDE.md`) at startup from:
 
 Use for project instructions, conventions, common commands. All matching files are concatenated.
 
+### Nested context auto-load
+
+The startup scan only walks **upward** from cwd, so an `AGENTS.md`/`CLAUDE.md` that lives in a **subdirectory** (e.g. a monorepo subpackage) is not loaded until the agent actually works there. When `context.autoLoadNested` is enabled (default), the first time any tool operates in a directory, dreb loads that directory's context files and appends them to the tool result — explaining why the load happened and headering each file with its source path.
+
+- Triggers on path-bearing tools (`read`, `edit`, `write`, `grep`, `find`, `ls`) and on `bash` commands that begin with `cd <dir>`.
+- Walks up from the target directory to a sensible ceiling: the session cwd (for in-tree targets), otherwise the outermost git repo root, otherwise the outermost directory containing a context file.
+- Each file is injected at most once per session, and files already loaded at startup are never repeated. Applies to subagents too — including subagents that work in a different repo than the parent.
+- Disable with `context.autoLoadNested: false` in settings, or toggle "Auto-load nested context" in `/settings`.
+
+**Security caution:** when working across untrusted or third-party repositories, their `AGENTS.md`/`CLAUDE.md` files may be auto-injected into the agent's context, which is a prompt-injection consideration. Auto-loaded content is secret-scrubbed before injection, but extension `tool_result` transforms do not see it because nested context is injected after those transforms for cache safety.
+
 ### System Prompt
 
 Replace the default system prompt with `.dreb/SYSTEM.md` (project) or `~/.dreb/agent/SYSTEM.md` (global). Append without replacing via `APPEND_SYSTEM.md`.

package/dist/core/agent-session.d.ts

@@ -185,6 +185,14 @@ export declare class AgentSession {
     private _customTools;
     private _baseToolDefinitions;
     private _cwd;
+    /**
+     * Per-session realpaths of context files already loaded (seeded lazily from the
+     * session-start context set). Ensures each nested AGENTS.md/CLAUDE.md is injected
+     * at most once. `undefined` until first use.
+     */
+    private _nestedContextLoaded;
+    /** Negative cache of target directories already scanned for nested context. */
+    private _nestedContextScannedDirs;
     private _extensionRunnerRef?;
     private _initialActiveToolNames?;
     private _baseToolsOverride?;
@@ -223,6 +231,14 @@ export declare class AgentSession {
      * happens here instead of in wrappers.
      */
     private _installAgentToolHooks;
+    /**
+     * Compute a nested-context injection block for a tool call, or `null` when nothing
+     * should be injected. Resolves the directory the tool operates in, walks up to a
+     * sensible ceiling collecting not-yet-loaded AGENTS.md/CLAUDE.md files, and formats
+     * them. Each directory is scanned at most once (negative cache) and each file is
+     * injected at most once per session (realpath dedup). Gated by `context.autoLoadNested`.
+     */
+    private _computeNestedContextBlock;
     /**
      * Install guardrails for background agent interactions:
      * - Layer B: Sentinel monitor — steer if the parent model generates suspicious tokens