feat(control-plane): host-side deploy runner — fix broken redeploy path

fix(control-plane): redeploy wykonywalny — dispatch do host-side deploy-runnera
Executor odpalal scripts/deploy/deploy-node.sh <node> <service> wewnatrz swojego kontenera: skrypt ignoruje oba argumenty i wymaga repo w ${HOME}/homelab-codex-ws (w kontenerze HOME=/home/homelab) -> exit 1 w 18. linii. Za tym brak git, brak klienta docker w obrazie, a gdyby przeszedl — deploy calego zestawu uslug hosta executora zamiast wezla z akcji. Kazdy redeploy padal (recon D14/D15; 18 pending / 0 completed). Redeploy idzie teraz ta sama sciezka pull co container_restart — VPS nigdy nie inicjuje polaczenia do wezla: executor -> actions/deploy/<node>/<id>.json -> deploy-runner (systemd na hoscie) rsync-pull, walidacja, deploy -> action_result event -> executor rozlicza completed/failed - scripts/deploy/deploy-service.sh: deploy jednej uslugi, wspoldzielony z deploy-node.sh, wiec inwokacja compose (a przez to nazwa projektu) jest identyczna jak przy deployu recznym - jobs/deploy-runner/: host-level, nie kontener — compose rozwiazuje wzgledne bindy i nazwe projektu tak jak przy deployu czlowieka; niezalezny od node-agenta, wiec potrafi zredeployowac takze jego - walidacja: tylko typ redeploy, node musi sie zgadzac, usluga musi byc w hosts/<node>/services.yaml, zadna tresc z payloadu nie trafia do shella - --force-recreate bez --build i bez --remove-orphans: redeploy to rekoncyliacja, nie wysylka kodu - executor: REDEPLOY_TIMEOUT_SECS=900, /repo zjechany do :ro (nieuzywany) 248 testow zielonych; deploy-node.sh przecwiczony na atrapie dockera — argv compose bez zmian. Instalacja unitow na wezlach i E2E: backlog. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-03 19:23:16 +02:00 · 2026-08-03 18:26:30 +02:00
18 changed files with 1681 additions and 87 deletions
--- a/CLAUDE.md
+++ b/CLAUDE.md
@ -90,9 +90,21 @@ The platform uses a multi-agent model with **human-in-the-loop** for destructive
 Agent → /opt/homelab/actions/pending/<id>.json
      → Telegram notification → Operator approves
      → /opt/homelab/actions/approved/<id>.json
-      → Executor runs → completed / failed
+      → Executor dispatches to the target node → completed / failed
 ```

+The executor never connects to a node (deliberate — see docs/backlog.md
+"Remediacja floty bez SSH"). It writes a dispatch file that the node collects:
+
+| Action type | Inbox | Executed on the node by |
+|---|---|---|
+| `container_restart` | `actions/dispatch/<node>/` | node-agent (own docker socket) |
+| `redeploy` | `actions/deploy/<node>/` | deploy-runner, host-level systemd (`jobs/deploy-runner/`) |
+
+Both report back with an `action_result` event, which the executor turns into
+`completed` / `failed`. `disk_cleanup` and `alert_only` still resolve inside the
+executor.
+
 Agents must never execute destructive actions (restarts, deploys, config changes) without a corresponding approved action file.

 ## Event System
@ -109,7 +121,7 @@ Normalized event types: `deployment_started/completed/failed`, `service_unhealth
 |---|---|---|---|
 | `containers_not_running` | stability-agent | `container_restart` | dedup via stable ID |
 | `healthcheck_failed` | node-agent | `container_restart` | dedup via stable ID |
-| `service_unhealthy` / other | stability-agent | `redeploy` (broken until etap 2 executor fix) | dedup via stable ID |
+| `service_unhealthy` / other | stability-agent | `redeploy` → dispatched to the node's deploy-runner (`jobs/deploy-runner/`) | dedup via stable ID |
 | `disk_pressure` (high) | stability-agent | `disk_cleanup` | dedup via stable ID |
 | `ha_websocket_dead` | ha-diag-agent | `container_restart` (homeassistant) | 30 min after completion |
 | `ha_websocket_recovered` | ha-diag-agent | cancels matching restart | — |
--- a/docs/backlog.md
+++ b/docs/backlog.md
@ -4,6 +4,60 @@ Centralny tracker tech-długu i znanych usterek. Wpisy ze sesji — dodawaj z da

 ---

+## deploy-runner: instalacja na węzłach + E2E redeployu (2026-08-03)
+
+**Data**: 2026-08-03
+**Źródło**: sesja `task/redeploy-fix` — recon D14/D15 + OPEN QUESTION 4 (nieopróżniona
+kolejka akcji).
+**Było**: `redeploy` nie mógł się wykonać nigdy. Executor odpalał
+`scripts/deploy/deploy-node.sh <node> <service>` **wewnątrz swojego kontenera** —
+skrypt ignoruje oba argumenty i wymaga repo w `${HOME}/homelab-codex-ws`
+(w kontenerze `HOME=/home/homelab`, katalog nie istnieje) → `exit 1` w 18. linii.
+Za tym stały jeszcze trzy blokady: brak `git`, brak klienta `docker` w obrazie oraz
+— gdyby przeszedł — deploy **całego zestawu usług hosta executora**, nie węzła
+z akcji. Stąd `healthcheck_failed` przekierowany 2026-07-29 na `container_restart`
+i 18 pending / 0 completed.
+**Zrobione w repo (ta sesja)**: `scripts/deploy/deploy-service.sh` (deploy jednej
+usługi, wspólny z `deploy-node.sh` — ta sama inwokacja compose, więc ta sama nazwa
+projektu), `jobs/deploy-runner/` (systemd na hoście: rsync-pull akcji, walidacja,
+deploy, `action_result` z powrotem), executor dispatchuje `redeploy` do
+`actions/deploy/<node>/` i rozlicza je jak `container_restart`
+(`REDEPLOY_TIMEOUT_SECS=900`). 248 testów zielonych.
+**Do zrobienia (runtime, wymaga operatora)**:
+1. Deploy control-plane na VPS (nowy executor + `../..:/repo:ro`).
+2. Instalacja `jobs/deploy-runner/` na vps, piha, solaria — patrz README
+   („Install (per node)"). Na VPS `VPS_EVENTS_HOST` musi zostać puste.
+3. E2E na benignej usłudze na PIHA (wzorzec `test-e2e-b` z 2026-07-23), potem
+   opróżnienie kolejki 18 pending — w tym `redeploy-vps-gokapi`.
+4. SOLARIA: `group_add: "996"` dla node-agenta wciąż niewdrożony (recon 641-649) —
+   dopóki nie wejdzie, `container_restart` tam nie działa (redeploy działa, bo nie
+   idzie przez node-agenta).
+
+---
+
+## disk_cleanup w executorze jest zepsuty tak samo jak był redeploy (2026-08-03)
+
+**Data**: 2026-08-03
+**Źródło**: sesja `task/redeploy-fix` (znalezione przy okazji, poza zakresem zadania).
+**Problem**: `executor._execute_disk_cleanup()` odpala `ssh oskar@<node> …`, a obraz
+control-plane (`python:3.11-slim` + `pip install pyyaml`) **nie ma klienta ssh** —
+`subprocess.run(["ssh", …])` leci `FileNotFoundError`, akcja ląduje w `failed/`.
+To ta sama klasa błędu co redeploy i sprzeczne z decyzją „bez SSH w executorze".
+**Fix**: przenieść `disk_cleanup` na model dispatch (node-agent albo deploy-runner —
+runner ma już hosta i uprawnienia) albo usunąć typ akcji. Do decyzji przy okazji
+opróżniania kolejki.
+
+---
+
+## `scripts/deploy/deploy-host.sh` to pusty plik (2026-08-03)
+
+**Data**: 2026-08-03
+**Źródło**: sesja `task/redeploy-fix`.
+**Problem**: 0-bajtowy, wykonywalny stub w `scripts/deploy/` — wygląda jak
+entrypoint, nie robi nic. Kandydat do skasowania (jak `deploy-role.sh` w etapie 0).
+
+---
+
 ## Cutover HA "ken": kontener piha to legacy, prawdziwy dom to RPi4/HAOS (2026-07-22)

 **Data**: 2026-07-22
--- a/jobs/deploy-runner/README.md
+++ b/jobs/deploy-runner/README.md
@ -0,0 +1,111 @@
+# deploy-runner — host-side executor for `redeploy` actions
+
+Runs on each node that should be able to execute an operator-approved
+`redeploy`. It is the missing half of the self-healing loop: the control-plane
+executor can generate and dispatch redeploys, but nothing on the node could ever
+perform one.
+
+## What was broken
+
+The executor ran `scripts/deploy/deploy-node.sh <node> <service>` **inside its own
+container**. That script ignores both arguments, and expects a repo at
+`${HOME}/homelab-codex-ws` — inside the container `HOME=/home/homelab`, so it
+exited at line 18 with `Error: Repository not found`. Behind that failure sat
+three more: no `git`, no `docker` CLI in the image, and — had it ever got that
+far — it would have deployed the **executor host's** entire service set, not the
+action's target node/service. Every `healthcheck_failed → redeploy` dead-ended
+(recon `docs/architecture/RECON-multiagent-2026-07-27.md`, D14/D15).
+
+## How it works now
+
+```
+supervisor      → actions/pending/<id>.json           (unchanged)
+operator        → actions/approved/<id>.json          (unchanged, HITL)
+executor (vps)  → actions/deploy/<node>/<id>.json     ← dispatch only, no execution
+deploy-runner   ← rsync-pull (node initiates; VPS never connects to a node)
+                → scripts/deploy/deploy-service.sh --force-recreate
+                → events/<node>/evt-…-action_result-….json → rsync-push
+executor        → completed / failed                  (via _reconcile_running_actions)
+```
+
+Design points, and why:
+
+- **Host-level, not a container.** Compose resolves relative bind mounts and the
+  project name against the filesystem of whatever runs it. In a container those
+  resolve to container paths that the daemon then interprets as host paths — a
+  silent way to produce broken mounts, or to land in a different Compose project
+  where `--remove-orphans` deletes the running stack. On the host it behaves
+  exactly like a human `deploy.sh`.
+- **Independent of node-agent** (own rsync pull, own result push) so it can
+  redeploy `node-agent` itself — the solaria case, where node-agent has been
+  docker-blind for weeks.
+- **Separate inbox** from `actions/dispatch/<node>/`: node-agent deletes and
+  failure-reports every file in its own inbox that is not `container_restart`.
+- **No `git pull`.** A redeploy reconciles the node to the checkout it already
+  has. Shipping new code stays a human `scripts/deploy/deploy.sh` action.
+- **No `--build`, no `--remove-orphans`, always `--force-recreate`.** Plain
+  `up -d` is a no-op when config is unchanged — exactly the unhealthy-container
+  case; building on the 4 GiB VPS mid-incident is an OOM risk.
+
+Every action is validated before anything runs (`action.py`): type must be
+`redeploy`, `node` must match this node, the service name must be a plain
+kebab-case name, the service must be listed in `hosts/<node>/services.yaml`, and
+its compose file must exist. Nothing from the action payload is ever executed —
+only a validated service *name* reaches the deploy script. Actions are
+idempotent (marker in `state/processed-deploy-actions/`), single-instance
+(`flock`), and time-boxed (`DEPLOY_TIMEOUT_SECS`, default 600 s, below the
+executor's `REDEPLOY_TIMEOUT_SECS` of 900 s so the node reports before the
+control plane gives up).
+
+## Install (per node)
+
+Not deployed by `deploy.sh` — it is a host-level systemd unit, like
+`jobs/documents-ingest/`. On the target node:
+
+```bash
+# 1. config
+sudo mkdir -p /opt/homelab/config/deploy-runner
+sudo cp ~/homelab-codex-ws/jobs/deploy-runner/env.example \
+        /opt/homelab/config/deploy-runner/env
+sudo chown oskar:oskar /opt/homelab/config/deploy-runner/env
+sudoedit /opt/homelab/config/deploy-runner/env      # set NODE_NAME; clear VPS_EVENTS_HOST on the VPS
+
+# 2. units
+sudo cp ~/homelab-codex-ws/jobs/deploy-runner/systemd/homelab-deploy-runner.{service,timer} \
+        /etc/systemd/system/
+sudo systemctl daemon-reload
+sudo systemctl enable --now homelab-deploy-runner.timer
+
+# 3. verify (no action queued yet → one clean, empty run)
+sudo systemctl start homelab-deploy-runner.service
+journalctl -u homelab-deploy-runner.service -n 30 --no-pager
+```
+
+Requirements on the node: repo checkout at `REPO_PATH`, the service user in the
+`docker` group, `python3` + PyYAML, `rsync`, and (remote nodes only) an ssh key
+that reaches the VPS — the same one node-agent already uses for event shipping.
+
+On the **VPS** leave `VPS_EVENTS_HOST` empty: the executor writes into the same
+`/opt/homelab` mount, so there is nothing to pull and nothing to push.
+
+## Operating it
+
+- Dispatched but not yet collected: `ls /opt/homelab/actions/deploy/<node>/` on the VPS.
+- Per-action deploy output: `/opt/homelab/logs/deploy-runner/<action_id>-<ts>.log` on the node.
+- Runner activity: `journalctl -u homelab-deploy-runner.service`.
+- A redeploy for a service with its own `deploy-local.sh` (control-plane) is
+  reported as **failed** with an explanatory message — those need an operator
+  deploy, by design.
+- To let an already-processed action run again, remove its marker:
+  `rm /opt/homelab/state/processed-deploy-actions/<action_id>.done`.
+
+## Tests
+
+`tests/` — validation and rejection cases, the event format contract against the
+executor's real parser, the `docker compose` argv contract of
+`scripts/deploy/deploy-service.sh`, and an end-to-end run of the runner itself
+with a stubbed `docker`.
+
+```bash
+python3 -m pytest jobs/deploy-runner/tests -q
+```
--- a/jobs/deploy-runner/action.py
+++ b/jobs/deploy-runner/action.py
@ -0,0 +1,221 @@
+#!/usr/bin/env python3
+"""Validation and result-reporting helper for the host-side deploy runner.
+
+The runner (deploy-runner.sh) is bash — it does flock, rsync and timeout, which
+bash does well. Everything that needs to parse untrusted JSON, read
+hosts/<node>/services.yaml or write an event in the exact node-agent format
+lives here, where it is testable (jobs/deploy-runner/tests/).
+
+Two subcommands, both designed to be consumed from bash:
+
+  validate --file <action.json> --node <name> --repo <path>
+      Prints shell assignments for `eval`:
+          OK='true'|'false'
+          ACTION_ID='...'      (best effort — set even when OK=false, so the
+                                runner can still report a result for it)
+          SERVICE='...'
+          ERROR='...'
+      Always exits 0 unless the arguments themselves are unusable; the verdict
+      is in OK. Nothing from the action file is ever executed or interpolated
+      into a command — only a validated service NAME is handed back.
+
+  emit-result --events-dir <dir> --node <n> --action-id <id> --service <svc>
+              --success true|false [--error <msg>]
+      Writes evt-<node>-<ts>-action_result-<slug>.json into the node's event
+      directory, byte-compatible with node_agent.emit_event() so the control
+      plane executor's _find_action_result() picks it up unchanged. Prints the
+      path it wrote.
+
+Validation rules (all must hold):
+  1. the file is JSON with type == "redeploy"
+  2. action["node"] == this node — defense in depth; the inbox is already
+     node-scoped, but a mis-delivered file must never deploy here
+  3. action_id and service match strict name patterns (no path traversal)
+  4. the service is listed in hosts/<node>/services.yaml — the repo's desired
+     state is the authority on what this node may run, so a rogue or stale
+     dispatch cannot pull an arbitrary stack onto the node
+  5. services/<service>/docker-compose.yml exists in the repo
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import re
+import shlex
+import sys
+import time
+from datetime import datetime, timezone
+from pathlib import Path
+
+ACTION_ID_RE = re.compile(r"^[A-Za-z0-9][A-Za-z0-9._-]{0,127}$")
+SERVICE_RE = re.compile(r"^[a-z0-9][a-z0-9._-]{0,63}$")
+ACTION_TYPE = "redeploy"
+
+
+class Invalid(Exception):
+    """Validation failure carrying a human-readable reason."""
+
+
+def _load_desired_services(repo: Path, node: str) -> set[str]:
+    """Service names listed in hosts/<node>/services.yaml."""
+    try:
+        import yaml
+    except ImportError as exc:  # pragma: no cover - environment problem
+        raise Invalid(f"PyYAML not available on this node: {exc}") from exc
+
+    path = repo / "hosts" / node / "services.yaml"
+    if not path.is_file():
+        raise Invalid(f"no desired-state manifest for node '{node}' ({path})")
+    try:
+        data = yaml.safe_load(path.read_text()) or {}
+    except Exception as exc:
+        raise Invalid(f"unreadable {path}: {exc}") from exc
+
+    services = data.get("services") or {}
+    if isinstance(services, dict):
+        return {str(k) for k in services.keys()}
+    if isinstance(services, list):
+        return {str(s) for s in services}
+    raise Invalid(f"unexpected 'services' shape in {path}: {type(services).__name__}")
+
+
+def validate(action_file: Path, node: str, repo: Path) -> tuple[str, str]:
+    """Return (action_id, service) for a valid redeploy action, else raise Invalid.
+
+    action_id is also returned inside the Invalid path via the caller's
+    best-effort parse, so a rejected action can still be reported back.
+    """
+    try:
+        action = json.loads(action_file.read_text())
+    except Exception as exc:
+        raise Invalid(f"unparseable action file: {exc}") from exc
+    if not isinstance(action, dict):
+        raise Invalid("action file is not a JSON object")
+
+    action_id = str(action.get("action_id") or action_file.stem)
+    if not ACTION_ID_RE.match(action_id):
+        raise Invalid(f"invalid action_id: {action_id!r}")
+
+    action_type = action.get("type")
+    if action_type != ACTION_TYPE:
+        raise Invalid(
+            f"action type {action_type!r} is not executable by the deploy runner "
+            f"(only {ACTION_TYPE!r})"
+        )
+
+    target_node = action.get("node")
+    if target_node != node:
+        raise Invalid(f"action addressed to node {target_node!r}, not {node!r} — refused")
+
+    service = action.get("service")
+    if not isinstance(service, str) or not SERVICE_RE.match(service):
+        raise Invalid(f"invalid service name: {service!r}")
+
+    desired = _load_desired_services(repo, node)
+    if service not in desired:
+        raise Invalid(
+            f"service '{service}' is not in hosts/{node}/services.yaml — "
+            f"refusing to deploy a service this node is not declared to run"
+        )
+
+    compose = repo / "services" / service / "docker-compose.yml"
+    if not compose.is_file():
+        raise Invalid(f"no compose file in repo for '{service}' ({compose})")
+
+    return action_id, service
+
+
+def _best_effort_action_id(action_file: Path) -> str:
+    try:
+        data = json.loads(action_file.read_text())
+        candidate = str(data.get("action_id") or action_file.stem)
+    except Exception:
+        candidate = action_file.stem
+    return candidate if ACTION_ID_RE.match(candidate) else ""
+
+
+def _emit_assignments(ok: bool, action_id: str, service: str, error: str) -> str:
+    return "\n".join(
+        f"{key}={shlex.quote(value)}"
+        for key, value in (
+            ("OK", "true" if ok else "false"),
+            ("ACTION_ID", action_id),
+            ("SERVICE", service),
+            ("ERROR", error),
+        )
+    )
+
+
+def emit_result(events_dir: Path, node: str, action_id: str, service: str,
+                success: bool, error: str) -> Path:
+    """Write an action_result event in node_agent.emit_event()'s exact format.
+
+    The control-plane executor globs evt-*-action_result-*.json in the node's
+    event directory and matches payload.action_id, comparing the unix timestamp
+    embedded in the filename against the action's started_at — so both the name
+    and the payload shape matter here (see executor._find_action_result).
+    """
+    ts = int(time.time())
+    slug = re.sub(r"[^a-z0-9]", "-", (service or "node").lower())[:32].strip("-")
+    event_id = f"evt-{node}-{ts}-action_result-{slug}"
+    event = {
+        "id": event_id,
+        "timestamp": ts,
+        "date": datetime.now(timezone.utc).isoformat(timespec="seconds").replace("+00:00", "Z"),
+        "type": "action_result",
+        "severity": "info" if success else "high",
+        "node": node,
+        "service": service or "",
+        "message": f"Action {action_id} {'succeeded' if success else 'failed'}",
+        "payload": {
+            "action_id": action_id,
+            "success": success,
+            "error": error,
+            "node": node,
+            "source": "deploy-runner",
+        },
+    }
+    events_dir.mkdir(parents=True, exist_ok=True)
+    path = events_dir / f"{event_id}.json"
+    path.write_text(json.dumps(event, indent=2))
+    return path
+
+
+def main(argv=None) -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    sub = parser.add_subparsers(dest="command", required=True)
+
+    p_val = sub.add_parser("validate")
+    p_val.add_argument("--file", required=True, type=Path)
+    p_val.add_argument("--node", required=True)
+    p_val.add_argument("--repo", required=True, type=Path)
+
+    p_emit = sub.add_parser("emit-result")
+    p_emit.add_argument("--events-dir", required=True, type=Path)
+    p_emit.add_argument("--node", required=True)
+    p_emit.add_argument("--action-id", required=True)
+    p_emit.add_argument("--service", default="")
+    p_emit.add_argument("--success", required=True, choices=["true", "false"])
+    p_emit.add_argument("--error", default="")
+
+    args = parser.parse_args(argv)
+
+    if args.command == "validate":
+        try:
+            action_id, service = validate(args.file, args.node, args.repo)
+        except Invalid as exc:
+            print(_emit_assignments(False, _best_effort_action_id(args.file), "", str(exc)))
+            return 0
+        print(_emit_assignments(True, action_id, service, ""))
+        return 0
+
+    path = emit_result(
+        args.events_dir, args.node, args.action_id, args.service,
+        args.success == "true", args.error,
+    )
+    print(path)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
--- a/jobs/deploy-runner/deploy-runner.sh
+++ b/jobs/deploy-runner/deploy-runner.sh
@ -0,0 +1,200 @@
+#!/usr/bin/env bash
+# jobs/deploy-runner/deploy-runner.sh — host-side executor for `redeploy` actions.
+#
+# WHY THIS EXISTS (docs/architecture/RECON-multiagent-2026-07-27.md D14/D15):
+# the control-plane executor could never carry out a redeploy — it ran
+# deploy-node.sh inside its own container, a script that ignores its arguments
+# and expects a repo at ${HOME}/homelab-codex-ws that does not exist there (nor
+# does git or the docker CLI). Every healthcheck_failed → redeploy action
+# dead-ended, which is the single biggest reason the self-healing loop had
+# 18 pending / 0 completed actions.
+#
+# The fix keeps the architecture decision from docs/backlog.md ("Remediacja
+# floty bez SSH"): the VPS never initiates a connection to a node. The executor
+# only WRITES an action file; this runner, on the target node, pulls it, runs
+# the deploy locally, and reports the outcome back through the existing event
+# pipeline.
+#
+#   executor (vps)  →  /opt/homelab/actions/deploy/<node>/<action_id>.json
+#     → this runner rsync-pulls it (--remove-source-files: collected once)
+#     → validates it (jobs/deploy-runner/action.py)
+#     → scripts/deploy/deploy-service.sh --force-recreate  (same compose
+#       invocation as a human deploy — identical project name)
+#     → writes an action_result event and rsync-pushes it back to the VPS
+#     → executor._reconcile_running_actions() → completed / failed
+#
+# It runs on the HOST, not in a container: compose then resolves relative bind
+# mounts and the project name exactly as it does for a human `deploy.sh`, and
+# no agent container needs a docker CLI. It is deliberately independent of
+# node-agent (own rsync, own result push) so it can redeploy node-agent itself —
+# the solaria case, where node-agent has been docker-blind for weeks.
+#
+# It NEVER runs `git pull`: a redeploy reconciles the node to the checkout it
+# already has. Shipping new code stays a human `scripts/deploy/deploy.sh` action.
+#
+# Config comes from the environment (systemd EnvironmentFile) — see env.example.
+# Install: see README.md.
+
+# NOT set -e: one malformed or failing action must be reported and skipped, not
+# abort the whole run.
+set -uo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+HELPER="${SCRIPT_DIR}/action.py"
+
+NODE_NAME="${NODE_NAME:-}"
+REPO_PATH="${REPO_PATH:-${HOME}/homelab-codex-ws}"
+RUNTIME_PATH="${RUNTIME_PATH:-/opt/homelab}"
+VPS_EVENTS_HOST="${VPS_EVENTS_HOST:-}"
+VPS_EVENTS_USER="${VPS_EVENTS_USER:-oskar}"
+VPS_EVENTS_PATH="${VPS_EVENTS_PATH:-/opt/homelab/events}"
+VPS_DEPLOY_PATH="${VPS_DEPLOY_PATH:-/opt/homelab/actions/deploy}"
+DEPLOY_TIMEOUT_SECS="${DEPLOY_TIMEOUT_SECS:-600}"
+
+INBOX="${RUNTIME_PATH}/actions/deploy/${NODE_NAME}"
+EVENTS_DIR="${RUNTIME_PATH}/events/${NODE_NAME}"
+STATE_DIR="${RUNTIME_PATH}/state"
+MARKER_DIR="${STATE_DIR}/processed-deploy-actions"
+LOG_DIR="${RUNTIME_PATH}/logs/deploy-runner"
+HOST_DIR="${REPO_PATH}/hosts/${NODE_NAME}"
+
+# Same ssh settings node-agent uses for its rsync channel: no ~/.ssh/config, no
+# host-key prompts, batch mode so a missing key fails fast instead of hanging.
+SSH_CMD="ssh -F /dev/null -o StrictHostKeyChecking=no -o UserKnownHostsFile=/dev/null -o ConnectTimeout=10 -o BatchMode=yes"
+if [[ -n "${SSH_KEY:-}" ]]; then
+    SSH_CMD="${SSH_CMD} -i ${SSH_KEY}"
+fi
+
+log() { echo "$(date -u +%Y-%m-%dT%H:%M:%SZ) [deploy-runner] $*"; }
+die() { log "ERROR: $*"; exit 1; }
+
+[[ -n "$NODE_NAME" ]] || die "NODE_NAME is not set (see env.example)"
+[[ -d "$REPO_PATH" ]] || die "repo not found at ${REPO_PATH} (set REPO_PATH)"
+[[ -d "$HOST_DIR" ]]  || die "no hosts/${NODE_NAME}/ directory in ${REPO_PATH}"
+[[ -x "${REPO_PATH}/scripts/deploy/deploy-service.sh" ]] \
+    || die "scripts/deploy/deploy-service.sh missing or not executable in ${REPO_PATH}"
+command -v python3 >/dev/null || die "python3 not found"
+
+mkdir -p "$INBOX" "$EVENTS_DIR" "$MARKER_DIR" "$LOG_DIR" || die "cannot create runtime dirs under ${RUNTIME_PATH}"
+
+# Single instance: the timer fires every 60s but a deploy may take minutes.
+exec 9>"${STATE_DIR}/deploy-runner.lock" || die "cannot open lock file"
+if ! flock -n 9; then
+    log "another deploy-runner run is in progress — exiting"
+    exit 0
+fi
+
+# ── transport ────────────────────────────────────────────────────────────────
+
+# Remote nodes fetch their inbox from the VPS. On the VPS itself the executor
+# writes straight into the shared /opt/homelab mount, so there is nothing to
+# pull and nothing to push — the executor reads the result event in place.
+is_remote() { [[ -n "$VPS_EVENTS_HOST" ]]; }
+
+pull_actions() {
+    is_remote || return 0
+    local rc=0
+    rsync -az --remove-source-files \
+        --omit-dir-times --no-perms --no-owner --no-group \
+        -e "$SSH_CMD" \
+        "${VPS_EVENTS_USER}@${VPS_EVENTS_HOST}:${VPS_DEPLOY_PATH}/${NODE_NAME}/" \
+        "${INBOX}/" || rc=$?
+    # 23/24 = "partial transfer"/"vanished source files": what an empty or
+    # concurrently-drained remote inbox looks like. Not worth logging each run.
+    if [[ $rc -ne 0 && $rc -ne 23 && $rc -ne 24 ]]; then
+        log "WARN: dispatch pull failed (rsync rc=${rc})"
+    fi
+}
+
+push_event() {
+    local event_file="$1"
+    is_remote || return 0
+    local rc=0
+    rsync -az --remove-source-files \
+        --omit-dir-times --no-perms --no-owner --no-group \
+        -e "$SSH_CMD" \
+        "$event_file" \
+        "${VPS_EVENTS_USER}@${VPS_EVENTS_HOST}:${VPS_EVENTS_PATH}/${NODE_NAME}/" || rc=$?
+    if [[ $rc -ne 0 && $rc -ne 23 && $rc -ne 24 ]]; then
+        # Left in place on failure: node-agent's own event shipping will pick it
+        # up on its next cycle, so a result is delayed, never lost.
+        log "WARN: result push failed (rsync rc=${rc}) — leaving ${event_file} for node-agent to ship"
+    fi
+}
+
+report() {
+    local action_id="$1" service="$2" success="$3" error="$4"
+    local event_file
+    event_file=$(python3 "$HELPER" emit-result \
+        --events-dir "$EVENTS_DIR" \
+        --node "$NODE_NAME" \
+        --action-id "$action_id" \
+        --service "$service" \
+        --success "$success" \
+        --error "$error") || { log "ERROR: could not write result event for ${action_id}"; return 1; }
+    log "reported ${action_id}: success=${success}${error:+ error=${error}}"
+    push_event "$event_file"
+}
+
+# ── main ─────────────────────────────────────────────────────────────────────
+
+pull_actions
+
+shopt -s nullglob
+for action_file in "${INBOX}"/*.json; do
+    OK="false"; ACTION_ID=""; SERVICE=""; ERROR=""
+    eval "$(python3 "$HELPER" validate --file "$action_file" --node "$NODE_NAME" --repo "$REPO_PATH")"
+
+    if [[ "$OK" != "true" ]]; then
+        log "rejected $(basename "$action_file"): ${ERROR}"
+        if [[ -n "$ACTION_ID" ]]; then
+            report "$ACTION_ID" "$SERVICE" "false" "$ERROR"
+        fi
+        rm -f "$action_file"
+        continue
+    fi
+
+    marker="${MARKER_DIR}/${ACTION_ID}.done"
+    if [[ -e "$marker" ]]; then
+        log "action ${ACTION_ID} already processed — skipping (idempotency)"
+        rm -f "$action_file"
+        continue
+    fi
+
+    log "deploying ${SERVICE} for action ${ACTION_ID}"
+    action_log="${LOG_DIR}/${ACTION_ID}-$(date -u +%Y%m%dT%H%M%SZ).log"
+    rc=0
+    timeout "$DEPLOY_TIMEOUT_SECS" \
+        "${REPO_PATH}/scripts/deploy/deploy-service.sh" \
+        --repo "$REPO_PATH" \
+        --host-dir "$HOST_DIR" \
+        --service "$SERVICE" \
+        --force-recreate \
+        >"$action_log" 2>&1 || rc=$?
+
+    case "$rc" in
+        0)
+            report "$ACTION_ID" "$SERVICE" "true" ""
+            ;;
+        3)
+            report "$ACTION_ID" "$SERVICE" "false" \
+                "${SERVICE} owns its deploy path (services/${SERVICE}/deploy-local.sh) — needs an operator deploy, not an automated redeploy"
+            ;;
+        124)
+            report "$ACTION_ID" "$SERVICE" "false" \
+                "deploy timed out after ${DEPLOY_TIMEOUT_SECS}s (log: ${action_log})"
+            ;;
+        *)
+            # Last 20 lines are the useful part of a compose failure; the full
+            # output stays in the per-action log on the node.
+            err=$(tail -n 20 "$action_log" 2>/dev/null | tr -d '\000' | tail -c 2000)
+            report "$ACTION_ID" "$SERVICE" "false" \
+                "deploy-service.sh exited ${rc} (log: ${action_log}): ${err}"
+            ;;
+    esac
+
+    touch "$marker" || log "WARN: could not write idempotency marker ${marker}"
+    rm -f "$action_file"
+done
+
+exit 0
--- a/jobs/deploy-runner/env.example
+++ b/jobs/deploy-runner/env.example
@ -0,0 +1,29 @@
+# deploy-runner runtime config — copy to /opt/homelab/config/deploy-runner/env
+# on each node and fill in. Read by homelab-deploy-runner.service
+# (EnvironmentFile), so: KEY=value, no quotes, no shell expansion.
+
+# Canonical topology node name — MUST match hosts/<node>/ and the `node` field
+# in dispatched actions. Not the OS hostname.
+NODE_NAME=piha
+
+# Repo checkout on this node (the runner never pulls it; deploys what is there).
+REPO_PATH=/home/oskar/homelab-codex-ws
+
+RUNTIME_PATH=/opt/homelab
+
+# VPS Tailscale IP — where the control plane writes dispatched actions and
+# expects result events. LEAVE EMPTY ON THE VPS ITSELF: there the executor
+# writes into the same /opt/homelab mount and no rsync hop is needed.
+VPS_EVENTS_HOST=100.95.58.48
+VPS_EVENTS_USER=oskar
+VPS_EVENTS_PATH=/opt/homelab/events
+VPS_DEPLOY_PATH=/opt/homelab/actions/deploy
+
+# Per-action wall-clock ceiling. Must stay below the executor's
+# REDEPLOY_TIMEOUT_SECS (default 900) so a stuck deploy is reported by the node
+# rather than silently timed out by the control plane.
+DEPLOY_TIMEOUT_SECS=600
+
+# Optional: explicit ssh key for the rsync channel. Omit when the service user's
+# default key (~/.ssh/id_*) already reaches the VPS — same key node-agent uses.
+# SSH_KEY=/home/oskar/.ssh/id_ed25519
--- a/jobs/deploy-runner/systemd/homelab-deploy-runner.service
+++ b/jobs/deploy-runner/systemd/homelab-deploy-runner.service
@ -0,0 +1,25 @@
+# homelab-deploy-runner.service — host-side executor for control-plane `redeploy`
+# actions (docs/architecture/RECON-multiagent-2026-07-27.md D14/D15).
+#
+# Host-level and NOT a container, deliberately: it runs `docker compose` the way
+# a human deploy does, so relative bind mounts and the compose project name
+# resolve identically. A containerised runner would resolve them against its own
+# filesystem and hand the daemon host paths that do not exist.
+#
+# Install: see ../README.md (copy this + the timer to /etc/systemd/system/,
+# create /opt/homelab/config/deploy-runner/env from env.example, daemon-reload,
+# enable the TIMER — not this unit).
+[Unit]
+Description=Homelab deploy runner (executes approved redeploy actions on this node)
+After=network-online.target docker.service
+Wants=network-online.target
+
+[Service]
+Type=oneshot
+User=oskar
+# Runtime config: NODE_NAME, REPO_PATH, VPS_EVENTS_HOST, ... (see env.example).
+EnvironmentFile=/opt/homelab/config/deploy-runner/env
+ExecStart=/home/oskar/homelab-codex-ws/jobs/deploy-runner/deploy-runner.sh
+# Slightly above the runner's own per-action DEPLOY_TIMEOUT_SECS (default 600)
+# so the script always gets to report a result before systemd intervenes.
+TimeoutStartSec=900
--- a/jobs/deploy-runner/systemd/homelab-deploy-runner.timer
+++ b/jobs/deploy-runner/systemd/homelab-deploy-runner.timer
@ -0,0 +1,16 @@
+# homelab-deploy-runner.timer — poll for dispatched redeploy actions.
+#
+# 60s matches node-agent's CHECK_INTERVAL, so remediation latency is the same on
+# both paths and well inside the executor's REDEPLOY_TIMEOUT_SECS (900).
+# Overlapping runs are prevented by the runner's own flock, not by the timer.
+[Unit]
+Description=Poll for dispatched redeploy actions every minute
+
+[Timer]
+OnBootSec=2min
+OnUnitActiveSec=60s
+AccuracySec=15s
+Unit=homelab-deploy-runner.service
+
+[Install]
+WantedBy=timers.target
--- a/jobs/deploy-runner/tests/test_action.py
+++ b/jobs/deploy-runner/tests/test_action.py
@ -0,0 +1,238 @@
+"""Tests for the deploy runner's validation / result-reporting helper.
+
+The validator is the only gate between a JSON file that arrived over the
+network and `docker compose up` running on the node, so its rejection cases
+matter more than its happy path. The emit-result tests pin the event format
+against the control-plane executor's actual parser — if either side drifts,
+redeploy actions would hang in running/ until they time out.
+"""
+from __future__ import annotations
+
+import json
+import shlex
+import sys
+import time
+from pathlib import Path
+
+import pytest
+
+_JOB_DIR = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(_JOB_DIR))
+
+import action as action_mod  # noqa: E402
+from action import Invalid, emit_result, validate  # noqa: E402
+
+
+@pytest.fixture
+def repo(tmp_path: Path) -> Path:
+    """Minimal repo: node piha declares vikunja, which has a compose file."""
+    repo = tmp_path / "repo"
+    (repo / "hosts" / "piha").mkdir(parents=True)
+    (repo / "hosts" / "piha" / "services.yaml").write_text(
+        "host: piha\nservices:\n  vikunja:\n    role: task-tracker\n  node-agent:\n    role: monitor\n"
+    )
+    (repo / "services" / "vikunja").mkdir(parents=True)
+    (repo / "services" / "vikunja" / "docker-compose.yml").write_text("services: {}\n")
+    (repo / "services" / "gokapi").mkdir(parents=True)
+    (repo / "services" / "gokapi" / "docker-compose.yml").write_text("services: {}\n")
+    return repo
+
+
+def _action_file(tmp_path: Path, **overrides) -> Path:
+    action = {
+        "action_id": "redeploy-piha-vikunja",
+        "type": "redeploy",
+        "node": "piha",
+        "service": "vikunja",
+    }
+    action.update(overrides)
+    path = tmp_path / "redeploy-piha-vikunja.json"
+    path.write_text(json.dumps(action))
+    return path
+
+
+# ---------------------------------------------------------------------------
+# validate
+# ---------------------------------------------------------------------------
+
+def test_valid_action_returns_id_and_service(tmp_path, repo):
+    action_id, service = validate(_action_file(tmp_path), "piha", repo)
+    assert action_id == "redeploy-piha-vikunja"
+    assert service == "vikunja"
+
+
+def test_action_for_another_node_is_refused(tmp_path, repo):
+    path = _action_file(tmp_path, node="solaria")
+    with pytest.raises(Invalid, match="addressed to node"):
+        validate(path, "piha", repo)
+
+
+def test_non_redeploy_type_is_refused(tmp_path, repo):
+    """The runner executes redeploys only — container_restart stays node-agent's
+    job, and anything else must never reach a shell."""
+    path = _action_file(tmp_path, type="container_restart")
+    with pytest.raises(Invalid, match="not executable by the deploy runner"):
+        validate(path, "piha", repo)
+
+
+def test_service_not_in_desired_state_is_refused(tmp_path, repo):
+    """gokapi has a compose file in the repo but is not in hosts/piha —
+    desired state is the authority on what this node may run."""
+    path = _action_file(tmp_path, service="gokapi")
+    with pytest.raises(Invalid, match="not in hosts/piha/services.yaml"):
+        validate(path, "piha", repo)
+
+
+@pytest.mark.parametrize("bad", ["../../etc/passwd", "/absolute", "Vikunja", "", "a b", "x" * 65])
+def test_malformed_service_names_are_refused(tmp_path, repo, bad):
+    path = _action_file(tmp_path, service=bad)
+    with pytest.raises(Invalid, match="invalid service name"):
+        validate(path, "piha", repo)
+
+
+def test_service_without_compose_file_is_refused(tmp_path, repo):
+    (repo / "services" / "vikunja" / "docker-compose.yml").unlink()
+    with pytest.raises(Invalid, match="no compose file"):
+        validate(_action_file(tmp_path), "piha", repo)
+
+
+def test_missing_manifest_for_own_node_is_refused(tmp_path, repo):
+    (repo / "hosts" / "piha" / "services.yaml").unlink()
+    with pytest.raises(Invalid, match="no desired-state manifest"):
+        validate(_action_file(tmp_path), "piha", repo)
+
+
+def test_services_as_list_is_supported(tmp_path, repo):
+    (repo / "hosts" / "piha" / "services.yaml").write_text("host: piha\nservices:\n  - vikunja\n")
+    _, service = validate(_action_file(tmp_path), "piha", repo)
+    assert service == "vikunja"
+
+
+def test_unparseable_file_is_refused(tmp_path, repo):
+    path = tmp_path / "broken.json"
+    path.write_text("{not json")
+    with pytest.raises(Invalid, match="unparseable"):
+        validate(path, "piha", repo)
+
+
+def test_invalid_action_id_is_refused(tmp_path, repo):
+    path = _action_file(tmp_path, action_id="../escape")
+    with pytest.raises(Invalid, match="invalid action_id"):
+        validate(path, "piha", repo)
+
+
+# ---------------------------------------------------------------------------
+# validate via the CLI (shell-assignment contract used by deploy-runner.sh)
+# ---------------------------------------------------------------------------
+
+def _eval_cli(argv) -> dict:
+    """Parse the KEY='value' lines the bash runner evaluates."""
+    import io
+    import contextlib
+
+    buf = io.StringIO()
+    with contextlib.redirect_stdout(buf):
+        rc = action_mod.main(argv)
+    assert rc == 0
+    out = {}
+    for line in buf.getvalue().splitlines():
+        key, _, raw = line.partition("=")
+        out[key] = shlex.split(raw)[0] if raw and raw != "''" else ""
+    return out
+
+
+def test_cli_valid_action_emits_ok_true(tmp_path, repo):
+    path = _action_file(tmp_path)
+    result = _eval_cli(["validate", "--file", str(path), "--node", "piha", "--repo", str(repo)])
+    assert result["OK"] == "true"
+    assert result["SERVICE"] == "vikunja"
+    assert result["ACTION_ID"] == "redeploy-piha-vikunja"
+    assert result["ERROR"] == ""
+
+
+def test_cli_rejected_action_still_reports_action_id(tmp_path, repo):
+    """So the runner can report a failure result instead of silently dropping
+    the action — a dropped action would sit in running/ until it times out."""
+    path = _action_file(tmp_path, service="gokapi")
+    result = _eval_cli(["validate", "--file", str(path), "--node", "piha", "--repo", str(repo)])
+    assert result["OK"] == "false"
+    assert result["ACTION_ID"] == "redeploy-piha-vikunja"
+    assert "not in hosts/piha/services.yaml" in result["ERROR"]
+
+
+def test_cli_quotes_hostile_error_content(tmp_path, repo):
+    """A crafted service name must not be able to inject shell syntax through
+    the eval'd ERROR assignment."""
+    path = _action_file(tmp_path, service="'; touch /tmp/pwned; '")
+    result = _eval_cli(["validate", "--file", str(path), "--node", "piha", "--repo", str(repo)])
+    assert result["OK"] == "false"
+    assert "touch /tmp/pwned" in result["ERROR"]  # inert data, not a command
+
+
+# ---------------------------------------------------------------------------
+# emit-result — format compatibility with the control-plane executor
+# ---------------------------------------------------------------------------
+
+def test_emit_result_shape(tmp_path):
+    events = tmp_path / "events" / "piha"
+    path = emit_result(events, "piha", "redeploy-piha-vikunja", "vikunja", True, "")
+    event = json.loads(path.read_text())
+    assert event["type"] == "action_result"
+    assert event["node"] == "piha"
+    assert event["severity"] == "info"
+    assert event["payload"]["action_id"] == "redeploy-piha-vikunja"
+    assert event["payload"]["success"] is True
+    assert event["payload"]["source"] == "deploy-runner"
+
+
+def test_emit_result_failure_carries_error(tmp_path):
+    events = tmp_path / "events" / "piha"
+    path = emit_result(events, "piha", "act-1", "vikunja", False, "compose exited 1")
+    event = json.loads(path.read_text())
+    assert event["severity"] == "high"
+    assert event["payload"]["success"] is False
+    assert event["payload"]["error"] == "compose exited 1"
+
+
+def test_executor_finds_the_emitted_result(tmp_path, monkeypatch):
+    """The real consumer: control-plane executor._find_action_result(). This is
+    the contract that makes a redeploy resolve instead of timing out."""
+    executor_src = _JOB_DIR.parents[1] / "services" / "control-plane" / "src"
+    sys.path.insert(0, str(executor_src))
+    import executor as executor_mod
+
+    events_root = tmp_path / "events"
+    monkeypatch.setattr(executor_mod, "EVENTS_DIR", events_root)
+    monkeypatch.setattr(executor_mod, "ACTIONS_DIR", tmp_path / "actions")
+    monkeypatch.setattr(executor_mod, "DISPATCH_DIR", tmp_path / "actions" / "dispatch")
+    monkeypatch.setattr(executor_mod, "DEPLOY_DISPATCH_DIR", tmp_path / "actions" / "deploy")
+
+    started_at = time.time()
+    emit_result(events_root / "piha", "piha", "redeploy-piha-vikunja", "vikunja", True, "")
+
+    payload = executor_mod.Executor()._find_action_result(
+        "piha", "redeploy-piha-vikunja", started_at
+    )
+    assert payload is not None
+    assert payload["success"] is True
+
+
+def test_executor_ignores_result_older_than_the_action(tmp_path, monkeypatch):
+    """Action IDs are deterministic and repeat, so a result from a previous run
+    must not resolve the current one."""
+    executor_src = _JOB_DIR.parents[1] / "services" / "control-plane" / "src"
+    sys.path.insert(0, str(executor_src))
+    import executor as executor_mod
+
+    events_root = tmp_path / "events"
+    monkeypatch.setattr(executor_mod, "EVENTS_DIR", events_root)
+    monkeypatch.setattr(executor_mod, "ACTIONS_DIR", tmp_path / "actions")
+    monkeypatch.setattr(executor_mod, "DISPATCH_DIR", tmp_path / "actions" / "dispatch")
+    monkeypatch.setattr(executor_mod, "DEPLOY_DISPATCH_DIR", tmp_path / "actions" / "deploy")
+
+    emit_result(events_root / "piha", "piha", "redeploy-piha-vikunja", "vikunja", True, "")
+    started_at = time.time() + 60  # action started after the event was written
+
+    assert executor_mod.Executor()._find_action_result(
+        "piha", "redeploy-piha-vikunja", started_at
+    ) is None
--- a/jobs/deploy-runner/tests/test_deploy_service.py
+++ b/jobs/deploy-runner/tests/test_deploy_service.py
@ -0,0 +1,150 @@
+"""Contract tests for scripts/deploy/deploy-service.sh.
+
+Why these live here: the deploy runner and the human whole-node deploy both go
+through this script, and what must stay identical between them is the argv it
+hands to `docker compose`. docker compose derives the project name from the
+directory of the FIRST -f file — a caller that assembles those arguments
+differently lands in a different project, where `--remove-orphans` tears down
+the running stack (the 2026-06-25 control-plane wipe on VPS).
+
+Each test runs the real script with a stub `docker` on PATH that records its
+argv, so nothing is executed against a daemon.
+"""
+from __future__ import annotations
+
+import os
+import subprocess
+from pathlib import Path
+
+import pytest
+
+REPO = Path(__file__).resolve().parents[3]
+SCRIPT = REPO / "scripts" / "deploy" / "deploy-service.sh"
+
+
+@pytest.fixture
+def env(tmp_path: Path):
+    """A fake repo + host dir + a `docker` stub that logs its arguments."""
+    repo = tmp_path / "repo"
+    (repo / "services" / "vikunja").mkdir(parents=True)
+    (repo / "services" / "vikunja" / "docker-compose.yml").write_text("services: {}\n")
+    host_dir = repo / "hosts" / "piha"
+    host_dir.mkdir(parents=True)
+
+    bindir = tmp_path / "bin"
+    bindir.mkdir()
+    argv_log = tmp_path / "docker-argv.txt"
+    stub = bindir / "docker"
+    stub.write_text(f'#!/usr/bin/env bash\nprintf "%s\\n" "$@" > {argv_log}\nexit 0\n')
+    stub.chmod(0o755)
+
+    return {"repo": repo, "host_dir": host_dir, "bindir": bindir, "argv_log": argv_log}
+
+
+def run(env, *args, service="vikunja"):
+    proc_env = dict(os.environ, PATH=f"{env['bindir']}:{os.environ['PATH']}")
+    return subprocess.run(
+        [str(SCRIPT), "--repo", str(env["repo"]), "--host-dir", str(env["host_dir"]),
+         "--service", service, *args],
+        capture_output=True, text=True, env=proc_env,
+    )
+
+
+def argv(env) -> list[str]:
+    return env["argv_log"].read_text().splitlines()
+
+
+def test_base_invocation(env):
+    result = run(env)
+    assert result.returncode == 0, result.stderr
+    assert argv(env) == [
+        "compose", "-f", str(env["repo"] / "services" / "vikunja" / "docker-compose.yml"),
+        "up", "-d",
+    ]
+
+
+def test_project_name_invariant(env):
+    """No -p and no --project-directory: the project name must come from the
+    compose file's own directory, exactly as it does for a human deploy."""
+    run(env, "--force-recreate")
+    assert "-p" not in argv(env)
+    assert "--project-directory" not in argv(env)
+    assert argv(env)[1] == "-f"  # first -f is the service's own compose file
+
+
+def test_override_is_second_f(env):
+    override = env["host_dir"] / "runtime" / "vikunja"
+    override.mkdir(parents=True)
+    (override / "docker-compose.override.yml").write_text("services: {}\n")
+
+    run(env)
+    args = argv(env)
+    assert args[1:5] == [
+        "-f", str(env["repo"] / "services" / "vikunja" / "docker-compose.yml"),
+        "-f", str(override / "docker-compose.override.yml"),
+    ]
+
+
+def test_env_file_is_passed_when_present(env):
+    (env["repo"] / "services" / "vikunja" / ".env").write_text("FOO=bar\n")
+    run(env)
+    args = argv(env)
+    assert "--env-file" in args
+    assert args[args.index("--env-file") + 1] == str(env["repo"] / "services" / "vikunja" / ".env")
+
+
+def test_force_recreate_flag(env):
+    """The remediation path needs it: plain `up -d` is a no-op when config is
+    unchanged, i.e. exactly the unhealthy-container case a redeploy targets."""
+    run(env, "--force-recreate")
+    assert "--force-recreate" in argv(env)
+
+
+def test_no_remove_orphans_by_default(env):
+    """The agent path must never be able to delete containers on a project-name
+    mismatch — only the human whole-node deploy opts in."""
+    run(env, "--force-recreate")
+    assert "--remove-orphans" not in argv(env)
+
+
+def test_remove_orphans_when_requested(env):
+    run(env, "--remove-orphans")
+    assert "--remove-orphans" in argv(env)
+
+
+def test_build_if_needed_only_builds_with_dockerfile(env):
+    run(env, "--build-if-needed")
+    assert "--build" not in argv(env)
+
+    (env["repo"] / "services" / "vikunja" / "Dockerfile").write_text("FROM scratch\n")
+    run(env, "--build-if-needed")
+    assert "--build" in argv(env)
+
+
+def test_redeploy_path_never_builds(env):
+    """The runner does not pass --build-if-needed: building on the 4 GiB VPS
+    mid-incident is an OOM risk, and a redeploy is a reconcile, not a code ship."""
+    (env["repo"] / "services" / "vikunja" / "Dockerfile").write_text("FROM scratch\n")
+    run(env, "--force-recreate")
+    assert "--build" not in argv(env)
+
+
+def test_service_with_own_deploy_path_exits_3(env):
+    (env["repo"] / "services" / "vikunja" / "deploy-local.sh").write_text("#!/bin/sh\n")
+    result = run(env)
+    assert result.returncode == 3
+    assert not env["argv_log"].exists()  # docker never invoked
+
+
+@pytest.mark.parametrize("bad", ["../../etc", "/abs", "Vikunja", "a b"])
+def test_invalid_service_name_exits_2(env, bad):
+    result = run(env, service=bad)
+    assert result.returncode == 2
+    assert not env["argv_log"].exists()
+
+
+def test_missing_compose_file_exits_2(env):
+    (env["repo"] / "services" / "vikunja" / "docker-compose.yml").unlink()
+    result = run(env)
+    assert result.returncode == 2
+    assert not env["argv_log"].exists()
--- a/jobs/deploy-runner/tests/test_runner_integration.py
+++ b/jobs/deploy-runner/tests/test_runner_integration.py
@ -0,0 +1,163 @@
+"""End-to-end test of deploy-runner.sh on a fake node.
+
+Runs the real bash runner against temp directories with a stubbed `docker`, in
+local (VPS) mode so no rsync/ssh is involved. Covers the full loop the control
+plane depends on: dispatch file in → validate → deploy → action_result event
+out → idempotency marker → inbox drained.
+"""
+from __future__ import annotations
+
+import json
+import os
+import shutil
+import subprocess
+from pathlib import Path
+
+import pytest
+
+REPO = Path(__file__).resolve().parents[3]
+RUNNER = REPO / "jobs" / "deploy-runner" / "deploy-runner.sh"
+
+
+@pytest.fixture
+def node(tmp_path: Path):
+    """A fake node: repo checkout, /opt/homelab-alike runtime, docker stub."""
+    repo = tmp_path / "repo"
+    (repo / "scripts" / "deploy").mkdir(parents=True)
+    shutil.copy(REPO / "scripts" / "deploy" / "deploy-service.sh",
+                repo / "scripts" / "deploy" / "deploy-service.sh")
+    os.chmod(repo / "scripts" / "deploy" / "deploy-service.sh", 0o755)
+
+    (repo / "hosts" / "piha").mkdir(parents=True)
+    (repo / "hosts" / "piha" / "services.yaml").write_text(
+        "host: piha\nservices:\n  vikunja:\n    role: task-tracker\n"
+    )
+    (repo / "services" / "vikunja").mkdir(parents=True)
+    (repo / "services" / "vikunja" / "docker-compose.yml").write_text("services: {}\n")
+
+    runtime = tmp_path / "runtime"
+    (runtime / "actions" / "deploy" / "piha").mkdir(parents=True)
+
+    bindir = tmp_path / "bin"
+    bindir.mkdir()
+    argv_log = tmp_path / "docker-argv.txt"
+    (bindir / "docker").write_text(
+        f'#!/usr/bin/env bash\nprintf "%s " "$@" >> {argv_log}\necho >> {argv_log}\n'
+        f'exit ${{DOCKER_STUB_EXIT:-0}}\n'
+    )
+    os.chmod(bindir / "docker", 0o755)
+
+    return {"repo": repo, "runtime": runtime, "bindir": bindir, "argv_log": argv_log,
+            "inbox": runtime / "actions" / "deploy" / "piha",
+            "events": runtime / "events" / "piha"}
+
+
+def dispatch(fake_node, action_id="redeploy-piha-vikunja", **overrides):
+    action = {"action_id": action_id, "type": "redeploy", "node": "piha",
+              "service": "vikunja", "dispatched_at": 0}
+    action.update(overrides)
+    (fake_node["inbox"] / f"{action_id}.json").write_text(json.dumps(action))
+
+
+def run_runner(fake_node, **extra_env):
+    env = dict(os.environ)
+    env.update({
+        "PATH": f"{fake_node['bindir']}:{os.environ['PATH']}",
+        "NODE_NAME": "piha",
+        "REPO_PATH": str(fake_node["repo"]),
+        "RUNTIME_PATH": str(fake_node["runtime"]),
+        "VPS_EVENTS_HOST": "",  # local mode: executor shares the same filesystem
+    })
+    env.update(extra_env)
+    return subprocess.run([str(RUNNER)], capture_output=True, text=True, env=env)
+
+
+def results(node) -> list[dict]:
+    if not node["events"].exists():
+        return []
+    return [json.loads(p.read_text())
+            for p in sorted(node["events"].glob("evt-*-action_result-*.json"))]
+
+
+def test_successful_redeploy_reports_and_drains(node):
+    dispatch(node)
+    proc = run_runner(node)
+    assert proc.returncode == 0, proc.stdout + proc.stderr
+
+    assert "compose -f" in node["argv_log"].read_text()
+    assert "--force-recreate" in node["argv_log"].read_text()
+
+    events = results(node)
+    assert len(events) == 1
+    assert events[0]["payload"] == {
+        "action_id": "redeploy-piha-vikunja", "success": True, "error": "",
+        "node": "piha", "source": "deploy-runner",
+    }
+
+    assert list(node["inbox"].glob("*.json")) == []
+    assert (node["runtime"] / "state" / "processed-deploy-actions"
+            / "redeploy-piha-vikunja.done").exists()
+
+
+def test_failed_deploy_reports_compose_output(node):
+    dispatch(node)
+    proc = run_runner(node, DOCKER_STUB_EXIT="1")
+    assert proc.returncode == 0  # the runner itself succeeds; the action fails
+
+    event = results(node)[0]
+    assert event["payload"]["success"] is False
+    assert "deploy-service.sh exited 1" in event["payload"]["error"]
+    assert event["severity"] == "high"
+    # Full output kept on the node for forensics
+    assert list((node["runtime"] / "logs" / "deploy-runner").glob("*.log"))
+
+
+def test_rejected_action_is_reported_not_silently_dropped(node):
+    """An action for a service this node does not declare must come back as a
+    failure — otherwise it sits in running/ until the executor times it out."""
+    dispatch(node, service="gokapi")
+    run_runner(node)
+
+    event = results(node)[0]
+    assert event["payload"]["success"] is False
+    assert "not in hosts/piha/services.yaml" in event["payload"]["error"]
+    assert not node["argv_log"].exists()  # docker never invoked
+    assert list(node["inbox"].glob("*.json")) == []
+
+
+def test_action_for_another_node_is_refused(node):
+    dispatch(node, node="solaria")
+    run_runner(node)
+
+    event = results(node)[0]
+    assert "addressed to node" in event["payload"]["error"]
+    assert not node["argv_log"].exists()
+
+
+def test_container_restart_is_not_executed_by_the_runner(node):
+    """node-agent's job. The runner refuses anything but redeploy."""
+    dispatch(node, type="container_restart")
+    run_runner(node)
+
+    event = results(node)[0]
+    assert "not executable by the deploy runner" in event["payload"]["error"]
+    assert not node["argv_log"].exists()
+
+
+def test_already_processed_action_is_not_redeployed(node):
+    dispatch(node)
+    run_runner(node)
+    first = len(results(node))
+
+    dispatch(node)  # same action_id arrives again
+    run_runner(node)
+
+    assert len(results(node)) == first  # no second result
+    assert len(node["argv_log"].read_text().splitlines()) == 1  # docker ran once
+    assert list(node["inbox"].glob("*.json")) == []
+
+
+def test_missing_node_name_fails_loudly(node):
+    proc = run_runner(node, NODE_NAME="")
+    assert proc.returncode == 1
+    assert "NODE_NAME is not set" in proc.stdout + proc.stderr
--- a/scripts/deploy/deploy-node.sh
+++ b/scripts/deploy/deploy-node.sh
@ -6,7 +6,9 @@ set -e

 # Configuration
 REPO_PATH="${HOME}/homelab-codex-ws"
-RUNTIME_PATH="/opt/homelab"
+# Env-overridable so the script can be exercised against a scratch runtime dir
+# without writing into the node's real /opt/homelab.
+RUNTIME_PATH="${RUNTIME_PATH:-/opt/homelab}"
 HOSTNAME=$(hostname | tr '[:lower:]' '[:upper:]')
 CURRENT_OS_HOST=$(hostname)

@ -81,57 +83,33 @@ if [ ${#SERVICES[@]} -eq 0 ]; then
 fi

 # 3. Deploy Services
+#
+# The per-service compose invocation lives in deploy-service.sh, shared with the
+# agent-driven single-service redeploy (services/deploy-runner/). Both callers
+# MUST assemble the -f arguments identically: docker compose derives the project
+# name from the first -f file's directory, and a mismatch plus --remove-orphans
+# is what wiped the control-plane on VPS on 2026-06-25. Services with their own
+# deploy-local.sh are skipped by deploy-service.sh (exit 3).
+DEPLOY_SERVICE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)/deploy-service.sh"
+
 for service in "${SERVICES[@]}"; do
-    # Services that ship their own deploy-local.sh have a dedicated, orchestrated
-    # deploy path (e.g. control-plane via deploy-control-plane.sh → deploy-local.sh).
-    # Deploying them in this loop too would invoke `docker compose up` with a
-    # different Compose project name than their dedicated path uses (this loop runs
-    # from ${REPO_PATH}; deploy-local.sh runs from services/<svc>/), so `up
-    # -d --remove-orphans` would Recreate and tear down the already-running
-    # containers — exactly what wiped the control-plane on VPS on 2026-06-25.
-    # Their own deploy path owns them; skip here.
-    if [ -f "${REPO_PATH}/services/${service}/deploy-local.sh" ]; then
-        echo "Skipping ${service}: ma własną ścieżkę deployu (services/${service}/deploy-local.sh), pomijam"
-        continue
-    fi
-
-    echo "Deploying service: ${service}..."
-
-    COMPOSE_FILE="${REPO_PATH}/services/${service}/docker-compose.yml"
-
-    if [ ! -f "$COMPOSE_FILE" ]; then
-        echo "Warning: Compose file not found for ${service} at ${COMPOSE_FILE}"
-        continue
-    fi
-
    TARGET_DIR="${RUNTIME_PATH}/services/${service}"
    mkdir -p "$TARGET_DIR"

-    OVERRIDE_FILE="${HOST_DIR}/runtime/${service}/docker-compose.override.yml"
+    rc=0
+    "$DEPLOY_SERVICE" \
+        --repo "$REPO_PATH" \
+        --host-dir "$HOST_DIR" \
+        --service "$service" \
+        --build-if-needed \
+        --remove-orphans || rc=$?

-    COMPOSE_CMD="docker compose -f ${COMPOSE_FILE}"
-    if [ -f "$OVERRIDE_FILE" ]; then
-        echo "Using override file for ${service}"
-        COMPOSE_CMD="${COMPOSE_CMD} -f ${OVERRIDE_FILE}"
-    fi
-
-    ENV_FILE="${REPO_PATH}/services/${service}/.env"
-    if [ -f "$ENV_FILE" ]; then
-        COMPOSE_CMD="${COMPOSE_CMD} --env-file ${ENV_FILE}"
-    fi
-
-    # Services with their own Dockerfile need a rebuild on every deploy —
-    # otherwise `up -d` sees "container running, image tag unchanged" and skips
-    # rebuilding even when src/ changed, silently leaving the old code running
-    # (backlog c858dbc). Services with a prebuilt image (no Dockerfile) don't
-    # need --build.
-    if [ -f "${REPO_PATH}/services/${service}/Dockerfile" ]; then
-        echo "Building ${service} (has Dockerfile)"
-        $COMPOSE_CMD up -d --build --remove-orphans
-    else
-        echo "Using prebuilt image for ${service}"
-        $COMPOSE_CMD up -d --remove-orphans
-    fi
+    case "$rc" in
+        0) ;;                       # deployed
+        3) ;;                       # owns its deploy path — already reported, skip
+        2) echo "Warning: skipping ${service} (validation failed, see above)" ;;
+        *) exit "$rc" ;;            # real deploy failure — preserve set -e semantics
+    esac
 done

 echo "--- Deployment Complete ---"
--- a/scripts/deploy/deploy-service.sh
+++ b/scripts/deploy/deploy-service.sh
@ -0,0 +1,134 @@
+#!/usr/bin/env bash
+# scripts/deploy/deploy-service.sh — deploy exactly ONE service on the local node.
+#
+# Extracted from deploy-node.sh so that the human whole-node deploy and the
+# agent-driven single-service redeploy (services/deploy-runner/) share ONE
+# compose invocation. That sharing is the whole point: docker compose derives
+# the PROJECT NAME from the directory of the first -f file (verified:
+# `docker compose -f services/node-agent/docker-compose.yml config` → name
+# "node-agent", independent of cwd; adding --project-directory changes it).
+# A caller that assembles the -f arguments differently lands in a DIFFERENT
+# project, where `--remove-orphans` tears down the already-running stack —
+# exactly what wiped the control-plane on VPS on 2026-06-25. Every caller goes
+# through here.
+#
+# Usage:
+#   deploy-service.sh --repo <repo-path> --host-dir <hosts/<node> path>
+#                     --service <name>
+#                     [--build-if-needed] [--force-recreate] [--remove-orphans]
+#
+# Flags:
+#   --build-if-needed  add --build when the service ships its own Dockerfile
+#                      (whole-node deploy path — see the rationale below)
+#   --force-recreate   recreate containers even when config is unchanged
+#                      (remediation path: plain `up -d` is a no-op for an
+#                      unhealthy-but-unchanged service, i.e. it would "succeed"
+#                      without doing anything)
+#   --remove-orphans   human whole-node deploy only; never used by the agent
+#                      path, where a project-name mismatch must not be able to
+#                      delete containers
+#
+# Exit codes:
+#   0  deployed
+#   2  usage / validation error
+#   3  service owns its deploy path (services/<svc>/deploy-local.sh) — skipped
+#   *  docker compose exit code
+
+set -euo pipefail
+
+REPO_PATH=""
+HOST_DIR=""
+SERVICE=""
+BUILD_IF_NEEDED=false
+FORCE_RECREATE=false
+REMOVE_ORPHANS=false
+
+usage() {
+    sed -n '2,30p' "${BASH_SOURCE[0]}" >&2
+    exit 2
+}
+
+while [[ $# -gt 0 ]]; do
+    case "$1" in
+        --repo)            REPO_PATH="${2:-}"; shift 2 ;;
+        --host-dir)        HOST_DIR="${2:-}"; shift 2 ;;
+        --service)         SERVICE="${2:-}"; shift 2 ;;
+        --build-if-needed) BUILD_IF_NEEDED=true; shift ;;
+        --force-recreate)  FORCE_RECREATE=true; shift ;;
+        --remove-orphans)  REMOVE_ORPHANS=true; shift ;;
+        -h|--help)         usage ;;
+        *) echo "deploy-service.sh: unknown argument: $1" >&2; usage ;;
+    esac
+done
+
+[[ -n "$REPO_PATH" && -n "$HOST_DIR" && -n "$SERVICE" ]] || {
+    echo "deploy-service.sh: --repo, --host-dir and --service are required" >&2
+    usage
+}
+
+# Service names are used to build filesystem paths. Reject anything that is not
+# a plain kebab-case service name so a malformed (or maliciously crafted)
+# dispatched action can never escape services/ via ../.
+if [[ ! "$SERVICE" =~ ^[a-z0-9][a-z0-9._-]{0,63}$ ]]; then
+    echo "deploy-service.sh: invalid service name: '${SERVICE}'" >&2
+    exit 2
+fi
+
+[[ -d "$REPO_PATH" ]] || { echo "deploy-service.sh: repo not found: ${REPO_PATH}" >&2; exit 2; }
+[[ -d "$HOST_DIR" ]]  || { echo "deploy-service.sh: host dir not found: ${HOST_DIR}" >&2; exit 2; }
+
+SERVICE_DIR="${REPO_PATH}/services/${SERVICE}"
+COMPOSE_FILE="${SERVICE_DIR}/docker-compose.yml"
+
+if [[ ! -f "$COMPOSE_FILE" ]]; then
+    echo "deploy-service.sh: compose file not found for ${SERVICE} at ${COMPOSE_FILE}" >&2
+    exit 2
+fi
+
+# Services that ship their own deploy-local.sh have a dedicated, orchestrated
+# deploy path (e.g. control-plane via deploy-control-plane.sh → deploy-local.sh)
+# which does host-level preparation (dirs, ownership, sudo) this script cannot
+# do. Deploying them from here would also run `up` against a stack whose own
+# path expects to own it. Their own deploy path owns them; skip.
+if [[ -f "${SERVICE_DIR}/deploy-local.sh" ]]; then
+    echo "Skipping ${SERVICE}: ma własną ścieżkę deployu (services/${SERVICE}/deploy-local.sh), pomijam"
+    exit 3
+fi
+
+COMPOSE_ARGS=(docker compose -f "$COMPOSE_FILE")
+
+OVERRIDE_FILE="${HOST_DIR}/runtime/${SERVICE}/docker-compose.override.yml"
+if [[ -f "$OVERRIDE_FILE" ]]; then
+    echo "Using override file for ${SERVICE}"
+    COMPOSE_ARGS+=(-f "$OVERRIDE_FILE")
+fi
+
+ENV_FILE="${SERVICE_DIR}/.env"
+if [[ -f "$ENV_FILE" ]]; then
+    COMPOSE_ARGS+=(--env-file "$ENV_FILE")
+fi
+
+COMPOSE_ARGS+=(up -d)
+
+# Services with their own Dockerfile need a rebuild on a whole-node deploy —
+# otherwise `up -d` sees "container running, image tag unchanged" and skips
+# rebuilding even when src/ changed, silently leaving the old code running
+# (backlog c858dbc). The agent redeploy path deliberately does NOT pass
+# --build-if-needed: a redeploy is a reconcile, not a code ship, and building
+# on the 4 GiB VPS mid-incident is an OOM risk.
+if [[ "$BUILD_IF_NEEDED" == "true" && -f "${SERVICE_DIR}/Dockerfile" ]]; then
+    echo "Building ${SERVICE} (has Dockerfile)"
+    COMPOSE_ARGS+=(--build)
+fi
+
+if [[ "$FORCE_RECREATE" == "true" ]]; then
+    COMPOSE_ARGS+=(--force-recreate)
+fi
+
+if [[ "$REMOVE_ORPHANS" == "true" ]]; then
+    COMPOSE_ARGS+=(--remove-orphans)
+fi
+
+echo "Deploying service: ${SERVICE}..."
+echo "+ ${COMPOSE_ARGS[*]}"
+"${COMPOSE_ARGS[@]}"
--- a/services/control-plane/docker-compose.yml
+++ b/services/control-plane/docker-compose.yml
@ -79,7 +79,11 @@ services:
    command: python src/executor.py
    volumes:
      - /opt/homelab:/opt/homelab
-      - ../..:/repo
+      # Read-only since the redeploy fix: the executor used to run
+      # scripts/deploy/deploy-node.sh out of this mount (it never worked — see
+      # jobs/deploy-runner/README.md). Deploys now happen on the node itself, so
+      # nothing here needs write access to the checkout.
+      - ../..:/repo:ro
      - /var/run/docker.sock:/var/run/docker.sock
    restart: unless-stopped
    environment:
--- a/services/control-plane/src/executor.py
+++ b/services/control-plane/src/executor.py
@ -21,6 +21,15 @@ RUNTIME_PATH = os.getenv("RUNTIME_PATH", "/opt/homelab")
 ACTIONS_DIR = Path(RUNTIME_PATH) / "actions"
 EVENTS_DIR = Path(RUNTIME_PATH) / "events"
 DISPATCH_DIR = ACTIONS_DIR / "dispatch"
+# Separate inbox for redeploy actions, consumed by the host-side deploy runner
+# (jobs/deploy-runner/) rather than by node-agent. It must NOT share
+# DISPATCH_DIR: node-agent's process_dispatched_actions() deletes and
+# failure-reports every file it finds in its inbox, so a redeploy landing there
+# would be killed before the runner ever saw it.
+DEPLOY_DISPATCH_DIR = ACTIONS_DIR / "deploy"
+# The executor no longer reads the repo at all (the old redeploy path ran a
+# script out of it). Kept only so an operator can still see which checkout the
+# container is wired to; nothing in this module resolves paths against it.
 REPO_ROOT = Path(os.getenv("REPO_ROOT", "/repo"))

 # SSH configuration
@ -41,6 +50,15 @@ SSH_OPTIONS = [
 # code change.
 ACTION_TIMEOUT_SECS = int(os.getenv("ACTION_TIMEOUT_SECS", "300"))

+# Redeploy gets its own, longer budget: the deploy runner polls once a minute
+# and a `docker compose up` may pull images before it can report anything.
+REDEPLOY_TIMEOUT_SECS = int(os.getenv("REDEPLOY_TIMEOUT_SECS", "900"))
+
+# Action types the executor hands to an on-node agent instead of resolving
+# synchronously. They stay in running/ until an action_result event arrives (or
+# the type's timeout expires) — see _reconcile_running_actions.
+DISPATCHED_ACTION_TYPES = {"container_restart", "redeploy"}
+
 # Matches evt-<node>-<unixts>-<type>-<svc>.json, same convention as
 # node_agent.py / observer.py / operator_ui.py.
 _EVENT_TS_RE = re.compile(r"-(\d{9,11})-")
@ -58,6 +76,7 @@ class Executor:
        for s in ["approved", "running", "completed", "failed", "rejected"]:
            (ACTIONS_DIR / s).mkdir(parents=True, exist_ok=True)
        DISPATCH_DIR.mkdir(parents=True, exist_ok=True)
+        DEPLOY_DISPATCH_DIR.mkdir(parents=True, exist_ok=True)

    def process_actions(self):
        # Update heartbeat
@ -104,19 +123,25 @@ class Executor:
            service = data.get("service")

            if action_type == "redeploy":
-                # Full service redeploy via the repo deploy script
-                cmd = [
-                    str(REPO_ROOT / "scripts" / "deploy" / "deploy-node.sh"),
-                    node,
-                    service
-                ]
-                logger.info(f"Running command: {' '.join(cmd)}")
-                result = subprocess.run(cmd, capture_output=True, text=True, cwd=str(REPO_ROOT))
-                if result.returncode == 0:
-                    success = True
-                else:
+                # Single-service redeploy, dispatched to the host-side deploy
+                # runner on the target node (jobs/deploy-runner/). Same pull
+                # architecture as container_restart — the executor never runs a
+                # deploy itself. It used to try: it ran deploy-node.sh inside
+                # this container, a script that ignores its arguments and needs
+                # a repo at ${HOME}/homelab-codex-ws (plus git and the docker
+                # CLI) that does not exist here, so every redeploy failed at the
+                # first line (recon D14/D15).
+                if not node or not service:
                    success = False
-                    error_msg = result.stderr or result.stdout
+                    error_msg = (
+                        f"redeploy requires both node and service "
+                        f"(node={node!r}, service={service!r})"
+                    )
+                else:
+                    self._dispatch_redeploy(action_id, node, service)
+                    # Stays in "running" until the runner reports an
+                    # action_result — same as container_restart below.
+                    return

            elif action_type == "container_restart":
                # No SSH from the executor (see CLAUDE.md / docs/backlog.md): the
@ -210,12 +235,47 @@ class Executor:
        except Exception as e:
            logger.error(f"Failed to dispatch {action_id} to {node}: {e}")

-    def _reconcile_running_actions(self):
-        """Resolve container_restart actions previously dispatched to a node-agent.
+    def _dispatch_redeploy(self, action_id, node, service):
+        """Write a redeploy action for the deploy runner on `node` to pick up.

-        Other action types (redeploy/disk_cleanup/alert_only) resolve
-        synchronously inside _execute_action and never linger in running/, so
-        they are not touched here.
+        Deliberately a different inbox from container_restart: node-agent
+        consumes actions/dispatch/<node>/ and rejects (and deletes) anything
+        that is not container_restart, so redeploys get their own
+        actions/deploy/<node>/ tree. The runner on the node pulls it over the
+        same rsync/ssh channel node-agent already uses, deploys locally with
+        scripts/deploy/deploy-service.sh, and reports back as an action_result
+        event — the executor still never initiates a connection to a node.
+
+        Does not resolve the action itself; _reconcile_running_actions() does.
+        """
+        inbox = DEPLOY_DISPATCH_DIR / node
+        inbox.mkdir(parents=True, exist_ok=True)
+        payload = {
+            "action_id": action_id,
+            "type": "redeploy",
+            "node": node,
+            "service": service,
+            "dispatched_at": time.time(),
+        }
+        try:
+            _atomic_write_json(inbox / f"{action_id}.json", payload)
+            logger.info(
+                f"Dispatched redeploy {action_id} (service={service}) "
+                f"to deploy runner on {node}"
+            )
+        except Exception as e:
+            logger.error(f"Failed to dispatch {action_id} to {node}: {e}")
+
+    def _reconcile_running_actions(self):
+        """Resolve actions previously dispatched to an on-node agent.
+
+        Covers both dispatched types: container_restart (executed by node-agent)
+        and redeploy (executed by the host-side deploy runner). Both report the
+        outcome the same way — an action_result event carrying the action_id —
+        so the resolution logic is identical; only the timeout differs.
+
+        Synchronous types (disk_cleanup/alert_only) resolve inside
+        _execute_action and never linger in running/, so they are not touched.
        """
        running_dir = ACTIONS_DIR / "running"
        if not running_dir.exists():
@ -229,24 +289,29 @@ class Executor:
                logger.error(f"Failed to read running action {action_file.name}: {e}")
                continue

-            if data.get("type") != "container_restart":
+            action_type = data.get("type")
+            if action_type not in DISPATCHED_ACTION_TYPES:
                continue

            action_id = data.get("action_id") or action_file.stem
            node = data.get("node")
            started_at = data.get("started_at") or 0
+            executor_name = "deploy runner" if action_type == "redeploy" else "node-agent"
+            timeout_secs = (
+                REDEPLOY_TIMEOUT_SECS if action_type == "redeploy" else ACTION_TIMEOUT_SECS
+            )

            result = self._find_action_result(node, action_id, started_at)
            if result is not None:
                success = bool(result.get("success"))
-                error_msg = "" if success else (result.get("error") or "node-agent reported failure")
+                error_msg = "" if success else (result.get("error") or f"{executor_name} reported failure")
                self._finalize_action(action_id, action_file, data, success, error_msg)
                continue

-            if started_at and (time.time() - started_at) > ACTION_TIMEOUT_SECS:
+            if started_at and (time.time() - started_at) > timeout_secs:
                error_msg = (
-                    f"Timed out after {ACTION_TIMEOUT_SECS}s waiting for node-agent "
-                    f"on '{node}' to report a result for container_restart "
+                    f"Timed out after {timeout_secs}s waiting for {executor_name} "
+                    f"on '{node}' to report a result for {action_type} "
                    f"(action_id={action_id})"
                )
                logger.error(f"Action {action_id} timed out: {error_msg}")
--- a/services/control-plane/src/supervisor.py
+++ b/services/control-plane/src/supervisor.py
@ -37,13 +37,14 @@ except Exception:
 # rather than a full redeploy: the container is present but not running, or
 # running with a failing health check — a restart plausibly heals both.
 # healthcheck_failed added 2026-07-29 (recon D14/D15): it used to route to
-# redeploy, which is broken as wired (executor calls deploy-node.sh with
-# arguments it ignores, at a path that does not exist in the container), so
-# 3376 healthcheck_failed events dead-ended. redeploy returns to the map only
-# once it actually works (etap 2). Everything not listed here
-# (service_unhealthy, deployment_failed, missing_service) still falls through
-# to redeploy — theoretical until etap 2 fixes the executor, kept as-is so the
-# drift stays visible in pending actions.
+# redeploy, which was broken as wired (the executor ran deploy-node.sh with
+# arguments it ignores, from a repo path that does not exist in its container),
+# so 3376 healthcheck_failed events dead-ended. Redeploy works as of 2026-08-03
+# (executor dispatch → jobs/deploy-runner/ on the node), but healthcheck_failed
+# deliberately STAYS here: restart-first is the cheaper, lower-risk remediation
+# for a container that is up with a failing probe. Everything not listed here
+# (service_unhealthy, deployment_failed, missing_service) falls through to
+# redeploy, which now actually executes.
 # mqtt_unreachable was removed 2026-07-28: the observer never creates incidents
 # with that trigger_type, so the branch was dead code (recon
 # docs/architecture/RECON-multiagent-2026-07-27.md, D15). stability-agent still
@ -446,10 +447,9 @@ class Supervisor:
        else:
            # Full redeploy: container is running but service is broken,
            # or the cause is unknown / not a simple restart candidate.
-            # NOTE: the redeploy action type is currently theoretical — the
-            # executor's redeploy path is broken until etap 2 (see
-            # CONTAINER_RESTART_TRIGGERS comment). Generated anyway so the
-            # drift is visible to the operator in pending actions.
+            # Executed on the target node by jobs/deploy-runner/ once an
+            # operator approves — `docker compose up -d --force-recreate` for
+            # this one service, from the checkout the node already has.
            action = {
                "action_id": action_id,
                "timestamp": time.time(),
--- a/services/control-plane/tests/test_executor_dispatch.py
+++ b/services/control-plane/tests/test_executor_dispatch.py
@ -31,6 +31,9 @@ def _setup_executor(tmp_path: Path, monkeypatch) -> Executor:
    monkeypatch.setattr(executor_module, "ACTIONS_DIR", actions)
    monkeypatch.setattr(executor_module, "EVENTS_DIR", events)
    monkeypatch.setattr(executor_module, "DISPATCH_DIR", actions / "dispatch")
+    # Must be patched too: _ensure_dirs() creates it, and the module default
+    # points at the real /opt/homelab.
+    monkeypatch.setattr(executor_module, "DEPLOY_DISPATCH_DIR", actions / "deploy")
    monkeypatch.setattr(executor_module, "REPO_ROOT", repo)
    return Executor()

@ -212,17 +215,19 @@ def test_reconcile_does_not_time_out_before_deadline(tmp_path, monkeypatch):
 # Other action types are untouched by the reconcile loop
 # ---------------------------------------------------------------------------

-def test_reconcile_ignores_non_container_restart_running_actions(tmp_path, monkeypatch):
+def test_reconcile_ignores_synchronous_running_actions(tmp_path, monkeypatch):
+    """Only dispatched types (container_restart, redeploy) are reconciled here.
+    disk_cleanup/alert_only resolve inside _execute_action and never linger."""
    ex = _setup_executor(tmp_path, monkeypatch)
-    running_path = tmp_path / "actions" / "running" / "redeploy-x.json"
+    running_path = tmp_path / "actions" / "running" / "disk-cleanup-x.json"
    running_path.write_text(json.dumps({
-        "action_id": "redeploy-x", "type": "redeploy", "node": "piha",
+        "action_id": "disk-cleanup-x", "type": "disk_cleanup", "node": "piha",
        "started_at": time.time() - 999_999,
    }))

    ex._reconcile_running_actions()

-    assert running_path.exists()  # untouched: redeploy resolves synchronously elsewhere
+    assert running_path.exists()  # untouched


 def test_alert_only_action_resolves_synchronously_not_via_dispatch(tmp_path, monkeypatch):
--- a/services/control-plane/tests/test_executor_redeploy.py
+++ b/services/control-plane/tests/test_executor_redeploy.py
@ -0,0 +1,189 @@
+"""Tests for the Executor redeploy path (recon D14/D15 — etap 2).
+
+Before this, `redeploy` ran scripts/deploy/deploy-node.sh inside the executor
+container: a script that ignores its arguments, expects a repo at
+${HOME}/homelab-codex-ws (absent there, along with git and the docker CLI), and
+would have deployed the EXECUTOR's own host's full service set rather than the
+action's target. Every redeploy failed, which is why healthcheck_failed had to
+be rerouted to container_restart and the action queue never drained.
+
+Now redeploy follows the same pull architecture as container_restart: the
+executor writes a dispatch file, the host-side deploy runner on the target node
+(jobs/deploy-runner/) executes it and reports back an action_result event.
+"""
+from __future__ import annotations
+
+import json
+import sys
+import time
+from pathlib import Path
+
+import pytest
+
+sys.path.insert(0, str(Path(__file__).parent.parent / "src"))
+import executor as executor_module
+from executor import Executor
+
+
+def _setup_executor(tmp_path: Path, monkeypatch) -> Executor:
+    actions = tmp_path / "actions"
+    events = tmp_path / "events"
+    for d in (actions, events):
+        d.mkdir(parents=True, exist_ok=True)
+    monkeypatch.setattr(executor_module, "ACTIONS_DIR", actions)
+    monkeypatch.setattr(executor_module, "EVENTS_DIR", events)
+    monkeypatch.setattr(executor_module, "DISPATCH_DIR", actions / "dispatch")
+    monkeypatch.setattr(executor_module, "DEPLOY_DISPATCH_DIR", actions / "deploy")
+    return Executor()
+
+
+def _write_approved(tmp_path, action_id, node="piha", service="vikunja"):
+    action = {
+        "action_id": action_id,
+        "type": "redeploy",
+        "node": node,
+        "service": service,
+        "status": "approved",
+        "timestamp": time.time(),
+    }
+    path = tmp_path / "actions" / "approved" / f"{action_id}.json"
+    path.write_text(json.dumps(action))
+    return path
+
+
+def _write_action_result_event(tmp_path, node, action_id, success, error="", ts=None):
+    ts = int(ts if ts is not None else time.time())
+    node_dir = tmp_path / "events" / node
+    node_dir.mkdir(parents=True, exist_ok=True)
+    event_id = f"evt-{node}-{ts}-action_result-{action_id}"
+    (node_dir / f"{event_id}.json").write_text(json.dumps({
+        "id": event_id,
+        "timestamp": ts,
+        "type": "action_result",
+        "node": node,
+        "payload": {"action_id": action_id, "success": success, "error": error,
+                    "node": node, "source": "deploy-runner"},
+    }))
+
+
+def _exists(tmp_path, state, action_id):
+    return (tmp_path / "actions" / state / f"{action_id}.json").exists()
+
+
+def _read(tmp_path, state, action_id):
+    return json.loads((tmp_path / "actions" / state / f"{action_id}.json").read_text())
+
+
+# ---------------------------------------------------------------------------
+# Dispatch
+# ---------------------------------------------------------------------------
+
+def test_redeploy_is_dispatched_not_executed_locally(tmp_path, monkeypatch):
+    ex = _setup_executor(tmp_path, monkeypatch)
+
+    def fail(*args, **kwargs):
+        raise AssertionError("executor must never run a deploy itself")
+
+    monkeypatch.setattr(executor_module.subprocess, "run", fail)
+
+    ex._execute_action(_write_approved(tmp_path, "redeploy-piha-vikunja"))
+
+    dispatch_file = tmp_path / "actions" / "deploy" / "piha" / "redeploy-piha-vikunja.json"
+    assert dispatch_file.exists()
+    payload = json.loads(dispatch_file.read_text())
+    assert payload["type"] == "redeploy"
+    assert payload["node"] == "piha"
+    assert payload["service"] == "vikunja"
+    assert "dispatched_at" in payload
+
+
+def test_redeploy_inbox_is_separate_from_node_agent_dispatch(tmp_path, monkeypatch):
+    """node-agent deletes and failure-reports anything in its own inbox that is
+    not container_restart, so a redeploy must never be written there."""
+    ex = _setup_executor(tmp_path, monkeypatch)
+    ex._execute_action(_write_approved(tmp_path, "redeploy-piha-vikunja"))
+
+    node_agent_inbox = tmp_path / "actions" / "dispatch"
+    assert not list(node_agent_inbox.glob("**/*.json"))
+
+
+def test_redeploy_stays_running_until_reported(tmp_path, monkeypatch):
+    ex = _setup_executor(tmp_path, monkeypatch)
+    ex._execute_action(_write_approved(tmp_path, "rd-1"))
+
+    assert _exists(tmp_path, "running", "rd-1")
+    assert not _exists(tmp_path, "completed", "rd-1")
+    assert not _exists(tmp_path, "failed", "rd-1")
+
+
+@pytest.mark.parametrize("node,service", [("", "vikunja"), ("piha", ""), ("", "")])
+def test_redeploy_without_node_or_service_fails_immediately(tmp_path, monkeypatch, node, service):
+    ex = _setup_executor(tmp_path, monkeypatch)
+    ex._execute_action(_write_approved(tmp_path, "rd-bad", node=node, service=service))
+
+    assert _exists(tmp_path, "failed", "rd-bad")
+    assert "requires both node and service" in _read(tmp_path, "failed", "rd-bad")["error"]
+    assert not list((tmp_path / "actions" / "deploy").glob("**/*.json"))
+
+
+# ---------------------------------------------------------------------------
+# Reconcile
+# ---------------------------------------------------------------------------
+
+def test_redeploy_completes_on_success_result(tmp_path, monkeypatch):
+    ex = _setup_executor(tmp_path, monkeypatch)
+    ex._execute_action(_write_approved(tmp_path, "rd-2"))
+
+    _write_action_result_event(tmp_path, "piha", "rd-2", success=True)
+    ex._reconcile_running_actions()
+
+    assert _exists(tmp_path, "completed", "rd-2")
+    assert not _exists(tmp_path, "running", "rd-2")
+
+
+def test_redeploy_fails_on_failure_result(tmp_path, monkeypatch):
+    ex = _setup_executor(tmp_path, monkeypatch)
+    ex._execute_action(_write_approved(tmp_path, "rd-3"))
+
+    _write_action_result_event(tmp_path, "piha", "rd-3", success=False,
+                               error="deploy-service.sh exited 1")
+    ex._reconcile_running_actions()
+
+    assert _read(tmp_path, "failed", "rd-3")["error"] == "deploy-service.sh exited 1"
+
+
+def test_redeploy_uses_its_own_longer_timeout(tmp_path, monkeypatch):
+    """A redeploy must not be timed out on the container_restart budget: the
+    runner polls once a minute and may pull images before reporting."""
+    monkeypatch.setattr(executor_module, "ACTION_TIMEOUT_SECS", 300)
+    monkeypatch.setattr(executor_module, "REDEPLOY_TIMEOUT_SECS", 900)
+    ex = _setup_executor(tmp_path, monkeypatch)
+    ex._execute_action(_write_approved(tmp_path, "rd-4"))
+
+    running_path = tmp_path / "actions" / "running" / "rd-4.json"
+    data = json.loads(running_path.read_text())
+    data["started_at"] = time.time() - 600  # past 300s, inside 900s
+    running_path.write_text(json.dumps(data))
+
+    ex._reconcile_running_actions()
+    assert _exists(tmp_path, "running", "rd-4")
+
+    data["started_at"] = time.time() - 1000  # past 900s
+    running_path.write_text(json.dumps(data))
+
+    ex._reconcile_running_actions()
+    assert _exists(tmp_path, "failed", "rd-4")
+    error = _read(tmp_path, "failed", "rd-4")["error"]
+    assert "Timed out after 900s" in error
+    assert "deploy runner" in error
+
+
+def test_stale_result_from_previous_run_does_not_resolve(tmp_path, monkeypatch):
+    ex = _setup_executor(tmp_path, monkeypatch)
+    _write_action_result_event(tmp_path, "piha", "rd-5", success=True,
+                               ts=int(time.time()) - 86_400)
+
+    ex._execute_action(_write_approved(tmp_path, "rd-5"))
+    ex._reconcile_running_actions()
+
+    assert _exists(tmp_path, "running", "rd-5")