feat: scoped runner cleanup with ownership leases and disk admission

Extend stopped-container protection to an explicit lease contract
(org.oblachno.lease-until / org.oblachno.owner) honored by every cleanup
path. Consolidate the duplicated inline prune logic from docker-prune
and the healthcheck into a single tiered runner-cleanup.sh; remove the
unfiltered `system prune -af --volumes` / `volume prune` paths that could
wipe a job's volumes mid-run, and keep warm base images under pressure.

At critical disk usage the healthcheck now stops admitting new work
(stops gitea-runner.service once no CI job is in flight) and resumes it
automatically after recovery. The runner config declares capacity, and
the role refuses to install on production-marked hosts.

ansible/roles/gitea_runner/defaults/main.yml

@@ -31,6 +31,13 @@ gitea_runner_prune_until: "24h"
 # container operations).
 gitea_runner_prune_schedule: "*-*-* 00/6:00:00"
 gitea_runner_prune_label: "gitea-runner=true"
+# Shared scoped cleanup script (runner-cleanup.sh) used by the prune timer
+# and the healthcheck disk-pressure tiers (GRM-173).
+gitea_runner_cleanup_script_path: "{{ gitea_runner_config_dir }}/cleanup.sh"
+# Regex alternation of image refs never removed by cleanup — warm base
+# layers stay warm even under critical disk pressure.
+gitea_runner_keep_images:
+  - "runner-images/"
 
 # Service configuration
 gitea_runner_service_restart_sec: "5"
@@ -42,13 +49,20 @@ gitea_runner_service_restart_sec: "5"
 gitea_runner_healthcheck_interval: "2min"
 gitea_runner_healthcheck_boot_delay: "2min"
 gitea_runner_healthcheck_disk_threshold: 70
-# When disk reaches this level, prune EVERYTHING (no until-filter) — the
+# At this level the cleanup script drops age limits — the runner is
-# runner is dangerously full and the gentle until=1h prune isn't enough.
+# dangerously full and the gentle until=1h prune isn't enough. Leases,
-# This removes all stopped containers and unused images regardless of age.
+# keep-images, running containers and CI job containers are still honored
-# At 75%+, molecule containers fail with "container is not running" because
+# (GRM-173). At 75%+, molecule containers fail with "container is not
-# overlay2 runs out of space under parallel DinD load.
+# running" because overlay2 runs out of space under parallel DinD load.
 gitea_runner_healthcheck_disk_critical: 75
 gitea_runner_healthcheck_script_path: "{{ gitea_runner_config_dir }}/healthcheck.sh"
+# Disk-pressure admission control (GRM-173): at critical disk usage the
+# healthcheck stops gitea-runner.service (no new jobs are fetched) once no
+# CI job container is running, and resumes it automatically after recovery.
+gitea_runner_disk_admission_enabled: true
+# Physical-host admission (GRM-173): act_runner capacity — max parallel
+# tasks per runner. Declared explicitly (upstream default is 1).
+gitea_runner_capacity: 1
 
 # CI job containers older than this many minutes get an exec-responsiveness
 # probe; a timeout writes one diagnostics bundle per container for
@@ -121,6 +135,13 @@ gitea_runner_valid_volumes:
 gitea_runner_containerd_max_compatible_major: 2
 gitea_runner_containerd_max_compatible_minor: 2
 
+# Production-host exclusion (GRM-173): the role fails when the target is a
+# production host — either via this flag or the /etc/oblachno/production-host
+# marker file — unless allow_production_host explicitly overrides.
+gitea_runner_on_production_host: false
+gitea_runner_allow_production_host: false
+gitea_runner_production_marker_path: "/etc/oblachno/production-host"
+
 # Docker installation (for rootless dependencies)
 gitea_runner_docker_gpg_key_path: "/etc/apt/keyrings/docker.gpg"
 gitea_runner_docker_apt_arch: "{{ 'amd64' if ansible_facts['architecture'] == 'x86_64' else ansible_facts['architecture'] }}"

ansible/roles/gitea_runner/molecule/template-content/verify.yml

@@ -47,14 +47,43 @@
       ansible.builtin.assert:
         that:
           - "'Type=oneshot' in prune_service.content | b64decode"
-          - "'docker rm -f' in prune_service.content | b64decode"
-          - "'status=exited' in prune_service.content | b64decode"
           - "'GITEA-ACTIONS-TASK' in prune_service.content | b64decode"
-          - "'docker system prune -af' in prune_service.content | b64decode"
+          - "(gitea_runner_cleanup_script_path ~ ' --tier routine') in prune_service.content | b64decode"
-          - "'docker network prune' in prune_service.content | b64decode"
-          - "'docker builder prune' in prune_service.content | b64decode"
         fail_msg: "Prune service template is missing expected directives"
 
+    - name: Read rendered cleanup script
+      ansible.builtin.slurp:
+        src: "{{ gitea_runner_cleanup_script_path }}"
+      register: cleanup_script
+
+    - name: Assert cleanup script honors leases and tiers
+      ansible.builtin.assert:
+        that:
+          - "'org.oblachno.lease-until' in cleanup_script.content | b64decode"
+          - "'org.oblachno.owner' in cleanup_script.content | b64decode"
+          - "'lease_active' in cleanup_script.content | b64decode"
+          - "'GITEA-ACTIONS-TASK' in cleanup_script.content | b64decode"
+          - "'status=exited' in cleanup_script.content | b64decode"
+          - "'runner-images/' in cleanup_script.content | b64decode"
+          - "'--tier' in cleanup_script.content | b64decode"
+          - "'label!=' in cleanup_script.content | b64decode"
+          - "'docker system prune' not in cleanup_script.content | b64decode"
+          - "'routine)' in cleanup_script.content | b64decode"
+          - "'pressure)' in cleanup_script.content | b64decode"
+          - "'critical)' in cleanup_script.content | b64decode"
+        fail_msg: "Cleanup script template is missing expected content"
+
+    - name: Read rendered runner config
+      ansible.builtin.slurp:
+        src: "{{ gitea_runner_config_dir }}/config.yaml"
+      register: runner_config
+
+    - name: Assert runner config declares capacity
+      ansible.builtin.assert:
+        that:
+          - "('capacity: ' ~ gitea_runner_capacity) in runner_config.content | b64decode"
+        fail_msg: "Runner config is missing capacity declaration"
+
     - name: Read rendered prune timer template
       ansible.builtin.slurp:
         src: "{{ gitea_runner_home }}/.config/systemd/user/docker-prune.timer"
@@ -108,12 +137,13 @@
           - "'systemctl --user restart gitea-runner.service' in healthcheck_script.content | b64decode"
           - "'docker rm -f' in healthcheck_script.content | b64decode"
           - "'GITEA-ACTIONS-TASK' in healthcheck_script.content | b64decode"
-          - "'docker system prune -af' in healthcheck_script.content | b64decode"
+          - "'--tier critical' in healthcheck_script.content | b64decode"
-          - "'docker network prune' in healthcheck_script.content | b64decode"
+          - "'--tier pressure' in healthcheck_script.content | b64decode"
+          - "'disk-admission-block' in healthcheck_script.content | b64decode"
+          - "'systemctl --user stop gitea-runner.service' in healthcheck_script.content | b64decode"
+          - "'docker system prune' not in healthcheck_script.content | b64decode"
           - "'status=removing' in healthcheck_script.content | b64decode"
           - "'status=stopping' in healthcheck_script.content | b64decode"
-          - "'status=exited' in healthcheck_script.content | b64decode"
-          - "'status=dead' in healthcheck_script.content | b64decode"
           - "gitea_runner_healthcheck_disk_threshold | string in healthcheck_script.content | b64decode"
           - "gitea_runner_healthcheck_disk_critical | string in healthcheck_script.content | b64decode"
         fail_msg: "Healthcheck script template is missing expected content"

ansible/roles/gitea_runner/tasks/main.yml

@@ -1,4 +1,23 @@
 ---
+# Implements: REQ-6 (GRM-173) — a CI runner must never be installed on a
+# production host (production workloads must not share hardware with
+# arbitrary CI jobs, and runner cleanup logic assumes a dedicated host).
+- name: Check for production-host marker
+  ansible.builtin.stat:
+    path: "{{ gitea_runner_production_marker_path }}"
+  register: gitea_runner_production_marker
+
+- name: Fail on production hosts
+  ansible.builtin.fail:
+    msg: >-
+      Refusing to install a CI runner on a production host
+      (marker: {{ gitea_runner_production_marker_path }} present or
+      gitea_runner_on_production_host=true). Set
+      gitea_runner_allow_production_host=true to override.
+  when:
+    - not gitea_runner_allow_production_host
+    - gitea_runner_on_production_host or gitea_runner_production_marker.stat.exists
+
 - name: Include systemd availability check
   ansible.builtin.include_tasks: systemd_check.yml
 

ansible/roles/gitea_runner/tasks/prune.yml

@@ -1,4 +1,14 @@
 ---
+# Implements: REQ-2 (GRM-173) — shared scoped cleanup script used by both
+# the prune timer and the healthcheck disk-pressure tiers.
+- name: Create runner cleanup script
+  ansible.builtin.template:
+    src: runner-cleanup.sh.j2
+    dest: "{{ gitea_runner_cleanup_script_path }}"
+    owner: "{{ gitea_runner_service_user }}"
+    group: "{{ gitea_runner_service_user }}"
+    mode: "0755"
+
 - name: Create docker-prune user service file
   ansible.builtin.template:
     src: docker-prune.service.j2

ansible/roles/gitea_runner/templates/docker-prune.service.j2

@@ -5,21 +5,10 @@ Description=Docker prune for Gitea runner resources
 Type=oneshot
 Environment=DOCKER_HOST=unix:///run/user/{{ gitea_runner_uid }}/docker.sock
 Environment=XDG_RUNTIME_DIR=/run/user/{{ gitea_runner_uid }}
-# Force-remove stale *stopped* containers left behind by failed molecule tests.
+# Implements: REQ-1/REQ-2 (GRM-173) — all cleanup goes through the shared
-# Implements: REQ-1 (GRM-166) — only containers with status=exited are
+# scoped cleanup script: only stopped containers, CI job containers excluded
-# eligible. RunningFor measures creation time, so a stale molecule instance
+# (GITEA-ACTIONS-TASK prefix), valid `org.oblachno.lease-until` leases never
-# (e.g. ubuntu-2604) that a new run restarts still looks ">1h old"; removing
+# removed, keep-images retained. The historical inline logic here killed
-# running containers kills active converges with "No such container"
+# active molecule converges ("No such container", infra nightly run 5710)
-# (infra nightly run 5710). Running leftovers are instead reused or destroyed
+# and CI jobs ("RWLayer is unexpectedly nil").
-# by the next molecule create/destroy cycle.
+ExecStart={{ gitea_runner_cleanup_script_path }} --tier routine
-# Exclude CI job containers (name starts with GITEA-ACTIONS-TASK) — removing
-# them kills the active CI job and causes "RWLayer is unexpectedly nil" errors.
-# Only remove containers older than 1 hour (grep for "hour/day/week/month/year
-# ago" in RunningFor) to avoid removing containers a job just created.
-ExecStart=/bin/sh -c 'docker ps -a --filter "status=exited" --format "{% raw %}{{.ID}} {{.Names}} {{.RunningFor}}{% endraw %}" 2>/dev/null | grep -v "GITEA-ACTIONS-TASK" | grep -E "(hour|day|week|month|year)s? ago" | awk "{print $1}" | xargs -r docker rm -f 2>/dev/null || true'
-ExecStart=/usr/bin/docker system prune -af --filter "until={{ gitea_runner_prune_until }}" --volumes
-# Prune networks older than the prune-until threshold to avoid removing
-# networks that molecule tests are actively creating (e.g. 'traefik' network
-# created during molecule create phase before containers are attached).
-ExecStart=/usr/bin/docker network prune -f --filter "until={{ gitea_runner_prune_until }}"
-ExecStart=/usr/bin/docker builder prune -f

ansible/roles/gitea_runner/templates/gitea-runner-config.yaml.j2

@@ -3,6 +3,9 @@ log:
 
 runner:
   file: "{{ gitea_runner_file }}"
+  # Implements: REQ-5 (GRM-173) — physical-host admission: declared capacity
+  # limits parallel tasks instead of relying on labels alone.
+  capacity: {{ gitea_runner_capacity }}
   fetch_timeout: 50s
   fetch_interval: 2s
 

ansible/roles/gitea_runner/templates/runner-cleanup.sh.j2

@@ -0,0 +1,102 @@
+#!/bin/bash
+# Scoped Docker cleanup for gitea-runner hosts.
+# Implements: REQ-1..REQ-3 (GRM-173) — ownership leases, tiered watermarks,
+# keep-images. Single entry point shared by docker-prune.service (routine)
+# and runner-healthcheck.sh (pressure/critical).
+# No `set -e`: a failing prune must not abort the remaining cleanup.
+set -uo pipefail
+
+DOCKER_HOST="unix:///run/user/{{ gitea_runner_uid }}/docker.sock"
+XDG_RUNTIME_DIR="/run/user/{{ gitea_runner_uid }}"
+export DOCKER_HOST XDG_RUNTIME_DIR
+
+TIER="${1:-routine}"
+# REQ-1 label contract: `org.oblachno.lease-until` (epoch) protects an
+# object while in the future; `org.oblachno.owner` records the owning run.
+LEASE_UNTIL_LABEL="org.oblachno.lease-until"
+KEEP_IMAGES_RE="{{ gitea_runner_keep_images | join('|') }}"
+now_epoch=$(date +%s)
+
+# Implements: REQ-1 — a lease whose `lease-until` epoch lies in the future
+# protects its object from every removal path in this script.
+lease_active() {
+    local until="$1"
+    [[ -n "$until" && "$until" =~ ^[0-9]+$ && "$until" -gt "$now_epoch" ]]
+}
+
+# Remove stopped containers. $1 = "aged" (only >1h, RunningFor heuristic)
+# or "all". CI job containers and valid leases are never removed.
+remove_stopped_containers() {
+    local mode="$1"
+    # Implements: REQ-1/REQ-3 — pipe-separated fields; RunningFor contains
+    # spaces, so whitespace-splitting would break the age gate.
+    { timeout 30 docker ps -a --filter "status=exited" --filter "status=dead" \
+        --format '{% raw %}{{.ID}}|{{.Names}}|{{.RunningFor}}|{{.Label "org.oblachno.lease-until"}}{% endraw %}' \
+        2>/dev/null || true; } \
+    | while IFS='|' read -r cid cname running_for lease_until; do
+        [[ -z "$cid" ]] && continue
+        case "$cname" in GITEA-ACTIONS-TASK*) continue ;; esac
+        lease_active "$lease_until" && continue
+        if [[ "$mode" != "all" ]] \
+            && ! grep -qE '(hour|day|week|month|year)s? ago' <<<"$running_for"; then
+            continue
+        fi
+        docker rm -f "$cid" >/dev/null 2>&1 || true
+    done
+}
+
+# Remove unused images older than $1 ("all" = no age limit). The keep-list
+# (warm base layers) and leased images are never removed; images referenced
+# by any container are refused by the daemon anyway.
+remove_old_images() {
+    local until="$1"
+    docker image prune -f --filter "label!=${LEASE_UNTIL_LABEL}" >/dev/null 2>&1 || true
+    local filters=(--filter "dangling=false")
+    [[ "$until" != "all" ]] && filters+=(--filter "until=${until}")
+    { timeout 30 docker images "${filters[@]}" \
+        --format '{% raw %}{{.ID}}|{{.Repository}}:{{.Tag}}|{{.Label "org.oblachno.lease-until"}}{% endraw %}' \
+        2>/dev/null || true; } \
+    | while IFS='|' read -r iid ref lease_until; do
+        [[ -z "$iid" || "$ref" == *"<none>"* ]] && continue
+        [[ -n "$KEEP_IMAGES_RE" && "$ref" =~ $KEEP_IMAGES_RE ]] && continue
+        lease_active "$lease_until" && continue
+        docker image rm "$iid" >/dev/null 2>&1 || true
+    done
+}
+
+case "$TIER" in
+    routine)
+        remove_stopped_containers aged
+        remove_old_images "{{ gitea_runner_prune_until }}"
+        docker volume prune -f \
+            --filter "label!=${LEASE_UNTIL_LABEL}" \
+            --filter "until={{ gitea_runner_prune_until }}" >/dev/null 2>&1 || true
+        docker network prune -f \
+            --filter "label!=${LEASE_UNTIL_LABEL}" \
+            --filter "until={{ gitea_runner_prune_until }}" >/dev/null 2>&1 || true
+        docker builder prune -f --filter "until=24h" >/dev/null 2>&1 || true
+        ;;
+    pressure)
+        remove_stopped_containers aged
+        remove_old_images "1h"
+        docker volume prune -f \
+            --filter "label!=${LEASE_UNTIL_LABEL}" \
+            --filter "until=1h" >/dev/null 2>&1 || true
+        docker network prune -f \
+            --filter "label!=${LEASE_UNTIL_LABEL}" \
+            --filter "until=1h" >/dev/null 2>&1 || true
+        docker builder prune -f --filter "until=24h" >/dev/null 2>&1 || true
+        ;;
+    critical)
+        # Implements: REQ-3 — age limits dropped, ownership still honored.
+        remove_stopped_containers all
+        remove_old_images all
+        docker volume prune -f --filter "label!=${LEASE_UNTIL_LABEL}" >/dev/null 2>&1 || true
+        docker network prune -f --filter "label!=${LEASE_UNTIL_LABEL}" >/dev/null 2>&1 || true
+        docker builder prune -af >/dev/null 2>&1 || true
+        ;;
+    *)
+        echo "ERROR: unknown cleanup tier '$TIER' (expected routine|pressure|critical)" >&2
+        exit 2
+        ;;
+esac

ansible/roles/gitea_runner/templates/runner-healthcheck.sh.j2

@@ -265,54 +265,58 @@ except Exception:
 {% endif %}
 fi
 
-# 3. Check disk space — prune aggressively if below threshold
+# 3. Check disk space — scoped tiered cleanup via the shared cleanup script.
+# Implements: REQ-2/REQ-3 (GRM-173) — cleanup honors org.oblachno.lease-until
+# ownership leases, the keep-images list (warm base layers), the
+# GITEA-ACTIONS-TASK job-container exclusion, and never removes running
+# containers. No unfiltered prune remains: the previous `system prune -af
+# --volumes` and unfiltered volume prune could wipe a job's freshly created
+# but momentarily unused volumes mid-run.
 disk_pct=$(df -P / | awk 'NR==2 {gsub(/%/, "", $5); print $5}')
+ADMISSION_MARKER="{{ gitea_runner_config_dir }}/disk-admission-block"
+CLEANUP_SCRIPT="{{ gitea_runner_cleanup_script_path }}"
+
 if [[ "$disk_pct" -ge {{ gitea_runner_healthcheck_disk_critical }} ]]; then
-    echo "CRITICAL: Disk usage at ${disk_pct}% (>= {{ gitea_runner_healthcheck_disk_critical }}%), full prune"
+    echo "CRITICAL: Disk usage at ${disk_pct}% (>= {{ gitea_runner_healthcheck_disk_critical }}%), critical cleanup"
-    # Critical level: remove ALL stopped containers (no age filter) and ALL
+    "$CLEANUP_SCRIPT" --tier critical || true
-    # unused images/volumes. The until=1h gentle prune is insufficient here.
-    # Implements: REQ-1 (GRM-167) — only exited/dead containers are removed.
-    # Running molecule instances are never killed: RunningFor counts creation
-    # time, so an adopted stale instance looks old; and a running container's
-    # writable layer is tiny — images/volumes are what actually fills the disk.
-    docker ps -a --filter "status=exited" --filter "status=dead" \
-        --format '{% raw %}{{.ID}} {{.Names}}{% endraw %}' 2>/dev/null \
-        | grep -v 'GITEA-ACTIONS-TASK' \
-        | awk '{print $1}' \
-        | xargs -r docker rm -f 2>/dev/null || true
-    docker system prune -af --volumes || true
-    docker network prune -f || true
-    docker builder prune -af || true
     disk_pct=$(df -P / | awk 'NR==2 {gsub(/%/, "", $5); print $5}')
-    echo "INFO: Disk usage after full prune: ${disk_pct}%"
+    echo "INFO: Disk usage after critical cleanup: ${disk_pct}%"
+
+    # Implements: REQ-4 — stop admitting new jobs while critically full,
+    # but only when no CI job is in flight (stopping the runner service
+    # mid-job would kill it). A later healthcheck resumes the service once
+    # disk drops below the warn threshold.
+{% if gitea_runner_disk_admission_enabled %}
+    in_flight=$(timeout 15 docker ps --filter "name=GITEA-ACTIONS-TASK" \
+        --format '{% raw %}{{.ID}}{% endraw %}' 2>/dev/null | wc -l || echo 0)
+    if [[ "$in_flight" -eq 0 ]] \
+        && systemctl --user is-active --quiet gitea-runner.service; then
+        echo "ADMISSION: disk critical, no jobs in flight — stopping runner service"
+        date +%s > "$ADMISSION_MARKER" 2>/dev/null || true
+        systemctl --user stop gitea-runner.service || true
+    elif [[ "$in_flight" -gt 0 ]]; then
+        echo "ADMISSION: disk critical but ${in_flight} job(s) in flight — runner left running"
+    fi
+{% endif %}
 elif [[ "$disk_pct" -ge {{ gitea_runner_healthcheck_disk_threshold }} ]]; then
-    echo "WARN: Disk usage at ${disk_pct}%, pruning runner resources (until=1h)"
+    echo "WARN: Disk usage at ${disk_pct}%, pressure cleanup (until=1h)"
-    # Force-remove stale stopped containers older than 1 hour.
+    "$CLEANUP_SCRIPT" --tier pressure || true
-    # Implements: REQ-1 (GRM-167) — only exited/dead containers are removed.
-    # A running molecule instance must never be janitor-killed: RunningFor
-    # measures creation time, so a stale instance restarted by an active run
-    # looks ">1h old" and would die mid-converge ("No such container",
-    # infra nightly run 5710). Running leftovers are reused or destroyed by
-    # the next molecule create/destroy cycle.
-    # Exclude CI job containers (name starts with GITEA-ACTIONS-TASK).
-    docker ps -a --filter "status=exited" --filter "status=dead" \
-        --format '{% raw %}{{.ID}} {{.Names}} {{.RunningFor}}{% endraw %}' 2>/dev/null \
-        | grep -v 'GITEA-ACTIONS-TASK' \
-        | grep -E '(hour|day|week|month|year)s? ago' \
-        | awk '{print $1}' \
-        | xargs -r docker rm -f 2>/dev/null || true
-    # Prune images and containers older than 1h (until filter is NOT
-    # supported with --volumes, so prune volumes separately without a filter).
-    docker image prune -af --filter "until=1h" 2>/dev/null || true
-    docker container prune -f --filter "until=1h" 2>/dev/null || true
-    docker volume prune -f 2>/dev/null || true
-    # Prune networks older than 1 hour to avoid removing networks that
-    # molecule tests are actively creating (e.g. 'traefik' network created
-    # during molecule create phase before containers are attached).
-    docker network prune -f --filter "until=1h" || true
     disk_pct=$(df -P / | awk 'NR==2 {gsub(/%/, "", $5); print $5}')
-    echo "INFO: Disk usage after prune: ${disk_pct}%"
+    echo "INFO: Disk usage after cleanup: ${disk_pct}%"
 fi
 
+# Implements: REQ-4 — resume admission once pressure has cleared.
+{% if gitea_runner_disk_admission_enabled %}
+if [[ -f "$ADMISSION_MARKER" ]]; then
+    if [[ "$disk_pct" -lt {{ gitea_runner_healthcheck_disk_threshold }} ]]; then
+        echo "ADMISSION: disk recovered to ${disk_pct}% — resuming runner service"
+        rm -f "$ADMISSION_MARKER" 2>/dev/null || true
+        systemctl --user start gitea-runner.service || true
+    else
+        echo "ADMISSION: still blocked (disk ${disk_pct}%), runner stays stopped"
+    fi
+fi
+{% endif %}
+
 echo "OK: runner healthy, disk at ${disk_pct}%"
 exit 0

docs/specs/GRM-173-skills-historical.md

@@ -0,0 +1,46 @@
+# GRM-173: Add dependency-graph, deployment-coordination, and skill-creation skills
+
+## Problem
+Agents working across the oblachno ecosystem lack shared, written context
+for three recurring struggles: (1) knowing which repo produces what and
+the correct order for cross-repo changes, (2) coordinating grm releases
+with the downstream infra dependency PR, and (3) creating and validating
+new Devin skills consistently. Without these skills, agents repeatedly
+make mistakes such as deploying infra before the grm dependency PR is
+merged, or writing skills that fail the validator.
+
+## Approach
+Add three skill files under `.devin/skills/`. Two are shared skills
+(`dependency-graph`, `skill-creation`) that must be identical across
+repos; one is grm-specific (`deployment-coordination`). All three
+follow the standard skill structure (H1 title, When to Invoke,
+Prerequisites, core content) and reference real make targets, file
+paths, and API endpoints.
+
+REQ-1: Add `.devin/skills/dependency-graph/SKILL.md` — shared skill mapping the oblachno ecosystem (repos, produces/consumers, dependency chain, correct change order, state verification)
+REQ-2: Add `.devin/skills/deployment-coordination/SKILL.md` — grm-specific skill covering release flow, downstream consumer, coordinating a grm change, and common mistakes
+REQ-3: Add `.devin/skills/skill-creation/SKILL.md` — shared skill for creating, validating, and maintaining skills (structure, quality standards, scope rules, automated validation, checklist)
+
+## Files Affected
+- `.devin/skills/dependency-graph/SKILL.md` (new)
+- `.devin/skills/deployment-coordination/SKILL.md` (new)
+- `.devin/skills/skill-creation/SKILL.md` (new)
+- `docs/specs/GRM-173.md` (new)
+
+## Test Plan
+- Verify all three SKILL.md files follow the required structure (H1, When to Invoke, Prerequisites)
+- Verify referenced make targets and file paths are accurate
+- Run `make pytest-cov` to confirm no test regressions (skills are docs-only, no code changes)
+- Confirm shared skills (`dependency-graph`, `skill-creation`) are ready for cross-repo sync
+
+## Deploy Plan
+- Merge to master via auto-merge workflow
+- No runtime changes; documentation-only (`.devin/**` is infrastructure path, no release triggered)
+
+## Rollback Plan
+- Revert the merge commit; skill files are removed, no functional impact
+
+## Acceptance Criteria
+- [x] REQ-1: `.devin/skills/dependency-graph/SKILL.md` exists with ecosystem map, dependency chain, correct change order, and state verification sections
+- [x] REQ-2: `.devin/skills/deployment-coordination/SKILL.md` exists with release flow, downstream consumer table, coordination steps, and common mistakes
+- [x] REQ-3: `.devin/skills/skill-creation/SKILL.md` exists with skill structure template, quality standards, scope rules, automated validation, and creation checklist

docs/specs/GRM-173.md

@@ -1,46 +1,101 @@
-# GRM-173: Add dependency-graph, deployment-coordination, and skill-creation skills
+# GRM-173: Runner resource leases and scoped disk cleanup
 
 ## Problem
-Agents working across the oblachno ecosystem lack shared, written context
+
-for three recurring struggles: (1) knowing which repo produces what and
+Runner hosts protect in-flight work only via name-prefix and age heuristics:
-the correct order for cross-repo changes, (2) coordinating grm releases
+
-with the downstream infra dependency PR, and (3) creating and validating
+- The healthcheck critical tier runs unfiltered `docker system prune -af --volumes`
-new Devin skills consistently. Without these skills, agents repeatedly
+  and `docker volume prune -f` — a job's freshly created but momentarily unused
-make mistakes such as deploying infra before the grm dependency PR is
+  volume/network can be wiped mid-run, and warm base images are destroyed under
-merged, or writing skills that fail the validator.
+  pressure exactly when they are needed most.
+- Container exclusion is a name-prefix (`GITEA-ACTIONS-TASK`) plus a
+  `RunningFor` text heuristic — molecule containers owned by a live job are
+  protected only by naming convention, not by an ownership claim.
+- Cleanup logic is duplicated inline between `docker-prune.service` and the
+  healthcheck script — the two paths already diverge (volume pruning exists in
+  one tier only).
+- There is no admission control: under disk pressure the runner keeps
+  accepting new jobs while cleanup races in-flight work.
+- The runner config never declares `capacity`, and nothing prevents
+  installing a runner on a production host.
 
 ## Approach
-Add three skill files under `.devin/skills/`. Two are shared skills
-(`dependency-graph`, `skill-creation`) that must be identical across
-repos; one is grm-specific (`deployment-coordination`). All three
-follow the standard skill structure (H1 title, When to Invoke,
-Prerequisites, core content) and reference real make targets, file
-paths, and API endpoints.
 
-REQ-1: Add `.devin/skills/dependency-graph/SKILL.md` — shared skill mapping the oblachno ecosystem (repos, produces/consumers, dependency chain, correct change order, state verification)
+REQ-1: Define an ownership-lease label contract. Producers tag containers,
-REQ-2: Add `.devin/skills/deployment-coordination/SKILL.md` — grm-specific skill covering release flow, downstream consumer, coordinating a grm change, and common mistakes
+images, volumes and networks with `org.oblachno.lease-until` (epoch seconds)
-REQ-3: Add `.devin/skills/skill-creation/SKILL.md` — shared skill for creating, validating, and maintaining skills (structure, quality standards, scope rules, automated validation, checklist)
+and `org.oblachno.owner` (free-form run/job id). All cleanup paths must never
+remove an object whose `lease-until` is in the future; expired leases are
+reclaimable. Existing `GITEA-ACTIONS-TASK` name-prefix and `status=exited`
+guards are retained for unlabeled objects.
 
-## Files Affected
+REQ-2: Introduce a single shared cleanup script
-- `.devin/skills/dependency-graph/SKILL.md` (new)
+(`runner-cleanup.sh`, templated next to the healthcheck script) invoked with
-- `.devin/skills/deployment-coordination/SKILL.md` (new)
+`--tier routine|pressure|critical`. It replaces all inline prune logic in
-- `.devin/skills/skill-creation/SKILL.md` (new)
+`docker-prune.service` and the healthcheck. Every prune is scoped: volumes and
-- `docs/specs/GRM-173.md` (new)
+networks get `label!=`/`until=` filters at every tier (no unfiltered volume or
+network prune remains); images matching `gitea_runner_keep_images` are never
+removed, so warm base layers survive critical pressure.
+
+REQ-3: Watermark-tiered behavior. `routine` (timer) prunes aged resources only.
+`pressure` (disk >= warn) prunes unowned resources older than 1h. `critical`
+(disk >= critical) drops age limits but still honors leases, keep-images, and
+never removes running or `GITEA-ACTIONS-TASK` containers.
+
+REQ-4: Admission control under disk pressure. When disk is >= critical and no
+`GITEA-ACTIONS-TASK` container is running, the healthcheck writes a marker
+file and stops `gitea-runner.service` (the runner simply stops fetching new
+jobs). The service is restarted and the marker cleared by a later healthcheck
+once disk drops below the warn threshold. In-flight jobs are never killed by
+the admission path. Controlled by `gitea_runner_disk_admission_enabled`.
+
+REQ-5: Declare physical-host capacity explicitly:
+`runner.capacity: {{ gitea_runner_capacity }}` in the act_runner config
+(default 1, matching upstream default).
+
+REQ-6: Production-host exclusion. The role fails early when the target host
+carries the production marker file `/etc/oblachno/production-host` or when
+`gitea_runner_on_production_host` is true, unless
+`gitea_runner_allow_production_host` overrides. Infra-side provisioning of the
+marker is a follow-up task.
+
+Historical spec for the colliding task ID:
+[GRM-173-skills-historical](GRM-173-skills-historical.md).
 
 ## Test Plan
-- Verify all three SKILL.md files follow the required structure (H1, When to Invoke, Prerequisites)
+
-- Verify referenced make targets and file paths are accurate
+- `template-content` molecule scenario: assert the prune service calls
-- Run `make pytest-cov` to confirm no test regressions (skills are docs-only, no code changes)
+  `runner-cleanup.sh`, assert lease filters and keep-images logic render in
-- Confirm shared skills (`dependency-graph`, `skill-creation`) are ready for cross-repo sync
+  the cleanup script, assert `capacity:` renders in the runner config.
+- `default` molecule scenario: assert the cleanup script is installed and
+  executable.
+- `bash -n` syntax check on rendered templates during development.
+- `make lint-all`, `make pytest-cov`, fast molecule for the changed role.
 
 ## Deploy Plan
-- Merge to master via auto-merge workflow
+
-- No runtime changes; documentation-only (`.devin/**` is infrastructure path, no release triggered)
+- Merge to master via auto-merge; post-merge publishes the package and
+  auto-creates the infra dependency-bump PR. Runner hosts pick up the change
+  on the next `grm install`/update run — no manual host action.
+- Producers emitting lease labels (molecule distribution, CI jobs) are a
+  separate devx-side change; until then the guards degrade gracefully to the
+  existing name-prefix/age behavior.
 
 ## Rollback Plan
-- Revert the merge commit; skill files are removed, no functional impact
+
+- Revert the merge commit; re-run `grm install` to redeploy the previous
+  prune/healthcheck units. No persistent state or data migration — the
+  marker file under `gitea_runner_data_dir` is removed by the previous
+  template's absence (or harmless if left behind).
 
 ## Acceptance Criteria
-- [x] REQ-1: `.devin/skills/dependency-graph/SKILL.md` exists with ecosystem map, dependency chain, correct change order, and state verification sections
+
-- [x] REQ-2: `.devin/skills/deployment-coordination/SKILL.md` exists with release flow, downstream consumer table, coordination steps, and common mistakes
+- [x] REQ-1: `org.oblachno.lease-until`/`org.oblachno.owner` labels are honored
-- [x] REQ-3: `.devin/skills/skill-creation/SKILL.md` exists with skill structure template, quality standards, scope rules, automated validation, and creation checklist
+  by every cleanup path; valid leases are never removed, expired leases are.
+- [x] REQ-2: single shared `runner-cleanup.sh` used by prune service and
+  healthcheck; no unfiltered `system prune --volumes`, `volume prune`, or
+  `network prune` remains; `gitea_runner_keep_images` never removed.
+- [x] REQ-3: three tiers behave as specified (routine/pressure/critical).
+- [x] REQ-4: critical pressure with zero in-flight job containers stops
+  admission via marker + service stop; recovery resumes automatically.
+- [x] REQ-5: `runner.capacity` rendered in `config.yaml`.
+- [x] REQ-6: role fails on production-marked hosts unless explicitly allowed.