diff --git a/src/features/mcp-oauth/AGENTS.md b/src/features/mcp-oauth/AGENTS.md
new file mode 100644
index 000000000..bca00331f
--- /dev/null
+++ b/src/features/mcp-oauth/AGENTS.md
@@ -0,0 +1,54 @@
+# src/features/mcp-oauth/ — OAuth 2.0 + PKCE + DCR for MCP Servers
+
+**Generated:** 2026-02-19
+
+## OVERVIEW
+
+18 files. Full OAuth 2.0 authorization flow for MCP servers requiring authentication. Implements PKCE (RFC 7636), Dynamic Client Registration (DCR, RFC 7591), and resource indicators (RFC 8707). Used by `bunx oh-my-opencode mcp-oauth login`.
+
+## AUTHORIZATION FLOW
+
+```
+1. discovery.ts → fetch /.well-known/oauth-authorization-server
+2. dcr.ts → Dynamic Client Registration (if server supports it)
+3. oauth-authorization-flow.ts → generate PKCE verifier/challenge
+4. callback-server.ts → local HTTP server on random port for redirect
+5. Open browser → authorization URL
+6. callback-server.ts → receive code + state
+7. provider.ts → exchange code for token (with PKCE verifier)
+8. storage.ts → persist token to ~/.config/opencode/mcp-oauth/
+9. step-up.ts → handle step-up auth if initial token insufficient
+```
+
+## KEY FILES
+
+| File | Purpose |
+|------|---------|
+| `oauth-authorization-flow.ts` | PKCE helpers: `generateCodeVerifier()`, `generateCodeChallenge()`, `buildAuthorizationUrl()` |
+| `callback-server.ts` | Local HTTP redirect server — listens for OAuth callback |
+| `provider.ts` | `OAuthProvider` — token exchange, refresh, revoke |
+| `discovery.ts` | Fetch + parse OAuth server metadata from well-known endpoint |
+| `dcr.ts` | Dynamic Client Registration — register this app with OAuth server |
+| `resource-indicator.ts` | RFC 8707 resource indicator handling |
+| `step-up.ts` | Handle step-up authentication challenges |
+| `storage.ts` | Persist tokens to `~/.config/opencode/mcp-oauth/{server-hash}.json` |
+| `schema.ts` | Zod schemas for OAuth server metadata, token response, DCR |
+
+## PKCE IMPLEMENTATION
+
+- Code verifier: 32 random bytes → base64url (no padding)
+- Code challenge: SHA-256(verifier) → base64url
+- Method: `S256`
+
+## TOKEN STORAGE
+
+Location: `~/.config/opencode/mcp-oauth/` — one JSON file per MCP server (keyed by server URL hash).
+Fields: `access_token`, `refresh_token`, `expires_at`, `client_id`.
+
+## CLI COMMANDS
+
+```bash
+bunx oh-my-opencode mcp-oauth login <server-url>   # Full PKCE flow
+bunx oh-my-opencode mcp-oauth logout <server-url>  # Revoke + delete token
+bunx oh-my-opencode mcp-oauth status               # List stored tokens
+```
diff --git a/src/hooks/atlas/AGENTS.md b/src/hooks/atlas/AGENTS.md
new file mode 100644
index 000000000..f541acf4c
--- /dev/null
+++ b/src/hooks/atlas/AGENTS.md
@@ -0,0 +1,64 @@
+# src/hooks/atlas/ — Master Boulder Orchestrator
+
+**Generated:** 2026-02-19
+
+## OVERVIEW
+
+17 files (~1976 LOC). The `atlasHook` — Continuation Tier hook that monitors session.idle events and forces continuation when boulder sessions (ralph-loop, task-spawned agents) have incomplete work. Also enforces write/edit policies for subagent sessions.
+
+## WHAT ATLAS DOES
+
+Atlas is the "keeper of sessions" — it tracks every session and decides:
+1. Should this session be forced to continue? (if boulder session with incomplete todos)
+2. Should write/edit be blocked? (policy enforcement for certain session types)
+3. Should a verification reminder be injected? (after tool execution)
+
+## DECISION GATE (session.idle)
+
+```
+session.idle event
+  → Is this a boulder/ralph/atlas session? (session-last-agent.ts)
+  → Is there an abort signal? (is-abort-error.ts)
+  → Failure count < max? (state.promptFailureCount)
+  → No running background tasks?
+  → Agent matches expected? (recent-model-resolver.ts)
+  → Plan complete? (todo status)
+  → Cooldown passed? (5s between injections)
+  → Inject continuation prompt (boulder-continuation-injector.ts)
+```
+
+## KEY FILES
+
+| File | Purpose |
+|------|---------|
+| `atlas-hook.ts` | `createAtlasHook()` — composes event + tool handlers, maintains session state |
+| `event-handler.ts` | `createAtlasEventHandler()` — decision gate for session.idle events |
+| `boulder-continuation-injector.ts` | Build + inject continuation prompt into session |
+| `system-reminder-templates.ts` | Templates for continuation reminder messages |
+| `tool-execute-before.ts` | Block write/edit based on session policy |
+| `tool-execute-after.ts` | Inject verification reminders post-tool |
+| `write-edit-tool-policy.ts` | Policy: which sessions can write/edit? |
+| `verification-reminders.ts` | Reminder content for verifying work |
+| `session-last-agent.ts` | Determine which agent owns the session |
+| `recent-model-resolver.ts` | Resolve model used in recent messages |
+| `subagent-session-id.ts` | Detect if session is a subagent session |
+| `sisyphus-path.ts` | Resolve `.sisyphus/` directory path |
+| `is-abort-error.ts` | Detect abort signals in session output |
+| `types.ts` | `SessionState`, `AtlasHookOptions`, `AtlasContext` |
+
+## STATE PER SESSION
+
+```typescript
+interface SessionState {
+  promptFailureCount: number  // Increments on failed continuations
+  // Resets on successful continuation
+}
+```
+
+Max consecutive failures before 5min pause: 5 (exponential backoff in todo-continuation-enforcer).
+
+## RELATIONSHIP TO OTHER HOOKS
+
+- **atlasHook** (Continuation Tier): Master orchestrator, handles boulder sessions
+- **todoContinuationEnforcer** (Continuation Tier): "Boulder" mechanism for main Sisyphus sessions
+- Both inject into session.idle but serve different session types
diff --git a/src/hooks/rules-injector/AGENTS.md b/src/hooks/rules-injector/AGENTS.md
new file mode 100644
index 000000000..70a13b2c5
--- /dev/null
+++ b/src/hooks/rules-injector/AGENTS.md
@@ -0,0 +1,53 @@
+# src/hooks/rules-injector/ — Conditional Rules Injection
+
+**Generated:** 2026-02-19
+
+## OVERVIEW
+
+19 files (~1604 LOC). The `rulesInjectorHook` — Tool Guard Tier hook that auto-injects AGENTS.md (and similar rule files) into context when a file in a directory is read, written, or edited. Proximity-based: closest rule file to the target path wins.
+
+## HOW IT WORKS
+
+```
+tool.execute.after (read/write/edit/multiedit)
+  → Extract file path from tool output
+  → Find rule files near that path (finder.ts)
+  → Already injected this session? (cache.ts)
+  → Inject rule content into tool output (injector.ts)
+```
+
+## TRACKED TOOLS
+
+`["read", "write", "edit", "multiedit"]` — triggers only on file manipulation tools.
+
+## KEY FILES
+
+| File | Purpose |
+|------|---------|
+| `hook.ts` | `createRulesInjectorHook()` — wires cache + injector, handles tool events |
+| `injector.ts` | `createRuleInjectionProcessor()` — orchestrates find → cache → inject |
+| `finder.ts` | `findRuleFiles()` + `calculateDistance()` — locate AGENTS.md near target path |
+| `rule-file-finder.ts` | Walk directory tree to find AGENTS.md / .rules files |
+| `rule-file-scanner.ts` | Scan for rule files in a directory |
+| `matcher.ts` | Match file paths against rule file scope |
+| `rule-distance.ts` | Calculate path distance between file and rule file |
+| `project-root-finder.ts` | Find project root (stops at .git, package.json) |
+| `output-path.ts` | Extract file paths from tool output text |
+| `cache.ts` | `createSessionCacheStore()` — per-session injection dedup |
+| `storage.ts` | Persist injected paths across tool calls |
+| `parser.ts` | Parse rule file content |
+| `constants.ts` | Rule file names: `AGENTS.md`, `.rules`, `CLAUDE.md` |
+| `types.ts` | `RuleFile`, `InjectionResult`, `RuleFileScope` |
+
+## RULE FILE DISCOVERY
+
+Priority (closest → farthest from target file):
+1. Same directory as target file
+2. Parent directories up to project root
+3. Project root itself
+
+Same-distance tie: all injected. Per-session dedup prevents re-injection.
+
+## TRUNCATION
+
+Uses `DynamicTruncator` — adapts injection size based on model context window (1M context models get full content, smaller models get truncated summaries).
diff --git a/src/tools/background-task/AGENTS.md b/src/tools/background-task/AGENTS.md
new file mode 100644
index 000000000..ee7404d94
--- /dev/null
+++ b/src/tools/background-task/AGENTS.md
@@ -0,0 +1,53 @@
+# src/tools/background-task/ — Background Task Tool Wrappers
+
+**Generated:** 2026-02-19
+
+## OVERVIEW
+
+18 files. Tool-layer wrappers for `background_output` and `background_cancel`. Does NOT implement the background execution engine — that lives in `src/features/background-agent/`. This directory provides the LLM-facing tool interface.
+
+## THREE TOOLS
+
+| Tool | Factory | Purpose |
+|------|---------|---------|
+| `background_output` | `createBackgroundOutput` | Get results from a running/completed background task |
+| `background_cancel` | `createBackgroundCancel` | Cancel running task(s) |
+| `createBackgroundTask` | internal | Shared factory used by both |
+
+## KEY FILES
+
+| File | Purpose |
+|------|---------|
+| `create-background-output.ts` | `background_output` tool: fetch task results by task_id |
+| `create-background-cancel.ts` | `background_cancel` tool: cancel by taskId or all=true |
+| `create-background-task.ts` | Shared tool factory with common params |
+| `clients.ts` | Client interfaces for background output and cancel |
+| `session-messages.ts` | Fetch session messages from OpenCode |
+| `full-session-format.ts` | Format full session output (messages, thinking blocks) |
+| `task-result-format.ts` | Format task result for LLM consumption |
+| `task-status-format.ts` | Format task status (running/completed/error) |
+| `message-dir.ts` | Temp directory for message exchange |
+| `truncate-text.ts` | Truncate large output to fit context |
+| `time-format.ts` | Human-readable duration formatting |
+| `delay.ts` | Polling delay utility |
+| `types.ts` | `BackgroundTaskOptions`, result/status types |
+| `constants.ts` | Timeout defaults, polling intervals |
+
+## BACKGROUND OUTPUT MODES
+
+```
+background_output(task_id, block=false)  → check current status/result
+background_output(task_id, block=true)   → wait until complete (timeout default: 120s)
+background_output(task_id, full_session=true) → return full session transcript
+background_output(task_id, message_limit=N) → last N messages only
+background_output(task_id, include_thinking=true) → include thinking blocks
+```
+
+## RELATIONSHIP TO BACKGROUND ENGINE
+
+```
+tools/background-task/  ← LLM tool interface
+features/background-agent/  ← execution engine (BackgroundManager)
+```
+
+`createBackgroundOutput` queries `BackgroundManager.getTask(task_id)` — it does not manage task state.
diff --git a/src/tools/call-omo-agent/AGENTS.md b/src/tools/call-omo-agent/AGENTS.md
new file mode 100644
index 000000000..8fc796fe7
--- /dev/null
+++ b/src/tools/call-omo-agent/AGENTS.md
@@ -0,0 +1,51 @@
+# src/tools/call-omo-agent/ — Direct Agent Invocation Tool
+
+**Generated:** 2026-02-19
+
+## OVERVIEW
+
+23 files. The `call_omo_agent` tool — direct invocation of named agents (explore, librarian only). Distinct from `delegate-task`: no category system, no skill loading, no model selection. Fixed agent set, same execution modes (background/sync).
+
+## DISTINCTION FROM delegate-task
+
+| Aspect | `call_omo_agent` | `delegate-task` (`task`) |
+|--------|-----------------|--------------------------|
+| Agent selection | Named agent (explore/librarian) | Category or subagent_type |
+| Skill loading | None | `load_skills[]` supported |
+| Model selection | From agent's fallback chain | From category config |
+| Use case | Quick contextual grep | Full delegation with skills |
+
+## ALLOWED AGENTS
+
+Only `explore` and `librarian` — enforced via `ALLOWED_AGENTS` constant in `constants.ts`. Case-insensitive validation.
+
+## EXECUTION MODES
+
+Same two modes as delegate-task:
+
+| Mode | File | Description |
+|------|------|-------------|
+| **Background** | `background-agent-executor.ts` | Async via `BackgroundManager` |
+| **Sync** | `sync-executor.ts` | Create session → wait for idle → return result |
+
+## KEY FILES
+
+| File | Purpose |
+|------|---------|
+| `tools.ts` | `createCallOmoAgent()` factory — validates agent, routes to executor |
+| `background-executor.ts` | Routes to background or sync based on `run_in_background` |
+| `background-agent-executor.ts` | Launch via `BackgroundManager.launch()` |
+| `sync-executor.ts` | Synchronous session: create → send prompt → poll → fetch result |
+| `session-creator.ts` | Create OpenCode session for sync execution |
+| `subagent-session-creator.ts` | Create session with agent-specific config |
+| `subagent-session-prompter.ts` | Inject prompt into session |
+| `completion-poller.ts` | Poll until session idle |
+| `session-completion-poller.ts` | Session-specific completion check |
+| `session-message-output-extractor.ts` | Extract last assistant message as result |
+| `message-processor.ts` | Process raw message content |
+| `message-dir.ts` + `message-storage-directory.ts` | Temp storage for message exchange |
+| `types.ts` | `CallOmoAgentArgs`, `AllowedAgentType`, `ToolContextWithMetadata` |
+
+## SESSION CONTINUATION
+
+Pass `session_id` to resume an existing session rather than create a new one — handled in both executors.
diff --git a/src/tools/lsp/AGENTS.md b/src/tools/lsp/AGENTS.md
new file mode 100644
index 000000000..9045bcf88
--- /dev/null
+++ b/src/tools/lsp/AGENTS.md
@@ -0,0 +1,70 @@
+# src/tools/lsp/ — LSP Tool Implementations
+
+**Generated:** 2026-02-19
+
+## OVERVIEW
+
+32 files. Full LSP (Language Server Protocol) client stack exposed as 6 tools. Custom implementation that manages server processes, opens files, and forwards requests — does NOT delegate to OpenCode's built-in LSP.
+
+## TOOL EXPOSURE
+
+| Tool | File | What It Does |
+|------|------|--------------|
+| `lsp_goto_definition` | `goto-definition-tool.ts` | Jump to symbol definition |
+| `lsp_find_references` | `find-references-tool.ts` | All usages of a symbol |
+| `lsp_symbols` | `symbols-tool.ts` | Document outline or workspace symbol search |
+| `lsp_diagnostics` | `diagnostics-tool.ts` | Errors/warnings from language server |
+| `lsp_prepare_rename` | `rename-tools.ts` | Validate rename before applying |
+| `lsp_rename` | `rename-tools.ts` | Apply safe rename across workspace |
+
+All 6 are direct `ToolDefinition` objects (not factory functions) — registered directly in `tool-registry.ts`.
+
+## ARCHITECTURE
+
+```
+tools.ts (6 ToolDefinition exports)
+  ↓ uses
+LspClientWrapper (lsp-client-wrapper.ts)
+  ↓ wraps
+LSPClient (lsp-client.ts) extends LSPClientConnection (lsp-client-connection.ts)
+  ↓ communicates via
+LSPClientTransport (lsp-client-transport.ts)
+  ↓ talks to
+LSPProcess (lsp-process.ts) — spawns server binary
+```
+
+## KEY FILES
+
+| File | Purpose |
+|------|---------|
+| `lsp-client-wrapper.ts` | High-level entry: resolves server, opens file, runs request |
+| `lsp-client.ts` | `LSPClient` — file tracking, document sync (`didOpen`/`didChange`) |
+| `lsp-client-connection.ts` | JSON-RPC request/response/notification layer |
+| `lsp-client-transport.ts` | stdin/stdout byte-stream framing |
+| `lsp-process.ts` | Spawn + cleanup of LSP server process |
+| `lsp-manager-process-cleanup.ts` | Reap orphan LSP processes on exit |
+| `lsp-manager-temp-directory-cleanup.ts` | Clean temp dirs used by some servers |
+| `server-definitions.ts` | 40+ builtin servers synced from OpenCode's `server.ts` |
+| `server-config-loader.ts` | Load custom server config from `.opencode/lsp.json` |
+| `server-resolution.ts` | Resolve which server handles a file extension |
+| `server-installation.ts` | Detect missing binaries, surface install hints |
+| `language-mappings.ts` | Extension → language ID mapping |
+| `lsp-formatters.ts` | Format LSP responses into human-readable strings |
+| `workspace-edit.ts` | Apply `WorkspaceEdit` results to disk (for rename) |
+| `types.ts` | `LSPServerConfig`, `Position`, `Range`, `Location`, `Diagnostic` etc. |
+
+## SERVER RESOLUTION
+
+```
+file.ts → extension (.ts) → language-mappings → server ID (typescript)
+  → server-resolution: check user config (.opencode/lsp.json) → fall back to server-definitions.ts
+  → server-installation: verify binary exists (warn with install hint if not)
+  → LSPProcess.spawn(command[])
+```
+
+## NOTES
+
+- File must be opened via `didOpen` before any LSP request — `LSPClient.openFile()` handles this
+- 1s delay after `didOpen` for server initialization before sending requests
+- `lsp_servers` tool was removed — duplicates OpenCode's built-in `LspServers` tool
+- Synced with OpenCode's `server.ts` — when adding servers, check upstream first