Merge pull request 'chore: gardener housekeeping' (#1029) from chore/gardener-20260418-2226 into main

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

AGENTS.md

@@ -1,4 +1,4 @@
-<!-- last-reviewed: 8fc3ba5b59cd6cb15bd01ca0658cfea2bcb12068 -->
+<!-- last-reviewed: 3fb2de4a8ab500707665adfbf954aa1921ae7775 -->
 # Disinto — Agent Instructions
 
 ## What this repo is
@@ -39,7 +39,7 @@ disinto/                 (code repo)
 │                  hooks/ — Claude Code session hooks (on-compact-reinject, on-idle-stop, on-phase-change, on-pretooluse-guard, on-session-end, on-stop-failure)
 │                  init/nomad/ — cluster-up.sh, install.sh, vault-init.sh, lib-systemd.sh (Nomad+Vault Step 0 installers, #821-#825); wp-oauth-register.sh (Forgejo OAuth2 app + Vault KV seeder for Woodpecker, S3.3); deploy.sh (dependency-ordered Nomad job deploy + health-wait, S4)
 ├── nomad/         server.hcl, client.hcl (allow_privileged for woodpecker-agent, S3-fix-5), vault.hcl — HCL configs deployed to /etc/nomad.d/ and /etc/vault.d/ by lib/init/nomad/cluster-up.sh
-│                  jobs/ — Nomad jobspecs: forgejo.hcl (Vault secrets via template, S2.4); woodpecker-server.hcl + woodpecker-agent.hcl (host-net, docker.sock, Vault KV, S3.1-S3.2); agents.hcl (7 roles, llama, Vault-templated bot tokens, S4.1); vault-runner.hcl (parameterized batch dispatch, S5.3); staging.hcl (Caddy file-server, S5.2); chat.hcl (Claude chat UI, Vault OAuth secrets, S5.2); edge.hcl (Caddy proxy + dispatcher sidecar, S5.1)
+│                  jobs/ — Nomad jobspecs: forgejo.hcl (Vault secrets via template, S2.4); woodpecker-server.hcl + woodpecker-agent.hcl (host-net, docker.sock, Vault KV, S3.1-S3.2); agents.hcl (7 roles, llama, Vault-templated bot tokens, S4.1); vault-runner.hcl (parameterized batch dispatch, S5.3); staging.hcl (Caddy file-server, dynamic port — edge discovers via service registration, S5.2); chat.hcl (Claude chat UI, tmpfs via mount block, Vault OAuth secrets, S5.2); edge.hcl (Caddy proxy + dispatcher sidecar, S5.1)
 ├── projects/      *.toml.example — templates; *.toml — local per-box config (gitignored)
 ├── formulas/      Issue templates (TOML specs for multi-step agent tasks)
 ├── docker/        Dockerfiles and entrypoints: reproduce, triage, edge dispatcher, chat (server.py, entrypoint-chat.sh, Dockerfile, ui/)

architect/AGENTS.md

@@ -1,4 +1,4 @@
-<!-- last-reviewed: 8fc3ba5b59cd6cb15bd01ca0658cfea2bcb12068 -->
+<!-- last-reviewed: 3fb2de4a8ab500707665adfbf954aa1921ae7775 -->
 # Architect — Agent Instructions
 
 ## What this agent is

dev/AGENTS.md

@@ -1,4 +1,4 @@
-<!-- last-reviewed: 8fc3ba5b59cd6cb15bd01ca0658cfea2bcb12068 -->
+<!-- last-reviewed: 3fb2de4a8ab500707665adfbf954aa1921ae7775 -->
 # Dev Agent
 
 **Role**: Implement issues autonomously — write code, push branches, address

docker/chat/Dockerfile

@@ -1,20 +1,22 @@
 # disinto-chat — minimal HTTP backend for Claude chat UI
 #
-# Small Debian slim base with Python runtime.
+# Small Debian slim base with Python runtime and Node.js.
 # Chosen for simplicity and small image size (~100MB).
 #
 # Image size: ~100MB (well under the 200MB ceiling)
 #
-# The claude binary is mounted from the host at runtime via docker-compose,
-# not baked into the image — same pattern as the agents container.
+# Claude CLI is baked into the image — same pattern as the agents container.
 
 FROM debian:bookworm-slim
 
-# Install Python (no build-time network access needed)
+# Install Node.js (required for Claude CLI) and Python
 RUN apt-get update && apt-get install -y --no-install-recommends \
-    python3 \
+    nodejs npm python3 \
     && rm -rf /var/lib/apt/lists/*
 
+# Install Claude Code CLI — chat backend runtime
+RUN npm install -g @anthropic-ai/claude-code@2.1.84
+
 # Non-root user — fixed UID 10001 for sandbox hardening (#706)
 RUN useradd -m -u 10001 -s /bin/bash chat
 

docker/edge/entrypoint-edge.sh

@@ -234,6 +234,13 @@ fi
   rm -f "$_fetch_log"
 done) &
 
+# Nomad template renders Caddyfile to /local/Caddyfile via service discovery;
+# copy it into the expected location if present (compose uses the mounted path).
+if [ -f /local/Caddyfile ]; then
+  cp /local/Caddyfile /etc/caddy/Caddyfile
+  echo "edge: using Nomad-rendered Caddyfile from /local/Caddyfile" >&2
+fi
+
 # Caddy as main process — run in foreground via wait so background jobs survive
 # (exec replaces the shell, which can orphan backgrounded subshells)
 caddy run --config /etc/caddy/Caddyfile --adapter caddyfile &

gardener/AGENTS.md

@@ -1,4 +1,4 @@
-<!-- last-reviewed: 8fc3ba5b59cd6cb15bd01ca0658cfea2bcb12068 -->
+<!-- last-reviewed: 3fb2de4a8ab500707665adfbf954aa1921ae7775 -->
 # Gardener Agent
 
 **Role**: Backlog grooming — detect duplicate issues, missing acceptance

gardener/dust.jsonl

@@ -1 +0,0 @@
-{"issue":915,"group":"lib/generators.sh","title":"remove no-op sed in generate_compose --build mode","reason":"sed replaces agents: with itself — no behavior change; single-line removal","ts":"2026-04-17T01:04:05Z"}

gardener/pending-actions.json

@@ -1,12 +1 @@
-[
-  {
-    "action": "edit_body",
-    "issue": 996,
-    "body": "Flagged by AI reviewer in PR #993.\n\n## Problem\n\nThe consul-template with/else/end pattern using aggressive whitespace trimming (e.g. `{{- with secret ... -}}` / `{{- else -}}` / `{{- end }}` then immediately `{{- with`) strips all newlines between consecutive single-variable env blocks at parse time. This would render the secrets env file as one concatenated line (`GITHUB_TOKEN=valCODEBERG_TOKEN=val...`), which Nomad's `env = true` cannot parse correctly.\n\n## Why not blocked\n\nagents.hcl has been runtime-tested (S4-fix-6 and S4-fix-7 made observable runtime fixes). If the env file were broken, all bot tokens would be absent — a loud, observable failure. This suggests consul-template may handle whitespace trimming differently from raw Go text/template. Needs runtime verification.\n\n## Verification\n\nDeploy either job and inspect the rendered secrets file:\n```\nnomad alloc exec <alloc-id> cat /secrets/bots.env\n```\nConfirm each KEY=VALUE pair is on its own line.\n\n---\n*Auto-created from AI review*\n\n## Affected files\n- `nomad/jobs/agents.hcl` — bots.env template (lines 147-189)\n- `nomad/jobs/vault-runner.hcl` — runner.env template (PR #993)\n\n## Acceptance criteria\n- [ ] Deploy `agents` or `vault-runner` job on factory host\n- [ ] Inspect rendered secrets file: `nomad alloc exec <alloc-id> cat /secrets/bots.env`\n- [ ] Confirm each KEY=VALUE pair is on its own line (not concatenated)\n- [ ] If broken: fix whitespace trimming to preserve newlines between blocks; if fine, close as not-a-bug"
-  },
-  {
-    "action": "add_label",
-    "issue": 996,
-    "label": "backlog"
-  }
-]
+[]

lib/AGENTS.md

@@ -1,4 +1,4 @@
-<!-- last-reviewed: 8fc3ba5b59cd6cb15bd01ca0658cfea2bcb12068 -->
+<!-- last-reviewed: 3fb2de4a8ab500707665adfbf954aa1921ae7775 -->
 # Shared Helpers (`lib/`)
 
 All agents source `lib/env.sh` as their first action. Additional helpers are

lib/generators.sh

@@ -657,7 +657,6 @@ COMPOSEEOF
 
   # In build mode, replace image: with build: for locally-built images
   if [ "$use_build" = true ]; then
-    sed -i 's|^\(  agents:\)|\1|' "$compose_file"
     sed -i '/^    image: ghcr\.io\/disinto\/agents:/{s|image: ghcr\.io/disinto/agents:.*|build:\n      context: .\n      dockerfile: docker/agents/Dockerfile\n    pull_policy: build|}' "$compose_file"
     sed -i '/^    image: ghcr\.io\/disinto\/edge:/{s|image: ghcr\.io/disinto/edge:.*|build: ./docker/edge\n    pull_policy: build|}' "$compose_file"
   fi

nomad/AGENTS.md

@@ -1,4 +1,4 @@
-<!-- last-reviewed: 8fc3ba5b59cd6cb15bd01ca0658cfea2bcb12068 -->
+<!-- last-reviewed: 3fb2de4a8ab500707665adfbf954aa1921ae7775 -->
 # nomad/ — Agent Instructions
 
 Nomad + Vault HCL for the factory's single-node cluster. These files are
@@ -19,8 +19,8 @@ see issues #821–#992 for the step breakdown.
 | `jobs/woodpecker-server.hcl` | submitted via `lib/init/nomad/deploy.sh` | Woodpecker CI server; host networking, Vault KV for `WOODPECKER_AGENT_SECRET` + Forgejo OAuth creds (S3.1) |
 | `jobs/woodpecker-agent.hcl` | submitted via `lib/init/nomad/deploy.sh` | Woodpecker CI agent; host networking, `docker.sock` mount, Vault KV for `WOODPECKER_AGENT_SECRET`; `WOODPECKER_SERVER` uses `${attr.unique.network.ip-address}:9000` (Nomad interpolation) — port binds to LXC alloc IP, not localhost (S3.2, S3-fix-6, #964) |
 | `jobs/agents.hcl` | submitted via `lib/init/nomad/deploy.sh` | All 7 agent roles (dev, review, gardener, planner, predictor, supervisor, architect) + llama variant; Vault-templated bot tokens via `service-agents` policy; `force_pull = false` — image is built locally by `bin/disinto --with agents`, no registry (S4.1, S4-fix-2, S4-fix-5, #955, #972, #978) |
-| `jobs/staging.hcl` | submitted via `lib/init/nomad/deploy.sh` | Caddy file-server mounting `docker/` as `/srv/site:ro`; no Vault integration; internal-only via edge proxy (S5.2, #989) |
-| `jobs/chat.hcl` | submitted via `lib/init/nomad/deploy.sh` | Claude chat UI; custom `disinto/chat:local` image; sandbox hardening (cap_drop ALL, tmpfs, pids_limit 128); Vault-templated OAuth secrets via `service-chat` policy (S5.2, #989) |
+| `jobs/staging.hcl` | submitted via `lib/init/nomad/deploy.sh` | Caddy file-server mounting `docker/` as `/srv/site:ro`; no Vault integration; **dynamic host port** (no static 80 — edge owns 80/443, collision fixed in S5-fix-7 #1018); edge discovers via Nomad service registration (S5.2, #989) |
+| `jobs/chat.hcl` | submitted via `lib/init/nomad/deploy.sh` | Claude chat UI; custom `disinto/chat:local` image; sandbox hardening (cap_drop ALL, **tmpfs via mount block** not `tmpfs=` arg — S5-fix-5 #1012, pids_limit 128); Vault-templated OAuth secrets via `service-chat` policy (S5.2, #989) |
 | `jobs/edge.hcl` | submitted via `lib/init/nomad/deploy.sh` | Caddy reverse proxy + dispatcher sidecar; routes /forge, /woodpecker, /staging, /chat; uses `disinto/edge:local` image built by `bin/disinto --with edge`; Vault-templated ops-repo creds via `service-dispatcher` policy (S5.1, #988) |
 
 Nomad auto-merges every `*.hcl` under `-config=/etc/nomad.d/`, so the

nomad/jobs/chat.hcl

@@ -89,18 +89,22 @@ job "chat" {
       config {
         image      = "disinto/chat:local"
         force_pull = false
-        # Sandbox hardening (#706): cap_drop ALL (no Linux capabilities)
-        # tmpfs /tmp for runtime files (64MB)
-        # pids_limit 128 (prevent fork bombs)
+        # Sandbox hardening (#706): cap_drop ALL, pids_limit 128, tmpfs /tmp
         # ReadonlyRootfs enforced via entrypoint script (fails if running as root)
+        cap_drop   = ["ALL"]
+        pids_limit = 128
+        mount {
+          type     = "tmpfs"
+          target   = "/tmp"
+          readonly = false
+          tmpfs_options {
+            size = 67108864  # 64MB in bytes
+          }
+        }
         # Security options for sandbox hardening
         # apparmor=unconfined needed for Claude CLI ptrace access
         # no-new-privileges prevents privilege escalation
-        cap_drop    = ["ALL"]
-        pids_limit  = 128
         security_opt = ["apparmor=unconfined", "no-new-privileges"]
-        # tmpfs mount via volumes config (Nomad Docker driver)
-        volumes = ["tmpfs:/tmp:size=64m"]
       }
 
       # ── Volume mounts ──────────────────────────────────────────────────────

nomad/jobs/edge.hcl

@@ -114,6 +114,58 @@ job "edge" {
         read_only   = false
       }
 
+      # ── Caddyfile via Nomad service discovery (S5-fix-7, issue #1018) ────
+      # Renders staging upstream from Nomad service registration instead of
+      # hardcoded staging:80. Caddy picks up /local/Caddyfile via entrypoint.
+      template {
+        destination = "local/Caddyfile"
+        change_mode = "restart"
+        data        = <<EOT
+# Caddyfile — edge proxy configuration (Nomad-rendered)
+# Staging upstream discovered via Nomad service registration.
+
+:80 {
+    # Redirect root to Forgejo
+    handle / {
+        redir /forge/ 302
+    }
+
+    # Reverse proxy to Forgejo
+    handle /forge/* {
+        reverse_proxy forgejo:3000
+    }
+
+    # Reverse proxy to Woodpecker CI
+    handle /ci/* {
+        reverse_proxy woodpecker:8000
+    }
+
+    # Reverse proxy to staging — dynamic port via Nomad service discovery
+    handle /staging/* {
+{{ range nomadService "staging" }}        reverse_proxy {{ .Address }}:{{ .Port }}
+{{ end }}    }
+
+    # Chat service — reverse proxy to disinto-chat backend (#705)
+    # OAuth routes bypass forward_auth — unauthenticated users need these (#709)
+    handle /chat/login {
+        reverse_proxy chat:8080
+    }
+    handle /chat/oauth/callback {
+        reverse_proxy chat:8080
+    }
+    # Defense-in-depth: forward_auth stamps X-Forwarded-User from session (#709)
+    handle /chat/* {
+        forward_auth chat:8080 {
+            uri /chat/auth/verify
+            copy_headers X-Forwarded-User
+            header_up X-Forward-Auth-Secret {$FORWARD_AUTH_SECRET}
+        }
+        reverse_proxy chat:8080
+    }
+}
+EOT
+      }
+
       # ── Non-secret env ───────────────────────────────────────────────────
       env {
         FORGE_URL       = "http://forgejo:3000"

nomad/jobs/staging.hcl

@@ -10,8 +10,8 @@
 #   served to staging environment users.
 #
 # Network:
-#   No external port exposed — edge proxy routes to it internally.
-#   Service discovery via Nomad native provider for internal routing.
+#   Dynamic host port — edge discovers via Nomad service registration.
+#   No static port to avoid collisions with edge (which owns 80/443).
 #
 # Not the runtime yet: docker-compose.yml is still the factory's live stack
 # until cutover. This file exists so CI can validate it and S5.2 can wire
@@ -27,11 +27,10 @@ job "staging" {
 
     # No Vault integration needed — no secrets required (static file server)
 
-    # Internal service — no external port. Edge proxy routes internally.
+    # Internal service — dynamic host port. Edge discovers via Nomad service.
     network {
       port "http" {
-        static = 80
-        to     = 80
+        to = 80
       }
     }
 

planner/AGENTS.md

@@ -1,4 +1,4 @@
-<!-- last-reviewed: 8fc3ba5b59cd6cb15bd01ca0658cfea2bcb12068 -->
+<!-- last-reviewed: 3fb2de4a8ab500707665adfbf954aa1921ae7775 -->
 # Planner Agent
 
 **Role**: Strategic planning using a Prerequisite Tree (Theory of Constraints),

predictor/AGENTS.md

@@ -1,4 +1,4 @@
-<!-- last-reviewed: 8fc3ba5b59cd6cb15bd01ca0658cfea2bcb12068 -->
+<!-- last-reviewed: 3fb2de4a8ab500707665adfbf954aa1921ae7775 -->
 # Predictor Agent
 
 **Role**: Abstract adversary (the "goblin"). Runs a 2-step formula

review/AGENTS.md

@@ -1,4 +1,4 @@
-<!-- last-reviewed: 8fc3ba5b59cd6cb15bd01ca0658cfea2bcb12068 -->
+<!-- last-reviewed: 3fb2de4a8ab500707665adfbf954aa1921ae7775 -->
 # Review Agent
 
 **Role**: AI-powered PR review — post structured findings and formal

supervisor/AGENTS.md

@@ -1,4 +1,4 @@
-<!-- last-reviewed: 8fc3ba5b59cd6cb15bd01ca0658cfea2bcb12068 -->
+<!-- last-reviewed: 3fb2de4a8ab500707665adfbf954aa1921ae7775 -->
 # Supervisor Agent
 
 **Role**: Health monitoring and auto-remediation, executed as a formula-driven

vault/policies/AGENTS.md

@@ -1,4 +1,4 @@
-<!-- last-reviewed: 8fc3ba5b59cd6cb15bd01ca0658cfea2bcb12068 -->
+<!-- last-reviewed: 3fb2de4a8ab500707665adfbf954aa1921ae7775 -->
 # vault/policies/ — Agent Instructions
 
 HashiCorp Vault ACL policies for the disinto factory. One `.hcl` file per