diff --git a/PROPOSED_CODE_FILE_REORGANIZATION_PLAN.md b/PROPOSED_CODE_FILE_REORGANIZATION_PLAN.md
new file mode 100644
index 0000000..4ce9599
--- /dev/null
+++ b/PROPOSED_CODE_FILE_REORGANIZATION_PLAN.md
@@ -0,0 +1,316 @@
+# Proposed Code File Reorganization Plan
+
+## Executive Summary
+
+After reading every source file in the project, analyzing all import graphs, and understanding how each module fits into the architecture, my assessment is: **the project is already reasonably well-organized**. The mixin-based decomposition of the handler, the dashboard's `components/utils/lib` split, and the test structure that mirrors source all reflect sound engineering.
+
+That said, there is one clear structural problem and a few smaller wins. This plan proposes **surgical, high-value changes** rather than a gratuitous restructure. The guiding principle: every change must make it easier for a developer (or agent) to find things and understand the architecture.
+
+---
+
+## Current Structure (Annotated)
+
+```
+amc/
+  amc_server/               # Python backend (2,571 LOC)
+    __init__.py              # Package init, exports main
+    server.py                # Server startup/shutdown (38 LOC)
+    handler.py               # Handler class composed from mixins (31 LOC)
+    context.py               # ** PROBLEM ** All config, constants, caches, locks, auth (121 LOC)
+    logging_utils.py         # Logging + signal handlers (31 LOC)
+    mixins/                  # Handler mixins (one per concern)
+      __init__.py            # Package comment (1 LOC)
+      http.py                # HTTP routing, static file serving (173 LOC)
+      state.py               # State aggregation, SSE, session collection, cleanup (440 LOC)
+      conversation.py        # Conversation parsing for Claude/Codex (278 LOC)
+      control.py             # Session dismiss/respond, Zellij pane injection (295 LOC)
+      discovery.py           # Codex session discovery, pane matching (347 LOC)
+      parsing.py             # JSONL parsing, context usage extraction (274 LOC)
+      skills.py              # Skill enumeration for autocomplete (184 LOC)
+      spawn.py               # Agent spawning in Zellij tabs (358 LOC)
+
+  dashboard/                 # Preact frontend (2,564 LOC)
+    index.html               # Entry HTML with Tailwind config
+    main.js                  # App mount point (7 LOC)
+    styles.css               # Custom styles
+    lib/                     # Third-party/shared
+      preact.js              # Preact re-exports
+      markdown.js            # Markdown rendering + syntax highlighting (159 LOC)
+    utils/                   # Pure utility functions
+      api.js                 # API constants + fetch helpers (39 LOC)
+      formatting.js          # Time/token formatting (66 LOC)
+      status.js              # Status metadata + session grouping (79 LOC)
+      autocomplete.js        # Autocomplete trigger detection (48 LOC)
+    components/              # UI components
+      App.js                 # Root component (616 LOC)
+      Sidebar.js             # Project nav sidebar (102 LOC)
+      SessionCard.js         # Session card (176 LOC)
+      Modal.js               # Full-screen modal wrapper (79 LOC)
+      ChatMessages.js        # Message list (39 LOC)
+      MessageBubble.js       # Individual message (54 LOC)
+      QuestionBlock.js       # AskUserQuestion UI (228 LOC)
+      SimpleInput.js         # Freeform text input (228 LOC)
+      OptionButton.js        # Option button (24 LOC)
+      AgentActivityIndicator.js  # Turn timer (115 LOC)
+      SpawnModal.js          # Spawn dropdown (241 LOC)
+      Toast.js               # Toast notifications (125 LOC)
+      EmptyState.js          # Empty state (18 LOC)
+      Header.js              # ** DEAD CODE ** (58 LOC, zero imports)
+      SessionGroup.js        # ** DEAD CODE ** (56 LOC, zero imports)
+
+  bin/                       # Shell/Python scripts
+    amc                      # Launcher (start/stop/status)
+    amc-hook                 # Hook script (standalone, writes session state)
+    amc-server               # Server launch script
+    amc-server-restart       # Server restart helper
+
+  tests/                     # Test suite (mirrors mixin structure)
+    test_context.py          # Context tests
+    test_control.py          # Control mixin tests
+    test_conversation.py     # Conversation parsing tests
+    test_conversation_mtime.py  # Conversation mtime tests
+    test_discovery.py        # Discovery mixin tests
+    test_hook.py             # Hook script tests
+    test_http.py             # HTTP mixin tests
+    test_parsing.py          # Parsing mixin tests
+    test_skills.py           # Skills mixin tests
+    test_spawn.py            # Spawn mixin tests
+    test_state.py            # State mixin tests
+    test_zellij_metadata.py  # Zellij metadata tests
+    e2e/                     # End-to-end tests
+      __init__.py
+      test_skills_endpoint.py
+      test_autocomplete_workflow.js
+    e2e_spawn.sh             # Spawn E2E script
+```
+
+---
+
+## Proposed Changes
+
+### Change 1: Split `context.py` into Focused Modules (HIGH VALUE)
+
+**Problem:** `context.py` is the classic "junk drawer" module. It contains:
+- Path constants for the server, Zellij, Claude, and Codex
+- Server configuration (port, timeouts)
+- 5 independent caches with their own size limits
+- 2 threading locks for unrelated concerns
+- Auth token generation/validation
+- Zellij binary resolution
+- Spawn-related config
+- Background thread management for projects cache
+
+Every mixin imports from it, but each only needs a subset. When a developer asks "where is the spawn rate limit configured?", they have to scan through an unrelated grab-bag of constants. When they ask "where's the Codex transcript cache?", same problem.
+
+**Proposed split:**
+
+```
+amc_server/
+  config.py          # Server-level constants: PORT, DATA_DIR, SESSIONS_DIR, EVENTS_DIR,
+                     #   DASHBOARD_DIR, PROJECT_DIR, STALE_EVENT_AGE, STALE_STARTING_AGE
+                     # These are the "universal" constants every module might need.
+
+  zellij.py          # Zellij integration: ZELLIJ_BIN resolution, ZELLIJ_PLUGIN path,
+                     #   ZELLIJ_SESSION, _zellij_cache (sessions cache + expiry)
+                     # Rationale: All Zellij-specific constants and helpers in one place.
+                     # Any developer working on Zellij integration knows exactly where to look.
+
+  agents.py          # Agent-specific paths and caches:
+                     #   CLAUDE_PROJECTS_DIR, CODEX_SESSIONS_DIR, CODEX_ACTIVE_WINDOW,
+                     #   _codex_pane_cache, _codex_transcript_cache, _CODEX_CACHE_MAX,
+                     #   _context_usage_cache, _CONTEXT_CACHE_MAX,
+                     #   _dismissed_codex_ids, _DISMISSED_MAX
+                     # Rationale: Agent data source configuration and caches that are only
+                     # relevant to discovery/parsing mixins, not the whole server.
+
+  auth.py            # Auth token: generate_auth_token(), validate_auth_token(), _auth_token
+                     # Rationale: Security-sensitive code in its own module. Small, but
+                     # architecturally clean separation from general config.
+
+  spawn_config.py    # Spawn feature config:
+                     #   PROJECTS_DIR, PENDING_SPAWNS_DIR, PENDING_SPAWN_TTL,
+                     #   _spawn_lock, _spawn_timestamps, SPAWN_COOLDOWN_SEC
+                     # + start_projects_watcher() (background refresh thread)
+                     # Rationale: Spawn feature has its own set of constants, lock, and
+                     # background thread. Currently scattered between context.py and spawn.py.
+                     # Consolidating makes the spawn feature self-contained.
+```
+
+Kept from current structure (unchanged):
+- `_state_lock` moves to `config.py` (it's a server-level concern)
+
+**Import changes required:**
+
+| File | Current import from `context` | New import from |
+|------|------|------|
+| `server.py` | `DATA_DIR, PORT, generate_auth_token, start_projects_watcher` | `config.DATA_DIR, config.PORT`, `auth.generate_auth_token`, `spawn_config.start_projects_watcher` |
+| `handler.py` | (none, uses mixins) | (unchanged) |
+| `mixins/http.py` | `DASHBOARD_DIR`, `ctx._auth_token` | `config.DASHBOARD_DIR`, `auth._auth_token` |
+| `mixins/state.py` | `EVENTS_DIR, SESSIONS_DIR, STALE_*, ZELLIJ_BIN, _state_lock, _zellij_cache` | `config.*`, `zellij.ZELLIJ_BIN, zellij._zellij_cache` |
+| `mixins/conversation.py` | `EVENTS_DIR` | `config.EVENTS_DIR` |
+| `mixins/control.py` | `SESSIONS_DIR, ZELLIJ_BIN, ZELLIJ_PLUGIN, _DISMISSED_MAX, _dismissed_codex_ids` | `config.SESSIONS_DIR`, `zellij.*`, `agents._DISMISSED_MAX, agents._dismissed_codex_ids` |
+| `mixins/discovery.py` | `CODEX_*, PENDING_SPAWNS_DIR, SESSIONS_DIR, _codex_*` | `agents.*`, `spawn_config.PENDING_SPAWNS_DIR`, `config.SESSIONS_DIR` |
+| `mixins/parsing.py` | `CLAUDE_PROJECTS_DIR, CODEX_SESSIONS_DIR, _*_cache, _*_MAX` | `agents.*` |
+| `mixins/spawn.py` | `PENDING_SPAWNS_DIR, PROJECTS_DIR, SESSIONS_DIR, ZELLIJ_*, _spawn_*, validate_auth_token` | `spawn_config.*`, `config.SESSIONS_DIR`, `zellij.*`, `auth.validate_auth_token` |
+
+**Why this is the right split:**
+
+1. **By domain, not by size.** Each new module groups constants + caches + helpers that serve one architectural concern. A developer working on Zellij integration opens `zellij.py`. Working on Codex discovery? `agents.py`. Spawn feature? `spawn_config.py`.
+
+2. **No circular imports.** The dependency graph is DAG: `config.py` is a leaf (imported by everything, imports nothing from `amc_server`). `zellij.py`, `agents.py`, `auth.py`, `spawn_config.py` import only from `config.py` (if at all). Mixins import from these.
+
+3. **No behavioral change.** Module-level caches and singletons work the same way whether they're in one file or five.
+
+---
+
+### Change 2: Delete Dead Dashboard Components (LOW EFFORT, HIGH CLARITY)
+
+**Problem:** `Header.js` (58 LOC) and `SessionGroup.js` (56 LOC) are completely unused. Zero imports anywhere in the codebase. They were replaced by the current Sidebar + grid layout but never cleaned up.
+
+**Action:** Delete both files.
+
+**Import changes required:** None (nothing imports them).
+
+**Rationale:** Dead code is noise. Anyone exploring the `components/` directory would reasonably assume these are active and try to understand how they fit. Removing them prevents that confusion.
+
+---
+
+### Change 3: No Changes to Dashboard Structure
+
+The dashboard is already well-organized:
+
+- `components/` - All React-like components
+- `utils/` - Pure utility functions (formatting, API, status, autocomplete)
+- `lib/` - Third-party wrappers (Preact, markdown rendering)
+
+This is a standard and intuitive layout. The `components/` directory has 13 files (15 before dead code removal), which is manageable. Creating sub-directories (e.g., `components/session/`, `components/layout/`) would add nesting without meaningful benefit at this scale.
+
+---
+
+### Change 4: No Changes to `mixins/` Structure
+
+The mixin decomposition is the project's architectural backbone. Each mixin handles one concern:
+
+| Mixin | Responsibility |
+|-------|---------------|
+| `http.py` | HTTP routing, static file serving, CORS |
+| `state.py` | State aggregation, SSE streaming, session collection |
+| `conversation.py` | Conversation history parsing (Claude + Codex JSONL) |
+| `control.py` | Session dismiss/respond, Zellij pane injection |
+| `discovery.py` | Codex session auto-discovery, pane matching |
+| `parsing.py` | JSONL tail reading, context usage extraction, caching |
+| `skills.py` | Skill enumeration for Claude/Codex autocomplete |
+| `spawn.py` | Agent spawning in Zellij tabs |
+
+All are 170-440 lines, which is reasonable. The largest (`state.py` at 440 lines) could theoretically be split, but its methods are tightly coupled around session collection. Splitting would create artificial seams.
+
+---
+
+### Change 5: No Changes to `tests/` Structure
+
+Tests already mirror the source structure (`test_state.py` tests `mixins/state.py`, etc.). This is the correct pattern.
+
+**One consideration:** After splitting `context.py`, `test_context.py` may need updates to import from the new module locations. The test file is small (755 bytes) and covers basic context constants, so the update would be trivial.
+
+---
+
+### Change 6: No Changes to `bin/` Scripts
+
+The `amc-hook` script intentionally duplicates `DATA_DIR`, `SESSIONS_DIR`, `EVENTS_DIR` from `context.py`. This is correct: the hook runs as a standalone process launched by Claude Code, not as part of the server. It must be self-contained with zero dependencies on the server package. Sharing code would create a fragile coupling.
+
+---
+
+## What I Explicitly Decided NOT to Do
+
+1. **Not creating a `src/` directory.** The project root is clean. Adding `src/` would be an extra nesting level with no benefit.
+
+2. **Not splitting any mixins.** `state.py` (440 LOC) and `spawn.py` (358 LOC) are the largest, but their methods are cohesive. Splitting would scatter related logic across files.
+
+3. **Not merging small files.** `EmptyState.js` (18 LOC), `OptionButton.js` (24 LOC), and `ChatMessages.js` (39 LOC) are tiny but each has a clear purpose and is imported independently. Merging them would violate component-per-file convention.
+
+4. **Not reorganizing dashboard components into sub-folders.** With 13 components, flat is fine. Sub-folders like `components/session/` and `components/layout/` become necessary at ~25+ components.
+
+5. **Not consolidating `api.js` + `formatting.js` + `status.js` + `autocomplete.js`.** Each is focused and independently imported. A combined `utils.js` would be a grab-bag (the exact problem we're fixing in `context.py`).
+
+6. **Not moving `markdown.js` out of `lib/`.** It uses third-party dependencies and provides rendering utilities. `lib/` is the correct location.
+
+---
+
+## Proposed Final Structure
+
+```
+amc/
+  amc_server/
+    __init__.py              # (unchanged)
+    server.py                # (updated imports)
+    handler.py               # (unchanged)
+    config.py                # NEW: Server constants, DATA_DIR, SESSIONS_DIR, EVENTS_DIR, PORT, etc.
+    zellij.py                # NEW: Zellij binary resolution, ZELLIJ_PLUGIN, ZELLIJ_SESSION, cache
+    agents.py                # NEW: Agent paths (Claude/Codex), transcript caches, dismissed cache
+    auth.py                  # NEW: Auth token generation/validation
+    spawn_config.py          # NEW: Spawn constants, locks, rate limiting, projects watcher
+    logging_utils.py         # (unchanged)
+    mixins/                  # (unchanged structure, updated imports)
+      __init__.py
+      http.py
+      state.py
+      conversation.py
+      control.py
+      discovery.py
+      parsing.py
+      skills.py
+      spawn.py
+
+  dashboard/
+    index.html               # (unchanged)
+    main.js                  # (unchanged)
+    styles.css               # (unchanged)
+    lib/
+      preact.js              # (unchanged)
+      markdown.js            # (unchanged)
+    utils/
+      api.js                 # (unchanged)
+      formatting.js          # (unchanged)
+      status.js              # (unchanged)
+      autocomplete.js        # (unchanged)
+    components/
+      App.js                 # (unchanged)
+      Sidebar.js             # (unchanged)
+      SessionCard.js         # (unchanged)
+      Modal.js               # (unchanged)
+      ChatMessages.js        # (unchanged)
+      MessageBubble.js       # (unchanged)
+      QuestionBlock.js       # (unchanged)
+      SimpleInput.js         # (unchanged)
+      OptionButton.js        # (unchanged)
+      AgentActivityIndicator.js  # (unchanged)
+      SpawnModal.js          # (unchanged)
+      Toast.js               # (unchanged)
+      EmptyState.js          # (unchanged)
+      [DELETED] Header.js
+      [DELETED] SessionGroup.js
+
+  bin/                       # (unchanged)
+  tests/                     # (minor import updates in test_context.py)
+```
+
+---
+
+## Implementation Order
+
+1. **Delete dead dashboard components** (`Header.js`, `SessionGroup.js`) - zero risk, instant clarity
+2. **Create new Python modules** (`config.py`, `zellij.py`, `agents.py`, `auth.py`, `spawn_config.py`) with the correct constants/functions
+3. **Update all mixin imports** to use new module locations
+4. **Update `server.py`** imports
+5. **Delete `context.py`**
+6. **Run full test suite** to verify nothing broke
+7. **Update `test_context.py`** if needed
+
+---
+
+## Risk Assessment
+
+- **Risk of breaking imports:** MEDIUM. There are many import statements to update across 8 mixin files + `server.py`. Mitigated by running the full test suite after changes.
+- **Risk of circular imports:** LOW. The new modules form a clean DAG (config <- zellij/agents/auth/spawn_config <- mixins).
+- **Risk to `bin/amc-hook`:** NONE. The hook is standalone and doesn't import from `amc_server`.
+- **Risk to dashboard:** NONE for dead code deletion. Zero imports to either file.
diff --git a/amc_server/agents.py b/amc_server/agents.py
new file mode 100644
index 0000000..baa5a98
--- /dev/null
+++ b/amc_server/agents.py
@@ -0,0 +1,28 @@
+"""Agent-specific paths, caches, and constants for Claude/Codex discovery."""
+
+from pathlib import Path
+
+# Claude Code conversation directory
+CLAUDE_PROJECTS_DIR = Path.home() / ".claude" / "projects"
+
+# Codex conversation directory
+CODEX_SESSIONS_DIR = Path.home() / ".codex" / "sessions"
+
+# Only discover recently-active Codex sessions (10 minutes)
+CODEX_ACTIVE_WINDOW = 600
+
+# Cache for Codex pane info (avoid running pgrep/ps/lsof on every request)
+_codex_pane_cache = {"pid_info": {}, "cwd_map": {}, "expires": 0}
+
+# Cache for parsed context usage by transcript file path + mtime/size
+_context_usage_cache = {}
+_CONTEXT_CACHE_MAX = 100
+
+# Cache mapping Codex session IDs to transcript paths (or None when missing)
+_codex_transcript_cache = {}
+_CODEX_CACHE_MAX = 200
+
+# Codex sessions dismissed during this server lifetime (prevents re-discovery)
+# Uses dict (not set) for O(1) lookup + FIFO eviction via insertion order (Python 3.7+)
+_dismissed_codex_ids = {}
+_DISMISSED_MAX = 500
diff --git a/amc_server/auth.py b/amc_server/auth.py
new file mode 100644
index 0000000..c1f557f
--- /dev/null
+++ b/amc_server/auth.py
@@ -0,0 +1,18 @@
+"""Auth token generation and validation for spawn endpoint security."""
+
+import secrets
+
+# Auth token for spawn endpoint
+_auth_token: str = ''
+
+
+def generate_auth_token():
+    """Generate a one-time auth token for this server instance."""
+    global _auth_token
+    _auth_token = secrets.token_urlsafe(32)
+    return _auth_token
+
+
+def validate_auth_token(request_token: str) -> bool:
+    """Validate the Authorization header token."""
+    return request_token == f'Bearer {_auth_token}'
diff --git a/amc_server/config.py b/amc_server/config.py
new file mode 100644
index 0000000..c5a7bbe
--- /dev/null
+++ b/amc_server/config.py
@@ -0,0 +1,20 @@
+"""Server-level constants: paths, port, timeouts, state lock."""
+
+import threading
+from pathlib import Path
+
+# Runtime data lives in XDG data dir
+DATA_DIR = Path.home() / ".local" / "share" / "amc"
+SESSIONS_DIR = DATA_DIR / "sessions"
+EVENTS_DIR = DATA_DIR / "events"
+
+# Source files live in project directory (relative to this module)
+PROJECT_DIR = Path(__file__).resolve().parent.parent
+DASHBOARD_DIR = PROJECT_DIR / "dashboard"
+
+PORT = 7400
+STALE_EVENT_AGE = 86400  # 24 hours in seconds
+STALE_STARTING_AGE = 3600  # 1 hour - sessions stuck in "starting" are orphans
+
+# Serialize state collection because it mutates session files/caches.
+_state_lock = threading.Lock()
diff --git a/amc_server/context.py b/amc_server/context.py
deleted file mode 100644
index a75c083..0000000
--- a/amc_server/context.py
+++ /dev/null
@@ -1,117 +0,0 @@
-import secrets
-import shutil
-from pathlib import Path
-import threading
-
-# Claude Code conversation directory
-CLAUDE_PROJECTS_DIR = Path.home() / ".claude" / "projects"
-
-# Codex conversation directory
-CODEX_SESSIONS_DIR = Path.home() / ".codex" / "sessions"
-
-# Plugin path for zellij-send-keys
-ZELLIJ_PLUGIN = Path.home() / ".config" / "zellij" / "plugins" / "zellij-send-keys.wasm"
-
-
-def _resolve_zellij_bin():
-    """Resolve zellij binary even when PATH is minimal (eg launchctl)."""
-    from_path = shutil.which("zellij")
-    if from_path:
-        return from_path
-
-    common_paths = (
-        "/opt/homebrew/bin/zellij",  # Apple Silicon Homebrew
-        "/usr/local/bin/zellij",     # Intel Homebrew
-        "/usr/bin/zellij",
-    )
-    for candidate in common_paths:
-        p = Path(candidate)
-        if p.exists() and p.is_file():
-            return str(p)
-    return "zellij"  # Fallback for explicit error reporting by subprocess
-
-
-ZELLIJ_BIN = _resolve_zellij_bin()
-
-# Runtime data lives in XDG data dir
-DATA_DIR = Path.home() / ".local" / "share" / "amc"
-SESSIONS_DIR = DATA_DIR / "sessions"
-EVENTS_DIR = DATA_DIR / "events"
-
-# Source files live in project directory (relative to this module)
-PROJECT_DIR = Path(__file__).resolve().parent.parent
-DASHBOARD_DIR = PROJECT_DIR / "dashboard"
-
-PORT = 7400
-STALE_EVENT_AGE = 86400  # 24 hours in seconds
-STALE_STARTING_AGE = 3600  # 1 hour - sessions stuck in "starting" are orphans
-CODEX_ACTIVE_WINDOW = 600  # 10 minutes - only discover recently-active Codex sessions
-
-# Cache for Zellij session list (avoid calling zellij on every request)
-_zellij_cache = {"sessions": None, "expires": 0}
-
-# Cache for Codex pane info (avoid running pgrep/ps/lsof on every request)
-_codex_pane_cache = {"pid_info": {}, "cwd_map": {}, "expires": 0}
-
-# Cache for parsed context usage by transcript file path + mtime/size
-# Limited to prevent unbounded memory growth
-_context_usage_cache = {}
-_CONTEXT_CACHE_MAX = 100
-
-# Cache mapping Codex session IDs to transcript paths (or None when missing)
-_codex_transcript_cache = {}
-_CODEX_CACHE_MAX = 200
-
-# Codex sessions dismissed during this server lifetime (prevents re-discovery)
-# Uses dict (not set) for O(1) lookup + FIFO eviction via insertion order (Python 3.7+)
-_dismissed_codex_ids = {}
-_DISMISSED_MAX = 500
-
-# Serialize state collection because it mutates session files/caches.
-_state_lock = threading.Lock()
-
-# Projects directory for spawning agents
-PROJECTS_DIR = Path.home() / 'projects'
-
-# Default Zellij session for spawning
-ZELLIJ_SESSION = 'infra'
-
-# Lock for serializing spawn operations (prevents Zellij race conditions)
-_spawn_lock = threading.Lock()
-
-# Rate limiting: track last spawn time per project (prevents spam)
-_spawn_timestamps: dict[str, float] = {}
-SPAWN_COOLDOWN_SEC = 10.0
-
-# Auth token for spawn endpoint
-_auth_token: str = ''
-
-
-def generate_auth_token():
-    """Generate a one-time auth token for this server instance."""
-    global _auth_token
-    _auth_token = secrets.token_urlsafe(32)
-    return _auth_token
-
-
-def validate_auth_token(request_token: str) -> bool:
-    """Validate the Authorization header token."""
-    return request_token == f'Bearer {_auth_token}'
-
-
-def start_projects_watcher():
-    """Start background thread to refresh projects cache every 5 minutes."""
-    import logging
-    from amc_server.mixins.spawn import load_projects_cache
-
-    def _watch_loop():
-        import time
-        while True:
-            try:
-                time.sleep(300)
-                load_projects_cache()
-            except Exception:
-                logging.exception('Projects cache refresh failed')
-
-    thread = threading.Thread(target=_watch_loop, daemon=True)
-    thread.start()
diff --git a/amc_server/mixins/control.py b/amc_server/mixins/control.py
index 65dc28a..8595070 100644
--- a/amc_server/mixins/control.py
+++ b/amc_server/mixins/control.py
@@ -3,7 +3,9 @@ import os
 import subprocess
 import time
 
-from amc_server.context import SESSIONS_DIR, ZELLIJ_BIN, ZELLIJ_PLUGIN, _DISMISSED_MAX, _dismissed_codex_ids
+from amc_server.agents import _DISMISSED_MAX, _dismissed_codex_ids
+from amc_server.config import SESSIONS_DIR
+from amc_server.zellij import ZELLIJ_BIN, ZELLIJ_PLUGIN
 from amc_server.logging_utils import LOGGER
 
 
diff --git a/amc_server/mixins/conversation.py b/amc_server/mixins/conversation.py
index b7c1637..b6e01e9 100644
--- a/amc_server/mixins/conversation.py
+++ b/amc_server/mixins/conversation.py
@@ -1,7 +1,7 @@
 import json
 import os
 
-from amc_server.context import EVENTS_DIR
+from amc_server.config import EVENTS_DIR
 
 
 class ConversationMixin:
diff --git a/amc_server/mixins/discovery.py b/amc_server/mixins/discovery.py
index 31ee280..3aa9bb5 100644
--- a/amc_server/mixins/discovery.py
+++ b/amc_server/mixins/discovery.py
@@ -5,18 +5,99 @@ import subprocess
 import time
 from datetime import datetime, timezone
 
-from amc_server.context import (
+from amc_server.agents import (
     CODEX_ACTIVE_WINDOW,
     CODEX_SESSIONS_DIR,
-    SESSIONS_DIR,
     _CODEX_CACHE_MAX,
     _codex_pane_cache,
     _codex_transcript_cache,
     _dismissed_codex_ids,
 )
+from amc_server.config import SESSIONS_DIR
+from amc_server.spawn_config import PENDING_SPAWNS_DIR
 from amc_server.logging_utils import LOGGER
 
 
+def _parse_session_timestamp(session_ts):
+    """Parse Codex session timestamp to Unix time. Returns None on failure."""
+    if not session_ts:
+        return None
+    try:
+        # Codex uses ISO format, possibly with Z suffix or +00:00
+        ts_str = session_ts.replace('Z', '+00:00')
+        dt = datetime.fromisoformat(ts_str)
+        return dt.timestamp()
+    except (ValueError, TypeError, AttributeError):
+        return None
+
+
+def _match_pending_spawn(session_cwd, session_start_ts):
+    """Match a Codex session to a pending spawn by CWD and timestamp.
+
+    Args:
+        session_cwd: The CWD of the Codex session
+        session_start_ts: The session's START timestamp (ISO string from Codex metadata)
+            IMPORTANT: Must be session start time, not file mtime, to avoid false
+            matches with pre-existing sessions that were recently active.
+
+    Returns:
+        spawn_id if matched (and deletes the pending file), None otherwise
+    """
+    if not PENDING_SPAWNS_DIR.exists():
+        return None
+
+    normalized_cwd = os.path.normpath(session_cwd) if session_cwd else ""
+    if not normalized_cwd:
+        return None
+
+    # Parse session start time - if we can't parse it, we can't safely match
+    session_start_unix = _parse_session_timestamp(session_start_ts)
+    if session_start_unix is None:
+        return None
+
+    try:
+        for pending_file in PENDING_SPAWNS_DIR.glob('*.json'):
+            try:
+                data = json.loads(pending_file.read_text())
+                if not isinstance(data, dict):
+                    continue
+
+                # Check agent type (only match codex to codex)
+                if data.get('agent_type') != 'codex':
+                    continue
+
+                # Check CWD match
+                pending_path = os.path.normpath(data.get('project_path', ''))
+                if normalized_cwd != pending_path:
+                    continue
+
+                # Check timing: session must have STARTED after spawn was initiated
+                # Using session start time (not mtime) prevents false matches with
+                # pre-existing sessions that happen to be recently active
+                spawn_ts = data.get('timestamp', 0)
+                if session_start_unix < spawn_ts:
+                    continue
+
+                # Match found - claim the spawn_id and delete the pending file
+                spawn_id = data.get('spawn_id')
+                try:
+                    pending_file.unlink()
+                except OSError:
+                    pass
+                LOGGER.info(
+                    'Matched Codex session (cwd=%s) to pending spawn_id=%s',
+                    session_cwd, spawn_id,
+                )
+                return spawn_id
+
+            except (json.JSONDecodeError, OSError):
+                continue
+    except OSError:
+        pass
+
+    return None
+
+
 class SessionDiscoveryMixin:
     def _discover_active_codex_sessions(self):
         """Find active Codex sessions and create/update session files with Zellij pane info."""
@@ -131,6 +212,13 @@ class SessionDiscoveryMixin:
                 session_ts = payload.get("timestamp", "")
                 last_event_at = datetime.fromtimestamp(mtime, tz=timezone.utc).isoformat()
 
+                # Check for spawn_id: preserve existing, or match to pending spawn
+                # Use session_ts (start time) not mtime to avoid false matches
+                # with pre-existing sessions that were recently active
+                spawn_id = existing.get("spawn_id")
+                if not spawn_id:
+                    spawn_id = _match_pending_spawn(cwd, session_ts)
+
                 session_data = {
                     "session_id": session_id,
                     "agent": "codex",
@@ -145,6 +233,8 @@ class SessionDiscoveryMixin:
                     "zellij_pane": zellij_pane or existing.get("zellij_pane", ""),
                     "transcript_path": str(jsonl_file),
                 }
+                if spawn_id:
+                    session_data["spawn_id"] = spawn_id
                 if context_usage:
                     session_data["context_usage"] = context_usage
                 elif existing.get("context_usage"):
diff --git a/amc_server/mixins/http.py b/amc_server/mixins/http.py
index 0a528ed..7ca092a 100644
--- a/amc_server/mixins/http.py
+++ b/amc_server/mixins/http.py
@@ -1,8 +1,8 @@
 import json
 import urllib.parse
 
-import amc_server.context as ctx
-from amc_server.context import DASHBOARD_DIR
+import amc_server.auth as auth
+from amc_server.config import DASHBOARD_DIR
 from amc_server.logging_utils import LOGGER
 
 
@@ -148,10 +148,10 @@ class HttpMixin:
             content_type = content_types.get(ext, "application/octet-stream")
 
             # Inject auth token into index.html for spawn endpoint security
-            if file_path == "index.html" and ctx._auth_token:
+            if file_path == "index.html" and auth._auth_token:
                 content = content.replace(
                     b"<!-- AMC_AUTH_TOKEN -->",
-                    f'<script>window.AMC_AUTH_TOKEN = "{ctx._auth_token}";</script>'.encode(),
+                    f'<script>window.AMC_AUTH_TOKEN = "{auth._auth_token}";</script>'.encode(),
                 )
 
             # No caching during development
diff --git a/amc_server/mixins/parsing.py b/amc_server/mixins/parsing.py
index 591dba8..0fdf9aa 100644
--- a/amc_server/mixins/parsing.py
+++ b/amc_server/mixins/parsing.py
@@ -2,7 +2,7 @@ import json
 import os
 from pathlib import Path
 
-from amc_server.context import (
+from amc_server.agents import (
     CLAUDE_PROJECTS_DIR,
     CODEX_SESSIONS_DIR,
     _CODEX_CACHE_MAX,
diff --git a/amc_server/mixins/skills.py b/amc_server/mixins/skills.py
index dc58bd4..c957f07 100644
--- a/amc_server/mixins/skills.py
+++ b/amc_server/mixins/skills.py
@@ -37,7 +37,7 @@ class SkillsMixin:
 
         Checks SKILL.md (canonical) first, then falls back to skill.md,
         prompt.md, README.md for description extraction. Parses YAML
-        frontmatter if present to extract the description field.
+        frontmatter if present to extract name and description fields.
 
         Returns:
             List of {name: str, description: str} dicts.
@@ -53,35 +53,41 @@ class SkillsMixin:
             if not skill_dir.is_dir() or skill_dir.name.startswith("."):
                 continue
 
-            description = ""
-            # Check files in priority order
+            meta = {"name": "", "description": ""}
+            # Check files in priority order, accumulating metadata
+            # (earlier files take precedence for each field)
             for md_name in ["SKILL.md", "skill.md", "prompt.md", "README.md"]:
                 md_file = skill_dir / md_name
                 if md_file.exists():
                     try:
                         content = md_file.read_text()
-                        description = self._extract_description(content)
-                        if description:
+                        parsed = self._parse_frontmatter(content)
+                        if not meta["name"] and parsed["name"]:
+                            meta["name"] = parsed["name"]
+                        if not meta["description"] and parsed["description"]:
+                            meta["description"] = parsed["description"]
+                        if meta["description"]:
                             break
                     except OSError:
                         pass
 
             skills.append({
-                "name": skill_dir.name,
-                "description": description or f"Skill: {skill_dir.name}",
+                "name": meta["name"] or skill_dir.name,
+                "description": meta["description"] or f"Skill: {skill_dir.name}",
             })
 
         return skills
 
-    def _extract_description(self, content: str) -> str:
-        """Extract description from markdown content.
+    def _parse_frontmatter(self, content: str) -> dict:
+        """Extract name and description from markdown YAML frontmatter.
 
-        Handles YAML frontmatter (looks for 'description:' field) and
-        falls back to first meaningful line after frontmatter.
+        Returns:
+            Dict with 'name' and 'description' keys (both str, may be empty).
         """
+        result = {"name": "", "description": ""}
         lines = content.splitlines()
         if not lines:
-            return ""
+            return result
 
         # Check for YAML frontmatter
         frontmatter_end = 0
@@ -91,33 +97,37 @@ class SkillsMixin:
                 if stripped == "---":
                     frontmatter_end = i + 1
                     break
-                # Look for description field in frontmatter
-                if stripped.startswith("description:"):
-                    # Extract value after colon
-                    desc = stripped[len("description:"):].strip()
-                    # Remove quotes if present
-                    if desc.startswith('"') and desc.endswith('"'):
-                        desc = desc[1:-1]
-                    elif desc.startswith("'") and desc.endswith("'"):
-                        desc = desc[1:-1]
-                    # Handle YAML multi-line indicators (>- or |-)
-                    if desc in (">-", "|-", ">", "|", ""):
-                        # Multi-line: read the next indented line
-                        if i + 1 < len(lines):
-                            next_line = lines[i + 1].strip()
-                            if next_line and not next_line.startswith("---"):
-                                return next_line[:100]
-                    elif desc:
-                        return desc[:100]
+                # Check each known frontmatter field
+                for field in ("name", "description"):
+                    if stripped.startswith(f"{field}:"):
+                        val = stripped[len(field) + 1:].strip()
+                        # Remove quotes if present
+                        if val.startswith('"') and val.endswith('"'):
+                            val = val[1:-1]
+                        elif val.startswith("'") and val.endswith("'"):
+                            val = val[1:-1]
+                        # Handle YAML multi-line indicators (>- or |-)
+                        if val in (">-", "|-", ">", "|", ""):
+                            if i + 1 < len(lines):
+                                next_line = lines[i + 1].strip()
+                                if next_line and not next_line.startswith("---"):
+                                    val = next_line
+                                else:
+                                    val = ""
+                            else:
+                                val = ""
+                        if val:
+                            result[field] = val[:100]
 
-        # Fall back to first meaningful line after frontmatter
-        for line in lines[frontmatter_end:]:
-            stripped = line.strip()
-            # Skip empty lines, headers, comments, and frontmatter delimiters
-            if stripped and not stripped.startswith("#") and not stripped.startswith("<!--") and stripped != "---":
-                return stripped[:100]
+        # Fall back to first meaningful line for description
+        if not result["description"]:
+            for line in lines[frontmatter_end:]:
+                stripped = line.strip()
+                if stripped and not stripped.startswith("#") and not stripped.startswith("<!--") and stripped != "---":
+                    result["description"] = stripped[:100]
+                    break
 
-        return ""
+        return result
 
     def _enumerate_codex_skills(self) -> list[dict]:
         """Enumerate Codex skills from cache and user directory.
@@ -161,19 +171,19 @@ class SkillsMixin:
                 if not skill_dir.is_dir() or skill_dir.name.startswith("."):
                     continue
 
-                description = ""
-                # Check SKILL.md for description
+                meta = {"name": "", "description": ""}
+                # Check SKILL.md for metadata
                 skill_md = skill_dir / "SKILL.md"
                 if skill_md.exists():
                     try:
                         content = skill_md.read_text()
-                        description = self._extract_description(content)
+                        meta = self._parse_frontmatter(content)
                     except OSError:
                         pass
 
                 skills.append({
-                    "name": skill_dir.name,
-                    "description": description or f"User skill: {skill_dir.name}",
+                    "name": meta["name"] or skill_dir.name,
+                    "description": meta["description"] or f"User skill: {skill_dir.name}",
                 })
 
         return skills
diff --git a/amc_server/mixins/spawn.py b/amc_server/mixins/spawn.py
index b974309..d7c70d1 100644
--- a/amc_server/mixins/spawn.py
+++ b/amc_server/mixins/spawn.py
@@ -5,13 +5,52 @@ import subprocess
 import time
 import uuid
 
-from amc_server.context import (
-    PROJECTS_DIR, SESSIONS_DIR, ZELLIJ_BIN, ZELLIJ_SESSION,
+from amc_server.auth import validate_auth_token
+from amc_server.config import SESSIONS_DIR
+from amc_server.spawn_config import (
+    PENDING_SPAWNS_DIR, PENDING_SPAWN_TTL,
+    PROJECTS_DIR,
     _spawn_lock, _spawn_timestamps, SPAWN_COOLDOWN_SEC,
-    validate_auth_token,
 )
+from amc_server.zellij import ZELLIJ_BIN, ZELLIJ_SESSION
 from amc_server.logging_utils import LOGGER
 
+
+def _write_pending_spawn(spawn_id, project_path, agent_type):
+    """Write a pending spawn record for later correlation by discovery.
+
+    This enables Codex session correlation since env vars don't propagate
+    through Zellij's pane spawn mechanism.
+    """
+    PENDING_SPAWNS_DIR.mkdir(parents=True, exist_ok=True)
+    pending_file = PENDING_SPAWNS_DIR / f'{spawn_id}.json'
+    data = {
+        'spawn_id': spawn_id,
+        'project_path': str(project_path),
+        'agent_type': agent_type,
+        'timestamp': time.time(),
+    }
+    try:
+        pending_file.write_text(json.dumps(data))
+    except OSError:
+        LOGGER.warning('Failed to write pending spawn file for %s', spawn_id)
+
+
+def _cleanup_stale_pending_spawns():
+    """Remove pending spawn files older than PENDING_SPAWN_TTL."""
+    if not PENDING_SPAWNS_DIR.exists():
+        return
+    now = time.time()
+    try:
+        for f in PENDING_SPAWNS_DIR.glob('*.json'):
+            try:
+                if now - f.stat().st_mtime > PENDING_SPAWN_TTL:
+                    f.unlink()
+            except OSError:
+                continue
+    except OSError:
+        pass
+
 # Agent commands (AC-8, AC-9: full autonomous permissions)
 AGENT_COMMANDS = {
     'claude': ['claude', '--dangerously-skip-permissions'],
@@ -215,6 +254,16 @@ class SpawnMixin:
 
     def _spawn_agent_in_project_tab(self, project, project_path, agent_type, spawn_id):
         """Spawn an agent in a project-named Zellij tab."""
+        # Clean up stale pending spawns opportunistically
+        _cleanup_stale_pending_spawns()
+
+        # For Codex, write pending spawn record before launching.
+        # Zellij doesn't propagate env vars to pane commands, so discovery
+        # will match the session to this record by CWD + timestamp.
+        # (Claude doesn't need this - amc-hook writes spawn_id directly)
+        if agent_type == 'codex':
+            _write_pending_spawn(spawn_id, project_path, agent_type)
+
         # Check session exists
         if not self._check_zellij_session_exists():
             return {
diff --git a/amc_server/server.py b/amc_server/server.py
index 73006cb..2ae9fc5 100644
--- a/amc_server/server.py
+++ b/amc_server/server.py
@@ -1,7 +1,9 @@
 import os
 from http.server import ThreadingHTTPServer
 
-from amc_server.context import DATA_DIR, PORT, generate_auth_token, start_projects_watcher
+from amc_server.auth import generate_auth_token
+from amc_server.config import DATA_DIR, PORT
+from amc_server.spawn_config import start_projects_watcher
 from amc_server.handler import AMCHandler
 from amc_server.logging_utils import LOGGER, configure_logging, install_signal_handlers
 from amc_server.mixins.spawn import load_projects_cache
diff --git a/amc_server/spawn_config.py b/amc_server/spawn_config.py
new file mode 100644
index 0000000..d409138
--- /dev/null
+++ b/amc_server/spawn_config.py
@@ -0,0 +1,40 @@
+"""Spawn feature config: paths, locks, rate limiting, projects watcher."""
+
+import threading
+from pathlib import Path
+
+from amc_server.config import DATA_DIR
+
+# Pending spawn registry
+PENDING_SPAWNS_DIR = DATA_DIR / "pending_spawns"
+
+# Pending spawn TTL: how long to keep unmatched spawn records (seconds)
+PENDING_SPAWN_TTL = 60
+
+# Projects directory for spawning agents
+PROJECTS_DIR = Path.home() / 'projects'
+
+# Lock for serializing spawn operations (prevents Zellij race conditions)
+_spawn_lock = threading.Lock()
+
+# Rate limiting: track last spawn time per project (prevents spam)
+_spawn_timestamps: dict[str, float] = {}
+SPAWN_COOLDOWN_SEC = 10.0
+
+
+def start_projects_watcher():
+    """Start background thread to refresh projects cache every 5 minutes."""
+    import logging
+    from amc_server.mixins.spawn import load_projects_cache
+
+    def _watch_loop():
+        import time
+        while True:
+            try:
+                time.sleep(300)
+                load_projects_cache()
+            except Exception:
+                logging.exception('Projects cache refresh failed')
+
+    thread = threading.Thread(target=_watch_loop, daemon=True)
+    thread.start()
diff --git a/amc_server/zellij.py b/amc_server/zellij.py
new file mode 100644
index 0000000..f5ee06c
--- /dev/null
+++ b/amc_server/zellij.py
@@ -0,0 +1,34 @@
+"""Zellij integration: binary resolution, plugin path, session name, cache."""
+
+import shutil
+from pathlib import Path
+
+# Plugin path for zellij-send-keys
+ZELLIJ_PLUGIN = Path.home() / ".config" / "zellij" / "plugins" / "zellij-send-keys.wasm"
+
+
+def _resolve_zellij_bin():
+    """Resolve zellij binary even when PATH is minimal (eg launchctl)."""
+    from_path = shutil.which("zellij")
+    if from_path:
+        return from_path
+
+    common_paths = (
+        "/opt/homebrew/bin/zellij",  # Apple Silicon Homebrew
+        "/usr/local/bin/zellij",     # Intel Homebrew
+        "/usr/bin/zellij",
+    )
+    for candidate in common_paths:
+        p = Path(candidate)
+        if p.exists() and p.is_file():
+            return str(p)
+    return "zellij"  # Fallback for explicit error reporting by subprocess
+
+
+ZELLIJ_BIN = _resolve_zellij_bin()
+
+# Default Zellij session for spawning
+ZELLIJ_SESSION = 'infra'
+
+# Cache for Zellij session list (avoid calling zellij on every request)
+_zellij_cache = {"sessions": None, "expires": 0}
diff --git a/tests/test_context.py b/tests/test_context.py
index 1488a52..ef9c3ce 100644
--- a/tests/test_context.py
+++ b/tests/test_context.py
@@ -1,18 +1,18 @@
 import unittest
 from unittest.mock import patch
 
-from amc_server.context import _resolve_zellij_bin
+from amc_server.zellij import _resolve_zellij_bin
 
 
 class ContextTests(unittest.TestCase):
     def test_resolve_zellij_bin_prefers_which(self):
-        with patch("amc_server.context.shutil.which", return_value="/custom/bin/zellij"):
+        with patch("amc_server.zellij.shutil.which", return_value="/custom/bin/zellij"):
             self.assertEqual(_resolve_zellij_bin(), "/custom/bin/zellij")
 
     def test_resolve_zellij_bin_falls_back_to_default_name(self):
-        with patch("amc_server.context.shutil.which", return_value=None), patch(
-            "amc_server.context.Path.exists", return_value=False
-        ), patch("amc_server.context.Path.is_file", return_value=False):
+        with patch("amc_server.zellij.shutil.which", return_value=None), patch(
+            "amc_server.zellij.Path.exists", return_value=False
+        ), patch("amc_server.zellij.Path.is_file", return_value=False):
             self.assertEqual(_resolve_zellij_bin(), "zellij")
 
 
diff --git a/tests/test_control.py b/tests/test_control.py
index 8fa14a3..0e0894c 100644
--- a/tests/test_control.py
+++ b/tests/test_control.py
@@ -336,7 +336,7 @@ class TestDismissSession(unittest.TestCase):
             # (if it existed, it would still exist)
 
     def test_tracks_dismissed_codex_session(self):
-        from amc_server.context import _dismissed_codex_ids
+        from amc_server.agents import _dismissed_codex_ids
         _dismissed_codex_ids.clear()
 
         with tempfile.TemporaryDirectory() as tmpdir:
diff --git a/tests/test_discovery.py b/tests/test_discovery.py
index 28b2a9d..95de599 100644
--- a/tests/test_discovery.py
+++ b/tests/test_discovery.py
@@ -27,7 +27,7 @@ class TestGetCodexPaneInfo(unittest.TestCase):
     def setUp(self):
         self.handler = DummyDiscoveryHandler()
         # Clear cache before each test
-        from amc_server.context import _codex_pane_cache
+        from amc_server.agents import _codex_pane_cache
         _codex_pane_cache["expires"] = 0
         _codex_pane_cache["pid_info"] = {}
         _codex_pane_cache["cwd_map"] = {}
@@ -80,7 +80,7 @@ class TestGetCodexPaneInfo(unittest.TestCase):
         self.assertEqual(pid_info["12345"]["zellij_session"], "myproject")
 
     def test_cache_used_when_fresh(self):
-        from amc_server.context import _codex_pane_cache
+        from amc_server.agents import _codex_pane_cache
         _codex_pane_cache["pid_info"] = {"cached": {"pane_id": "1", "zellij_session": "s"}}
         _codex_pane_cache["cwd_map"] = {"/cached/path": {"session": "s", "pane_id": "1"}}
         _codex_pane_cache["expires"] = time.time() + 100
@@ -203,7 +203,7 @@ class TestDiscoverActiveCodexSessions(unittest.TestCase):
     def setUp(self):
         self.handler = DummyDiscoveryHandler()
         # Clear caches
-        from amc_server.context import _codex_transcript_cache, _dismissed_codex_ids
+        from amc_server.agents import _codex_transcript_cache, _dismissed_codex_ids
         _codex_transcript_cache.clear()
         _dismissed_codex_ids.clear()
 
@@ -236,7 +236,7 @@ class TestDiscoverActiveCodexSessions(unittest.TestCase):
 
     def test_skips_dismissed_sessions(self):
         """Sessions in _dismissed_codex_ids should be skipped."""
-        from amc_server.context import _dismissed_codex_ids
+        from amc_server.agents import _dismissed_codex_ids
 
         with tempfile.TemporaryDirectory() as tmpdir:
             codex_dir = Path(tmpdir)
diff --git a/tests/test_parsing.py b/tests/test_parsing.py
index 006e3b5..7f55cc5 100644
--- a/tests/test_parsing.py
+++ b/tests/test_parsing.py
@@ -228,7 +228,7 @@ class TestFindCodexTranscriptFile(unittest.TestCase):
     def test_codex_sessions_dir_missing_returns_none(self):
         with patch("amc_server.mixins.parsing.CODEX_SESSIONS_DIR", Path("/nonexistent")):
             # Clear cache to force discovery
-            from amc_server.context import _codex_transcript_cache
+            from amc_server.agents import _codex_transcript_cache
             _codex_transcript_cache.clear()
             result = self.handler._find_codex_transcript_file("abc123")
             self.assertIsNone(result)
@@ -238,7 +238,7 @@ class TestFindCodexTranscriptFile(unittest.TestCase):
             transcript_file = Path(tmpdir) / "abc123.jsonl"
             transcript_file.write_text('{"type": "session_meta"}\n')
 
-            from amc_server.context import _codex_transcript_cache
+            from amc_server.agents import _codex_transcript_cache
             _codex_transcript_cache["abc123"] = str(transcript_file)
 
             result = self.handler._find_codex_transcript_file("abc123")
@@ -248,7 +248,7 @@ class TestFindCodexTranscriptFile(unittest.TestCase):
             _codex_transcript_cache.clear()
 
     def test_cache_hit_with_deleted_file_returns_none(self):
-        from amc_server.context import _codex_transcript_cache
+        from amc_server.agents import _codex_transcript_cache
         _codex_transcript_cache["deleted-session"] = "/nonexistent/file.jsonl"
 
         result = self.handler._find_codex_transcript_file("deleted-session")
@@ -257,7 +257,7 @@ class TestFindCodexTranscriptFile(unittest.TestCase):
         _codex_transcript_cache.clear()
 
     def test_cache_hit_with_none_returns_none(self):
-        from amc_server.context import _codex_transcript_cache
+        from amc_server.agents import _codex_transcript_cache
         _codex_transcript_cache["cached-none"] = None
 
         result = self.handler._find_codex_transcript_file("cached-none")
@@ -557,7 +557,7 @@ class TestGetCachedContextUsage(unittest.TestCase):
     def setUp(self):
         self.handler = DummyParsingHandler()
         # Clear cache before each test
-        from amc_server.context import _context_usage_cache
+        from amc_server.agents import _context_usage_cache
         _context_usage_cache.clear()
 
     def test_nonexistent_file_returns_none(self):
diff --git a/tests/test_spawn.py b/tests/test_spawn.py
index 5337c19..457399b 100644
--- a/tests/test_spawn.py
+++ b/tests/test_spawn.py
@@ -290,7 +290,7 @@ class TestRateLimiting(unittest.TestCase):
 
     def test_first_spawn_allowed(self):
         """First spawn for a project should not be rate-limited."""
-        from amc_server.context import _spawn_timestamps
+        from amc_server.spawn_config import _spawn_timestamps
         _spawn_timestamps.clear()
 
         handler = self._make_handler('fresh-project')
@@ -317,7 +317,7 @@ class TestRateLimiting(unittest.TestCase):
 
     def test_rapid_spawn_same_project_rejected(self):
         """Spawning the same project within cooldown returns 429."""
-        from amc_server.context import _spawn_timestamps
+        from amc_server.spawn_config import _spawn_timestamps
         _spawn_timestamps.clear()
         # Pretend we just spawned this project
         _spawn_timestamps['rapid-project'] = time.monotonic()
@@ -339,7 +339,7 @@ class TestRateLimiting(unittest.TestCase):
 
     def test_spawn_different_project_allowed(self):
         """Spawning a different project while one is on cooldown succeeds."""
-        from amc_server.context import _spawn_timestamps
+        from amc_server.spawn_config import _spawn_timestamps
         _spawn_timestamps.clear()
         _spawn_timestamps['project-a'] = time.monotonic()
 
@@ -360,7 +360,7 @@ class TestRateLimiting(unittest.TestCase):
 
     def test_spawn_after_cooldown_allowed(self):
         """Spawning the same project after cooldown expires succeeds."""
-        from amc_server.context import _spawn_timestamps, SPAWN_COOLDOWN_SEC
+        from amc_server.spawn_config import _spawn_timestamps, SPAWN_COOLDOWN_SEC
         _spawn_timestamps.clear()
         # Set timestamp far enough in the past
         _spawn_timestamps['cooled-project'] = time.monotonic() - SPAWN_COOLDOWN_SEC - 1