diff --git a/AGENTS.md b/AGENTS.md
index ab05c6a..e10f88c 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -3,6 +3,8 @@
 **Scope:** Repository-wide.
 **Child files:** `src/lib/AGENTS.md`, `src/lib/handlers/AGENTS.md`, `tests/AGENTS.md`, `tests/integration/AGENTS.md`.
 
+Child files define only local deltas; they inherit this file and do not override repository-wide rules unless they say so explicitly.
+
 ## Repository overview
 
 JavaScript GitHub Action (Node 20 runtime, `dist/index.js` entrypoint) with three event flows:
diff --git a/src/lib/handlers/AGENTS.md b/src/lib/handlers/AGENTS.md
index da2dbf9..66b8f81 100644
--- a/src/lib/handlers/AGENTS.md
+++ b/src/lib/handlers/AGENTS.md
@@ -1,9 +1,19 @@
 # HANDLER MODULE GUIDE
 
-## OVERVIEW
+## Scope and inheritance
+
+Applies to: `src/lib/handlers/` and descendants.
+
+Inherits repository-wide guidance from `../../../AGENTS.md` and services-layer guidance from `../AGENTS.md`.
+
+This file defines only local deltas for the command-handler subtree.
+
+## Overview
+
 Command handlers implement `/zai` behavior only after parsing + authorization; each module owns prompt construction, API call wiring, and response formatting. The `scheduled` handler is distinct: it executes scheduled tasks defined in `.zai-scheduled.yml` (and the manual `/zai update-agents` command) rather than responding to a standard review command.
 
-## WHERE TO LOOK
+## Where to look
+
 | Command | File | Notes |
 |---------|------|-------|
 | `/zai ask` | `src/lib/handlers/ask.js` | Uses continuity state and broad PR context |
@@ -16,7 +26,8 @@ Command handlers implement `/zai` behavior only after parsing + authorization; e
 | scheduled tasks | `src/lib/handlers/scheduled.js` | Largest module; cron-driven `.zai-scheduled.yml` tasks; grounded + validated AGENTS.md upgrades |
 | Handler registry | `src/lib/handlers/index.js` | Dispatcher map consumed by runtime (note: `scheduled` is exported but not in the `/zai` HANDLERS map) |
 
-## SCHEDULED MODULE (`scheduled.js`) KEY SYMBOLS
+## Scheduled module (`scheduled.js`) key symbols
+
 - `handleScheduledEvent` (entry) → `executeScheduledTask` (per-task) → `buildExecutionContext` → `getScheduledHandler` (registry lookup).
 - `handleUpdateAgentsTask` (grounded flow): gist command → `collectRepositoryContext` (`../repository-context.js`: real tree + existing AGENTS.md discovery + key files) → `buildAgentsUpgradePrompt` (embeds context, tells model it has NO live repo access) → `callZaiApiWithRetry` → `parseFileUpdatesFromResponse` (multi-format JSON) → `validateGeneratedAgentFiles` (`../agents-validation.js`: rejects non-AGENTS paths, out-of-scope writes, hallucinated content referencing non-existent files) → diff vs repo files → `createPR`.
 - Registry: `SCHEDULED_HANDLERS` (const) + `registerScheduledHandler`/`getAllScheduledHandlers` for extension.
@@ -24,25 +35,29 @@ Command handlers implement `/zai` behavior only after parsing + authorization; e
 - Config consumed from `src/lib/config/scheduled-config.js` (`loadScheduledConfig`, `getTasksToRun`, `getGistUrl`, `validateAgentsConfig`).
 - Scoping config (all optional, per-task in `.zai-scheduled.yml`): `context_paths`, `target_paths`, `exclude_paths`, `max_context_chars`, `max_file_chars`, `max_files_to_fetch`, `allow_create_new`, `update_existing_only`.
 
-## CONVENTIONS
+## Conventions
+
 - Keep command argument parsing explicit and reject invalid formats early.
 - Always use threaded replies (`replyToId`) for command results.
-- Reactions should reflect lifecycle: acknowledge -> work -> success/failure.
+- Reactions should reflect lifecycle: acknowledge → work → success/failure.
 - Keep prompts bounded via context truncation helpers; never pass raw unbounded patches.
 - Return user-safe failures; log internal details through shared logging helpers.
 
-## TESTING
+## Testing
+
 - Local handler unit coverage exists in `tests/handlers/`: `ask.test.js`, `explain.test.js`, `impact.test.js`, `review.test.js`, `scheduled.test.js`.
 - Scheduled pipeline coverage: `tests/handlers/scheduled.test.js` (registry, PR creation, parse, grounded `handleUpdateAgentsTask` flow incl. hallucination rejection), `tests/scheduled-config.test.js` (config + `validateAgentsConfig` scoping fields), `tests/repository-context.test.js` (tree/AGENTS.md discovery/budgets/globs), `tests/agents-validation.test.js` (path/hallucination/target-path guards incl. PR #15 regression).
 - End-to-end command pipeline behavior is validated in `tests/integration/command-pipeline.test.js`.
 - When changing parsing or output contracts, update both unit and integration assertions.
 
-## ANTI-PATTERNS
+## Anti-patterns
+
 - Parsing arguments with loose heuristics that silently alter user intent.
 - Posting top-level comments for command replies (breaks conversational threading).
 - Bypassing `auth.checkForkAuthorization` in a handler.
 - Embedding duplicate parser/auth logic that already exists upstream.
 
-## NOTES
+## Notes
+
 - Prefer adding helper functions within a handler module before introducing cross-handler coupling.
 - Keep marker constants stable once tests depend on them.
diff --git a/tests/AGENTS.md b/tests/AGENTS.md
index 4fb56f8..afd680a 100644
--- a/tests/AGENTS.md
+++ b/tests/AGENTS.md
@@ -1,9 +1,19 @@
 # TEST SUITE GUIDE
 
-## OVERVIEW
+## Scope and inheritance
+
+Applies to: `tests/` and descendants.
+
+Inherits repository-wide guidance from `../AGENTS.md`.
+
+This file defines only local deltas for the test suite. End-to-end integration specifics live in the child file `tests/integration/AGENTS.md`.
+
+## Overview
+
 Repository test coverage mixes module-focused tests in `tests/*.test.js` and scenario-driven flows in `tests/integration/*`.
 
-## STRUCTURE
+## Structure
+
 ```text
 tests/
 ├── *.test.js        # Module-level tests for lib/runtime units
@@ -11,10 +21,11 @@ tests/
 ├── helpers/         # Shared mocks and fixtures utilities
 ├── lib/             # Shared test helpers + code-scope tests
 ├── fixtures/        # Static test payloads (issue-comment-event.json, pr-event.json)
-└── integration/     # End-to-end command/review pipeline checks
+└── integration/     # End-to-end command/review pipeline checks (see child AGENTS)
 ```
 
-## WHERE TO LOOK
+## Where to look
+
 | Task | Location | Notes |
 |------|----------|-------|
 | Parser/auth/comment unit behavior | `tests/commands.test.js`, `tests/auth.test.js`, `tests/comments.test.js` | Fast regression checks |
@@ -24,23 +35,26 @@ tests/
 | Continuity and events | `tests/continuity.test.js`, `tests/events.test.js` | Hidden-marker state, event-type detection |
 | Code scope and window extraction | `tests/lib/code-scope.test.js` | Token budget, enclosing block, window extraction |
 | Describe handler | `tests/describe.test.js` | PR description generation |
-| Scheduled pipeline | `tests/handlers/scheduled.test.js`, `tests/scheduled-config.test.js`, `tests/repository-context.test.js`, `tests/agents-validation.test.js` | Config load + `validateAgentsConfig` scoping; `parseFileUpdatesFromResponse`; grounded `handleUpdateAgentsTask` flow incl. hallucination rejection; repo-context collection (tree/budgets/globs); validation guards incl. PR #15 regression |
+| Scheduled pipeline (units) | `tests/handlers/scheduled.test.js`, `tests/scheduled-config.test.js`, `tests/repository-context.test.js`, `tests/agents-validation.test.js` | Config load + `validateAgentsConfig` scoping; `parseFileUpdatesFromResponse`; grounded `handleUpdateAgentsTask` flow incl. hallucination rejection; repo-context collection (tree/budgets/globs); validation guards incl. PR #15 regression |
 | Scheduled pipeline (integration) | (pending) | End-to-end schedule event → context → Z.ai mock → validated PR is still a gap; unit coverage of the grounded flow exists via the `handleUpdateAgentsTask` seam (`__callZaiForTest`). |
-| Full command pipeline | `tests/integration/command-pipeline.test.js` | Parse -> auth -> handler -> output contract |
+| Full command pipeline | `tests/integration/command-pipeline.test.js` | Parse → auth → handler → output contract |
 | PR auto-review behavior | `tests/integration/pr-auto-review.test.js` | Marker upsert and PR event lifecycle |
 
-## CONVENTIONS
+## Conventions
+
 - Test framework: Vitest v3 (uses vitest globals: describe/test/expect); configured via `vitest.config.js`.
 - Keep tests deterministic with explicit mock payloads and marker assertions.
 - Prefer scenario names that encode trigger + expected visible outcome.
 - When changing comment markers or command UX, update integration snapshots/assertions immediately.
 
-## ANTI-PATTERNS
+## Anti-patterns
+
 - Deleting integration assertions to make behavior changes pass.
 - Asserting only internal calls without validating user-visible output.
 - Duplicating large fixtures inline when reusable fixtures already exist.
 
-## NOTES
+## Notes
+
 - Test command: `npm test` → `vitest run --coverage`.
 - Coverage uploaded to Codecov.
 - Integration tests are the safety net for command threading and marker idempotency.