diff --git a/.env.example b/.env.example
index a1f24d5..c1c0b98 100644
--- a/.env.example
+++ b/.env.example
@@ -32,10 +32,13 @@ FORGE_URL=http://localhost:3000             # [CONFIG] local Forgejo instance
 #   - FORGE_PASS_DEV_QWEN2
 # Name conversion: tr 'a-z-' 'A-Z_' (lowercase→UPPER, hyphens→underscores).
 # The compose generator looks these up via the agent's `forge_user` field in
-# the project TOML. Configure local-model agents via [agents.X] sections in
-# projects/*.toml — this is the canonical activation path.
+# the project TOML. The pre-existing `dev-qwen` llama agent uses
+# FORGE_TOKEN_LLAMA / FORGE_PASS_LLAMA (kept for backwards-compat with the
+# legacy `ENABLE_LLAMA_AGENT=1` single-agent path).
 FORGE_TOKEN=                               # [SECRET] dev-bot API token (default for all agents)
 FORGE_PASS=                                # [SECRET] dev-bot password for git HTTP push (#361)
+FORGE_TOKEN_LLAMA=                         # [SECRET] dev-qwen API token (for agents-llama)
+FORGE_PASS_LLAMA=                          # [SECRET] dev-qwen password for git HTTP push
 FORGE_REVIEW_TOKEN=                        # [SECRET] review-bot API token
 FORGE_REVIEW_PASS=                         # [SECRET] review-bot password for git HTTP push
 FORGE_PLANNER_TOKEN=                       # [SECRET] planner-bot API token
@@ -104,6 +107,13 @@ FORWARD_AUTH_SECRET=                      # [SECRET] Shared secret for Caddy ↔
 # Store all project secrets here so formulas reference env vars, never hardcode.
 BASE_RPC_URL=                              # [SECRET] on-chain RPC endpoint
 
+# ── Local Qwen dev agent (optional) ──────────────────────────────────────
+# Set ENABLE_LLAMA_AGENT=1 to emit agents-llama in docker-compose.yml.
+# Requires a running llama-server reachable at ANTHROPIC_BASE_URL.
+# See docs/agents-llama.md for details.
+ENABLE_LLAMA_AGENT=0                       # [CONFIG] 1 = enable agents-llama service
+ANTHROPIC_BASE_URL=                        # [CONFIG] e.g. http://host.docker.internal:8081
+
 # ── Tuning ────────────────────────────────────────────────────────────────
 CLAUDE_TIMEOUT=7200                        # [CONFIG] max seconds per Claude invocation
 
diff --git a/AGENTS.md b/AGENTS.md
index ad3867b..ef5f00d 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -122,7 +122,8 @@ bash dev/phase-test.sh
 | Reproduce | `docker/reproduce/` | Bug reproduction using Playwright MCP | `formulas/reproduce.toml` |
 | Triage | `docker/reproduce/` | Deep root cause analysis | `formulas/triage.toml` |
 | Edge dispatcher | `docker/edge/` | Polls ops repo for vault actions, executes via Claude sessions | `docker/edge/dispatcher.sh` |
-| Local-model agents | `docker/agents/` (same image) | Local llama-server agents configured via `[agents.X]` sections in project TOML | [docs/agents-llama.md](docs/agents-llama.md) |
+| agents-llama | `docker/agents/` (same image) | Local-Qwen dev agent (`AGENT_ROLES=dev`), gated on `ENABLE_LLAMA_AGENT=1` | [docs/agents-llama.md](docs/agents-llama.md) |
+| agents-llama-all | `docker/agents/` (same image) | Local-Qwen all-roles agent (all 7 roles), profile `agents-llama-all` | [docs/agents-llama.md](docs/agents-llama.md) |
 
 > **Vault:** Being redesigned as a PR-based approval workflow (issues #73-#77).
 > See [docs/VAULT.md](docs/VAULT.md) for the vault PR workflow details.
diff --git a/bin/disinto b/bin/disinto
index 2b676a3..a752bac 100755
--- a/bin/disinto
+++ b/bin/disinto
@@ -684,21 +684,13 @@ _disinto_init_nomad() {
     exit 1
   fi
 
-  # --empty short-circuits after cluster-up: no policies, no auth, no
-  # import, no deploy. It's the "cluster-only escape hatch" for debugging
-  # (docs/nomad-migration.md). Caller-side validation already rejects
-  # --empty combined with --with or any --import-* flag, so reaching
-  # this branch with those set is a bug in the caller.
-  #
-  # On the default (non-empty) path, vault-apply-policies.sh and
-  # vault-nomad-auth.sh are invoked unconditionally — they are idempotent
-  # and cheap to re-run, and subsequent --with deployments depend on
-  # them. vault-import.sh is invoked only when an --import-* flag is set.
+  # Step 2/3/4 scripts must exist as soon as any --import-* flag is set,
+  # since we unconditionally invoke policies+auth and optionally import.
   local import_any=false
   if [ -n "$import_env" ] || [ -n "$import_sops" ]; then
     import_any=true
   fi
-  if [ "$empty" != "true" ]; then
+  if [ "$import_any" = true ]; then
     if [ ! -x "$vault_policies_sh" ]; then
       echo "Error: ${vault_policies_sh} not found or not executable" >&2
       exit 1
@@ -707,7 +699,7 @@ _disinto_init_nomad() {
       echo "Error: ${vault_auth_sh} not found or not executable" >&2
       exit 1
     fi
-    if [ "$import_any" = true ] && [ ! -x "$vault_import_sh" ]; then
+    if [ ! -x "$vault_import_sh" ]; then
       echo "Error: ${vault_import_sh} not found or not executable" >&2
       exit 1
     fi
@@ -730,13 +722,6 @@ _disinto_init_nomad() {
     "${cmd[@]}" || true
     echo ""
 
-    # --empty skips policies/auth/import/deploy — cluster-up only, no
-    # workloads. The operator-visible dry-run plan must match the real
-    # run, so short-circuit here too.
-    if [ "$empty" = "true" ]; then
-      exit 0
-    fi
-
     # Vault policies + auth are invoked on every nomad real-run path
     # regardless of --import-* flags (they're idempotent; S2.1 + S2.3).
     # Mirror that ordering in the dry-run plan so the operator sees the
@@ -808,12 +793,6 @@ _disinto_init_nomad() {
     sudo -n -- "${cluster_cmd[@]}" || exit $?
   fi
 
-  # --empty short-circuits here: cluster-up only, no policies/auth/import
-  # and no deploy. Matches the dry-run plan above and the docs/runbook.
-  if [ "$empty" = "true" ]; then
-    exit 0
-  fi
-
   # Apply Vault policies (S2.1) — idempotent, safe to re-run.
   echo ""
   echo "── Applying Vault policies ────────────────────────────"
@@ -1026,15 +1005,6 @@ disinto_init() {
     exit 1
   fi
 
-  # --empty is the cluster-only escape hatch — it skips policies, auth,
-  # import, and deploy. Pairing it with --import-* silently does nothing,
-  # which is a worse failure mode than a clear error. Reject explicitly.
-  if [ "$empty" = true ] \
-     && { [ -n "$import_env" ] || [ -n "$import_sops" ] || [ -n "$age_key" ]; }; then
-    echo "Error: --empty and --import-env/--import-sops/--age-key are mutually exclusive" >&2
-    exit 1
-  fi
-
   # Dispatch on backend — the nomad path runs lib/init/nomad/cluster-up.sh
   # (S0.4). The default and --empty variants are identical today; Step 1
   # will branch on $empty to add job deployment to the default path.
@@ -1154,6 +1124,7 @@ p.write_text(text)
     echo ""
     echo "[ensure] Forgejo admin user 'disinto-admin'"
     echo "[ensure] 8 bot users: dev-bot, review-bot, planner-bot, gardener-bot, vault-bot, supervisor-bot, predictor-bot, architect-bot"
+    echo "[ensure] 2 llama bot users: dev-qwen, dev-qwen-nightly"
     echo "[ensure] .profile repos for all bots"
     echo "[ensure] repo ${forge_repo} on Forgejo with collaborators"
     echo "[run]    preflight checks"
@@ -1349,6 +1320,19 @@ p.write_text(text)
     echo "Config:  CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 saved to .env"
   fi
 
+  # Write local-Qwen dev agent env keys with safe defaults (#769)
+  if ! grep -q '^ENABLE_LLAMA_AGENT=' "$env_file" 2>/dev/null; then
+    cat >> "$env_file" <<'LLAMAENVEOF'
+
+# Local Qwen dev agent (optional) — set to 1 to enable
+ENABLE_LLAMA_AGENT=0
+FORGE_TOKEN_LLAMA=
+FORGE_PASS_LLAMA=
+ANTHROPIC_BASE_URL=
+LLAMAENVEOF
+    echo "Config:  ENABLE_LLAMA_AGENT keys written to .env (disabled by default)"
+  fi
+
   # Create labels on remote
   create_labels "$forge_repo" "$forge_url"
 
diff --git a/docker/agents/entrypoint.sh b/docker/agents/entrypoint.sh
index 7c58674..f838c15 100644
--- a/docker/agents/entrypoint.sh
+++ b/docker/agents/entrypoint.sh
@@ -17,38 +17,6 @@ set -euo pipefail
 #   - predictor: every 24 hours (288 iterations * 5 min)
 #   - supervisor: every SUPERVISOR_INTERVAL seconds (default: 1200 = 20 min)
 
-# ── Migration check: reject ENABLE_LLAMA_AGENT ───────────────────────────────
-# #846: The legacy ENABLE_LLAMA_AGENT env flag is no longer supported.
-# Activation is now done exclusively via [agents.X] sections in project TOML.
-# If this legacy flag is detected, fail immediately with a migration message.
-if [ "${ENABLE_LLAMA_AGENT:-}" = "1" ]; then
-  cat <<'MIGRATION_ERR'
-FATAL: ENABLE_LLAMA_AGENT is no longer supported.
-
-The legacy ENABLE_LLAMA_AGENT=1 flag has been removed (#846).
-Activation is now done exclusively via [agents.X] sections in projects/*.toml.
-
-To migrate:
-  1. Remove ENABLE_LLAMA_AGENT from your .env or .env.enc file
-  2. Add an [agents.<name>] section to your project TOML:
-
-     [agents.dev-qwen]
-     base_url = "http://your-llama-server:8081"
-     model = "unsloth/Qwen3.5-35B-A3B"
-     api_key = "sk-no-key-required"
-     roles = ["dev"]
-     forge_user = "dev-qwen"
-     compact_pct = 60
-     poll_interval = 60
-
-  3. Run: disinto init
-  4. Start the agent: docker compose up -d agents-dev-qwen
-
-See docs/agents-llama.md for full details.
-MIGRATION_ERR
-  exit 1
-fi
-
 DISINTO_BAKED="/home/agent/disinto"
 DISINTO_LIVE="/home/agent/repos/_factory"
 DISINTO_DIR="$DISINTO_BAKED"  # start with baked copy; switched to live checkout after bootstrap
diff --git a/docs/agents-llama.md b/docs/agents-llama.md
index b3a1334..bc973b7 100644
--- a/docs/agents-llama.md
+++ b/docs/agents-llama.md
@@ -2,12 +2,9 @@
 
 Local-model agents run the same agent code as the Claude-backed agents, but
 connect to a local llama-server (or compatible OpenAI-API endpoint) instead of
-the Anthropic API. This document describes the canonical activation flow using
+the Anthropic API. This document describes the current activation flow using
 `disinto hire-an-agent` and `[agents.X]` TOML configuration.
 
-> **Note:** The legacy `ENABLE_LLAMA_AGENT=1` env flag has been removed (#846).
-> Activation is now done exclusively via `[agents.X]` sections in project TOML.
-
 ## Overview
 
 Local-model agents are configured via `[agents.<name>]` sections in
diff --git a/docs/nomad-migration.md b/docs/nomad-migration.md
index 02ff023..8984b10 100644
--- a/docs/nomad-migration.md
+++ b/docs/nomad-migration.md
@@ -60,9 +60,6 @@ This runs, in order:
 - `--age-key` without `--import-sops` → error.
 - `--import-env` alone (no sops) → OK (imports just the plaintext `.env`).
 - `--backend=docker` with any `--import-*` flag → error.
-- `--empty` with any `--import-*` flag → error (mutually exclusive: `--empty`
-  skips the import step, so pairing them silently discards the import
-  intent).
 
 ## Idempotency
 
diff --git a/formulas/release.sh b/formulas/release.sh
index 6526d1a..b8c4eb6 100644
--- a/formulas/release.sh
+++ b/formulas/release.sh
@@ -178,8 +178,8 @@ log "Tagged disinto/agents:${RELEASE_VERSION}"
 
 log "Step 6/6: Restarting agent containers"
 
-docker compose stop agents 2>/dev/null || true
-docker compose up -d agents
+docker compose stop agents agents-llama 2>/dev/null || true
+docker compose up -d agents agents-llama
 log "Agent containers restarted"
 
 # ── Done ─────────────────────────────────────────────────────────────────
diff --git a/formulas/release.toml b/formulas/release.toml
index ccd7f95..f702f42 100644
--- a/formulas/release.toml
+++ b/formulas/release.toml
@@ -189,10 +189,10 @@ Restart agent containers to use the new image.
    - docker compose pull agents
 
 2. Stop and remove existing agent containers:
-   - docker compose down agents
+   - docker compose down agents agents-llama 2>/dev/null || true
 
 3. Start agents with new image:
-   - docker compose up -d agents
+   - docker compose up -d agents agents-llama
 
 4. Wait for containers to be healthy:
    - for i in {1..30}; do
@@ -203,7 +203,7 @@ Restart agent containers to use the new image.
    - done
 
 5. Verify containers are running:
-   - docker compose ps agents
+   - docker compose ps agents agents-llama
 
 6. Log restart:
    - echo "Restarted agents containers"
diff --git a/lib/action-vault.sh b/lib/action-vault.sh
index 7602a39..6348cc6 100644
--- a/lib/action-vault.sh
+++ b/lib/action-vault.sh
@@ -128,6 +128,7 @@ vault_request() {
   # Validate TOML content
   local tmp_toml
   tmp_toml=$(mktemp /tmp/vault-XXXXXX.toml)
+  trap 'rm -f "$tmp_toml"' RETURN
 
   printf '%s' "$toml_content" > "$tmp_toml"
 
@@ -135,7 +136,6 @@ vault_request() {
   local vault_env="${FACTORY_ROOT:-$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)}/action-vault/vault-env.sh"
   if [ ! -f "$vault_env" ]; then
     echo "ERROR: vault-env.sh not found at $vault_env" >&2
-    rm -f "$tmp_toml"
     return 1
   fi
 
@@ -145,15 +145,11 @@ vault_request() {
   if ! source "$vault_env"; then
     FORGE_TOKEN="${_saved_forge_token:-}"
     echo "ERROR: failed to source vault-env.sh" >&2
-    rm -f "$tmp_toml"
     return 1
   fi
   # Restore caller's FORGE_TOKEN after validation
   FORGE_TOKEN="${_saved_forge_token:-}"
 
-  # Set trap AFTER sourcing vault-env.sh to avoid RETURN trap firing during source
-  trap 'rm -f "$tmp_toml"' RETURN
-
   # Run validation
   if ! validate_vault_action "$tmp_toml"; then
     echo "ERROR: TOML validation failed" >&2
diff --git a/lib/forge-setup.sh b/lib/forge-setup.sh
index 2f8b117..2b7b697 100644
--- a/lib/forge-setup.sh
+++ b/lib/forge-setup.sh
@@ -356,6 +356,16 @@ setup_forge() {
     [predictor-bot]="FORGE_PREDICTOR_PASS"
     [architect-bot]="FORGE_ARCHITECT_PASS"
   )
+  # Llama bot users (local-model agents) — separate from main agents
+  # Each llama agent gets its own Forgejo user, token, and password
+  local -A llama_token_vars=(
+    [dev-qwen]="FORGE_TOKEN_LLAMA"
+    [dev-qwen-nightly]="FORGE_TOKEN_LLAMA_NIGHTLY"
+  )
+  local -A llama_pass_vars=(
+    [dev-qwen]="FORGE_PASS_LLAMA"
+    [dev-qwen-nightly]="FORGE_PASS_LLAMA_NIGHTLY"
+  )
 
   local bot_user bot_pass token token_var pass_var
 
@@ -505,12 +515,159 @@ setup_forge() {
     fi
   done
 
+  # Create llama bot users and tokens (local-model agents)
+  # These are separate from the main agents and get their own credentials
+  echo ""
+  echo "── Setting up llama bot users ────────────────────────────"
+
+  local llama_user llama_pass llama_token llama_token_var llama_pass_var
+  for llama_user in "${!llama_token_vars[@]}"; do
+    llama_token_var="${llama_token_vars[$llama_user]}"
+    llama_pass_var="${llama_pass_vars[$llama_user]}"
+
+    # Check if token already exists in .env
+    local token_exists=false
+    if _token_exists_in_env "$llama_token_var" "$env_file"; then
+      token_exists=true
+    fi
+
+    # Check if password already exists in .env
+    local pass_exists=false
+    if _pass_exists_in_env "$llama_pass_var" "$env_file"; then
+      pass_exists=true
+    fi
+
+    # Check if llama bot user exists on Forgejo
+    local llama_user_exists=false
+    if curl -sf --max-time 5 \
+      -H "Authorization: token ${admin_token}" \
+      "${forge_url}/api/v1/users/${llama_user}" >/dev/null 2>&1; then
+      llama_user_exists=true
+    fi
+
+    # Skip token/password regeneration if both exist in .env and not forcing rotation
+    if [ "$token_exists" = true ] && [ "$pass_exists" = true ] && [ "$rotate_tokens" = false ]; then
+      echo "  ${llama_user} token and password preserved (use --rotate-tokens to force)"
+      # Still export the existing token for use within this run
+      local existing_token existing_pass
+      existing_token=$(grep "^${llama_token_var}=" "$env_file" | head -1 | cut -d= -f2-)
+      existing_pass=$(grep "^${llama_pass_var}=" "$env_file" | head -1 | cut -d= -f2-)
+      export "${llama_token_var}=${existing_token}"
+      export "${llama_pass_var}=${existing_pass}"
+      continue
+    fi
+
+    # Generate new credentials if:
+    # - Token doesn't exist (first run)
+    # - Password doesn't exist (first run)
+    # - --rotate-tokens flag is set (explicit rotation)
+    if [ "$llama_user_exists" = false ]; then
+      # User doesn't exist - create it
+      llama_pass="llama-$(head -c 16 /dev/urandom | base64 | tr -dc 'a-zA-Z0-9' | head -c 20)"
+      echo "Creating llama bot user: ${llama_user}"
+      local create_output
+      if ! create_output=$(_forgejo_exec forgejo admin user create \
+        --username "${llama_user}" \
+        --password "${llama_pass}" \
+        --email "${llama_user}@disinto.local" \
+        --must-change-password=false 2>&1); then
+        echo "Error: failed to create llama bot user '${llama_user}':" >&2
+        echo "  ${create_output}" >&2
+        exit 1
+      fi
+      # Forgejo 11.x ignores --must-change-password=false on create;
+      # explicitly clear the flag so basic-auth token creation works.
+      _forgejo_exec forgejo admin user change-password \
+        --username "${llama_user}" \
+        --password "${llama_pass}" \
+        --must-change-password=false
+
+      # Verify llama bot user was actually created
+      if ! curl -sf --max-time 5 \
+        -H "Authorization: token ${admin_token}" \
+        "${forge_url}/api/v1/users/${llama_user}" >/dev/null 2>&1; then
+        echo "Error: llama bot user '${llama_user}' not found after creation" >&2
+        exit 1
+      fi
+      echo "  ${llama_user} user created"
+    else
+      # User exists - reset password if needed
+      echo "  ${llama_user} user exists"
+      if [ "$rotate_tokens" = true ] || [ "$pass_exists" = false ]; then
+        llama_pass="llama-$(head -c 16 /dev/urandom | base64 | tr -dc 'a-zA-Z0-9' | head -c 20)"
+        _forgejo_exec forgejo admin user change-password \
+          --username "${llama_user}" \
+          --password "${llama_pass}" \
+          --must-change-password=false || {
+          echo "Error: failed to reset password for existing llama bot user '${llama_user}'" >&2
+          exit 1
+        }
+        echo "  ${llama_user} password reset for token generation"
+      else
+        # Password exists, get it from .env
+        llama_pass=$(grep "^${llama_pass_var}=" "$env_file" | head -1 | cut -d= -f2-)
+      fi
+    fi
+
+    # Generate token via API (basic auth as the llama user)
+    # First, delete any existing tokens to avoid name collision
+    local existing_llama_token_ids
+    existing_llama_token_ids=$(curl -sf \
+      -u "${llama_user}:${llama_pass}" \
+      "${forge_url}/api/v1/users/${llama_user}/tokens" 2>/dev/null \
+      | jq -r '.[].id // empty' 2>/dev/null) || existing_llama_token_ids=""
+
+    # Delete any existing tokens for this user
+    if [ -n "$existing_llama_token_ids" ]; then
+      while IFS= read -r tid; do
+        [ -n "$tid" ] && curl -sf -X DELETE \
+          -u "${llama_user}:${llama_pass}" \
+          "${forge_url}/api/v1/users/${llama_user}/tokens/${tid}" >/dev/null 2>&1 || true
+      done <<< "$existing_llama_token_ids"
+    fi
+
+    llama_token=$(curl -sf -X POST \
+      -u "${llama_user}:${llama_pass}" \
+      -H "Content-Type: application/json" \
+      "${forge_url}/api/v1/users/${llama_user}/tokens" \
+      -d "{\"name\":\"disinto-${llama_user}-token\",\"scopes\":[\"all\"]}" 2>/dev/null \
+      | jq -r '.sha1 // empty') || llama_token=""
+
+    if [ -z "$llama_token" ]; then
+      echo "Error: failed to create API token for '${llama_user}'" >&2
+      exit 1
+    fi
+
+    # Store token in .env under the llama-specific variable name
+    if grep -q "^${llama_token_var}=" "$env_file" 2>/dev/null; then
+      sed -i "s|^${llama_token_var}=.*|${llama_token_var}=${llama_token}|" "$env_file"
+    else
+      printf '%s=%s\n' "$llama_token_var" "$llama_token" >> "$env_file"
+    fi
+    export "${llama_token_var}=${llama_token}"
+    echo "  ${llama_user} token generated and saved (${llama_token_var})"
+
+    # Store password in .env for git HTTP push (#361)
+    # Forgejo 11.x API tokens don't work for git push; password auth does.
+    if grep -q "^${llama_pass_var}=" "$env_file" 2>/dev/null; then
+      sed -i "s|^${llama_pass_var}=.*|${llama_pass_var}=${llama_pass}|" "$env_file"
+    else
+      printf '%s=%s\n' "$llama_pass_var" "$llama_pass" >> "$env_file"
+    fi
+    export "${llama_pass_var}=${llama_pass}"
+    echo "  ${llama_user} password saved (${llama_pass_var})"
+  done
+
   # Create .profile repos for all bot users (if they don't already exist)
   # This runs the same logic as hire-an-agent Step 2-3 for idempotent setup
   echo ""
   echo "── Setting up .profile repos ────────────────────────────"
 
   local -a bot_users=(dev-bot review-bot planner-bot gardener-bot vault-bot supervisor-bot predictor-bot architect-bot)
+  # Add llama bot users to .profile repo creation
+  for llama_user in "${!llama_token_vars[@]}"; do
+    bot_users+=("$llama_user")
+  done
   local bot_user
 
   for bot_user in "${bot_users[@]}"; do
@@ -618,6 +775,15 @@ setup_forge() {
         -d "{\"permission\":\"${bot_perm}\"}" >/dev/null 2>&1 || true
     done
 
+    # Add llama bot users as write collaborators for local-model agents
+    for llama_user in "${!llama_token_vars[@]}"; do
+      curl -sf -X PUT \
+        -H "Authorization: token ${admin_token:-${FORGE_TOKEN}}" \
+        -H "Content-Type: application/json" \
+        "${forge_url}/api/v1/repos/${repo_slug}/collaborators/${llama_user}" \
+        -d '{"permission":"write"}' >/dev/null 2>&1 || true
+    done
+
     # Add disinto-admin as admin collaborator
     curl -sf -X PUT \
       -H "Authorization: token ${admin_token:-${FORGE_TOKEN}}" \
diff --git a/lib/generators.sh b/lib/generators.sh
index 0df5725..3f88e39 100644
--- a/lib/generators.sh
+++ b/lib/generators.sh
@@ -438,6 +438,136 @@ services:
 
 COMPOSEEOF
 
+  # ── Conditional agents-llama block (ENABLE_LLAMA_AGENT=1) ──────────────
+  # Local-Qwen dev agent — gated on ENABLE_LLAMA_AGENT so factories without
+  # a local llama endpoint don't try to start it.  See docs/agents-llama.md.
+  if [ "${ENABLE_LLAMA_AGENT:-0}" = "1" ]; then
+    cat >> "$compose_file" <<'LLAMAEOF'
+
+  agents-llama:
+    build:
+      context: .
+      dockerfile: docker/agents/Dockerfile
+    # Rebuild on every up (#887): makes docker/agents/ source changes reach this
+    # container without a manual \`docker compose build\`. Cache-fast when clean.
+    pull_policy: build
+    container_name: disinto-agents-llama
+    restart: unless-stopped
+    security_opt:
+      - apparmor=unconfined
+    volumes:
+      - agent-data:/home/agent/data
+      - project-repos:/home/agent/repos
+      - ${CLAUDE_SHARED_DIR:-/var/lib/disinto/claude-shared}:${CLAUDE_SHARED_DIR:-/var/lib/disinto/claude-shared}
+      - ${CLAUDE_CONFIG_FILE:-${HOME}/.claude.json}:/home/agent/.claude.json:ro
+      - ${CLAUDE_BIN_DIR}:/usr/local/bin/claude:ro
+      - ${AGENT_SSH_DIR:-${HOME}/.ssh}:/home/agent/.ssh:ro
+      - ${SOPS_AGE_DIR:-${HOME}/.config/sops/age}:/home/agent/.config/sops/age:ro
+      - woodpecker-data:/woodpecker-data:ro
+    environment:
+      FORGE_URL: http://forgejo:3000
+      FORGE_REPO: ${FORGE_REPO:-disinto-admin/disinto}
+      FORGE_TOKEN: ${FORGE_TOKEN_LLAMA:-}
+      FORGE_PASS: ${FORGE_PASS_LLAMA:-}
+      FORGE_BOT_USERNAMES: ${FORGE_BOT_USERNAMES:-}
+      WOODPECKER_TOKEN: ${WOODPECKER_TOKEN:-}
+      CLAUDE_TIMEOUT: ${CLAUDE_TIMEOUT:-7200}
+      CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC: ${CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC:-1}
+      CLAUDE_AUTOCOMPACT_PCT_OVERRIDE: "60"
+      ANTHROPIC_API_KEY: ${ANTHROPIC_API_KEY:-}
+      ANTHROPIC_BASE_URL: ${ANTHROPIC_BASE_URL:-}
+      FORGE_ADMIN_PASS: ${FORGE_ADMIN_PASS:-}
+      DISINTO_CONTAINER: "1"
+      PROJECT_NAME: ${PROJECT_NAME:-project}
+      PROJECT_REPO_ROOT: /home/agent/repos/${PROJECT_NAME:-project}
+      WOODPECKER_DATA_DIR: /woodpecker-data
+      WOODPECKER_REPO_ID: "PLACEHOLDER_WP_REPO_ID"
+      CLAUDE_CONFIG_DIR: ${CLAUDE_CONFIG_DIR:-/var/lib/disinto/claude-shared/config}
+      POLL_INTERVAL: ${POLL_INTERVAL:-300}
+      AGENT_ROLES: dev
+    healthcheck:
+      test: ["CMD", "pgrep", "-f", "entrypoint.sh"]
+      interval: 60s
+      timeout: 5s
+      retries: 3
+      start_period: 30s
+    depends_on:
+      forgejo:
+        condition: service_healthy
+    networks:
+      - disinto-net
+
+  agents-llama-all:
+    build:
+      context: .
+      dockerfile: docker/agents/Dockerfile
+    # Rebuild on every up (#887): makes docker/agents/ source changes reach this
+    # container without a manual \`docker compose build\`. Cache-fast when clean.
+    pull_policy: build
+    container_name: disinto-agents-llama-all
+    restart: unless-stopped
+    profiles: ["agents-llama-all"]
+    security_opt:
+      - apparmor=unconfined
+    volumes:
+      - agent-data:/home/agent/data
+      - project-repos:/home/agent/repos
+      - ${CLAUDE_SHARED_DIR:-/var/lib/disinto/claude-shared}:${CLAUDE_SHARED_DIR:-/var/lib/disinto/claude-shared}
+      - ${CLAUDE_CONFIG_FILE:-${HOME}/.claude.json}:/home/agent/.claude.json:ro
+      - ${CLAUDE_BIN_DIR}:/usr/local/bin/claude:ro
+      - ${AGENT_SSH_DIR:-${HOME}/.ssh}:/home/agent/.ssh:ro
+      - ${SOPS_AGE_DIR:-${HOME}/.config/sops/age}:/home/agent/.config/sops/age:ro
+      - woodpecker-data:/woodpecker-data:ro
+    environment:
+      FORGE_URL: http://forgejo:3000
+      FORGE_REPO: ${FORGE_REPO:-disinto-admin/disinto}
+      FORGE_TOKEN: ${FORGE_TOKEN_LLAMA:-}
+      FORGE_PASS: ${FORGE_PASS_LLAMA:-}
+      FORGE_REVIEW_TOKEN: ${FORGE_REVIEW_TOKEN:-}
+      FORGE_PLANNER_TOKEN: ${FORGE_PLANNER_TOKEN:-}
+      FORGE_GARDENER_TOKEN: ${FORGE_GARDENER_TOKEN:-}
+      FORGE_VAULT_TOKEN: ${FORGE_VAULT_TOKEN:-}
+      FORGE_SUPERVISOR_TOKEN: ${FORGE_SUPERVISOR_TOKEN:-}
+      FORGE_PREDICTOR_TOKEN: ${FORGE_PREDICTOR_TOKEN:-}
+      FORGE_ARCHITECT_TOKEN: ${FORGE_ARCHITECT_TOKEN:-}
+      FORGE_FILER_TOKEN: ${FORGE_FILER_TOKEN:-}
+      FORGE_BOT_USERNAMES: ${FORGE_BOT_USERNAMES:-}
+      WOODPECKER_TOKEN: ${WOODPECKER_TOKEN:-}
+      CLAUDE_TIMEOUT: ${CLAUDE_TIMEOUT:-7200}
+      CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC: ${CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC:-1}
+      CLAUDE_AUTOCOMPACT_PCT_OVERRIDE: "60"
+      CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS: "1"
+      ANTHROPIC_API_KEY: ${ANTHROPIC_API_KEY:-}
+      ANTHROPIC_BASE_URL: ${ANTHROPIC_BASE_URL:-}
+      FORGE_ADMIN_PASS: ${FORGE_ADMIN_PASS:-}
+      DISINTO_CONTAINER: "1"
+      PROJECT_NAME: ${PROJECT_NAME:-project}
+      PROJECT_REPO_ROOT: /home/agent/repos/${PROJECT_NAME:-project}
+      WOODPECKER_DATA_DIR: /woodpecker-data
+      WOODPECKER_REPO_ID: "PLACEHOLDER_WP_REPO_ID"
+      CLAUDE_CONFIG_DIR: ${CLAUDE_CONFIG_DIR:-/var/lib/disinto/claude-shared/config}
+      POLL_INTERVAL: ${POLL_INTERVAL:-300}
+      GARDENER_INTERVAL: ${GARDENER_INTERVAL:-21600}
+      ARCHITECT_INTERVAL: ${ARCHITECT_INTERVAL:-21600}
+      PLANNER_INTERVAL: ${PLANNER_INTERVAL:-43200}
+      SUPERVISOR_INTERVAL: ${SUPERVISOR_INTERVAL:-1200}
+      AGENT_ROLES: review,dev,gardener,architect,planner,predictor,supervisor
+    healthcheck:
+      test: ["CMD", "pgrep", "-f", "entrypoint.sh"]
+      interval: 60s
+      timeout: 5s
+      retries: 3
+      start_period: 30s
+    depends_on:
+      forgejo:
+        condition: service_healthy
+      woodpecker:
+        condition: service_started
+    networks:
+      - disinto-net
+LLAMAEOF
+  fi
+
   # Resume the rest of the compose file (runner onward)
   cat >> "$compose_file" <<'COMPOSEEOF'
 
diff --git a/tests/disinto-init-nomad.bats b/tests/disinto-init-nomad.bats
index f38805e..30c7f7c 100644
--- a/tests/disinto-init-nomad.bats
+++ b/tests/disinto-init-nomad.bats
@@ -280,33 +280,3 @@ setup_file() {
   [ "$status" -eq 0 ]
   [[ "$output" == *"env file:  /tmp/.env"* ]]
 }
-
-# --empty short-circuits after cluster-up: no policies, no auth, no
-# import, no deploy. The dry-run plan must match that — cluster-up plan
-# appears, but none of the S2.x section banners do.
-@test "disinto init --backend=nomad --empty --dry-run skips policies/auth/import sections" {
-  run "$DISINTO_BIN" init placeholder/repo --backend=nomad --empty --dry-run
-  [ "$status" -eq 0 ]
-  # Cluster-up still runs (it's what --empty brings up).
-  [[ "$output" == *"Cluster-up dry-run"* ]]
-  # Policies + auth + import must NOT appear under --empty.
-  [[ "$output" != *"Vault policies dry-run"* ]]
-  [[ "$output" != *"Vault auth dry-run"* ]]
-  [[ "$output" != *"Vault import dry-run"* ]]
-  [[ "$output" != *"no --import-env/--import-sops"* ]]
-}
-
-# --empty + any --import-* flag silently does nothing (import is skipped),
-# so the CLI rejects the combination up front rather than letting it
-# look like the import "succeeded".
-@test "disinto init --backend=nomad --empty --import-env errors" {
-  run "$DISINTO_BIN" init placeholder/repo --backend=nomad --empty --import-env /tmp/.env --dry-run
-  [ "$status" -ne 0 ]
-  [[ "$output" == *"--empty and --import-env/--import-sops/--age-key are mutually exclusive"* ]]
-}
-
-@test "disinto init --backend=nomad --empty --import-sops --age-key errors" {
-  run "$DISINTO_BIN" init placeholder/repo --backend=nomad --empty --import-sops /tmp/.env.vault.enc --age-key /tmp/keys.txt --dry-run
-  [ "$status" -ne 0 ]
-  [[ "$output" == *"--empty and --import-env/--import-sops/--age-key are mutually exclusive"* ]]
-}