From 09051881fa767867d06287aa27073b3c7c45ef0a Mon Sep 17 00:00:00 2001 From: HananNouman Date: Fri, 22 May 2026 22:53:45 +0300 Subject: [PATCH] fix(hermes): host-side chown for CRD child agent PVCs on Linux k3d PR #481 only repaired hermes- volumes after hermes.Sync (master agent). Child agents live under agent- and are provisioned by the controller or agent-factory without that path, so hermes-data stayed 1000:1000 while Hermes runs as 10000:10000 and crash-looped on Permission denied under /data/.hermes. Extend EnsureHermesDataPVCOwnership to agent-/hermes-data, call it from obol agent new and obol sell demo quant, and add obol agent repair-perms for factory-only creates that cannot docker-exec the k3d node from in-cluster. Co-authored-by: Cursor --- cmd/obol/agent.go | 27 ++++++++++ cmd/obol/agent_crd.go | 4 ++ cmd/obol/sell_agent.go | 1 + internal/agentcrd/agent.go | 11 +++- internal/agentcrd/agent_test.go | 13 +++++ internal/embed/skills/agent-factory/SKILL.md | 13 +++++ internal/hermes/hermes.go | 56 +++++++++++--------- plans/crd-agent-hermes-pvc-chown.md | 44 +++++++++++++++ 8 files changed, 141 insertions(+), 28 deletions(-) create mode 100644 plans/crd-agent-hermes-pvc-chown.md diff --git a/cmd/obol/agent.go b/cmd/obol/agent.go index f3062f66..748de8ab 100644 --- a/cmd/obol/agent.go +++ b/cmd/obol/agent.go @@ -200,6 +200,33 @@ Hermes/OpenClaw onboard flow used by the master agent.`, return listAgentInstances(cfg, cmd.String("runtime"), getUI(cmd)) }, }, + { + Name: "repair-perms", + Usage: "Fix Hermes data PVC ownership for a CRD child agent (Linux k3d)", + ArgsUsage: "", + Description: `Host-side chown for agent-/hermes-data so the child Hermes pod can write under /data. + +Use when a factory-spawned or CLI-created child agent crash-loops with +Permission denied under /data/.hermes on Linux k3d (KubeletInUserNamespace). + +This is a no-op on backends where in-pod init chown already works.`, + Action: func(ctx context.Context, cmd *cli.Command) error { + if cmd.NArg() != 1 { + return fmt.Errorf("agent name required: obol agent repair-perms ") + } + name := strings.TrimSpace(cmd.Args().First()) + if err := agentcrd.ValidateName(name); err != nil { + return err + } + if err := kubectl.EnsureCluster(cfg); err != nil { + return fmt.Errorf("Obol Stack is not running. Start it with `obol stack up` first") + } + u := getUI(cmd) + hermes.EnsureCRDAgentHermesPVCOwnership(cfg, name, u) + u.Successf("Repaired Hermes data volume ownership for agent %s", name) + return nil + }, + }, { Name: "delete", Usage: "Remove an agent instance and its cluster resources", diff --git a/cmd/obol/agent_crd.go b/cmd/obol/agent_crd.go index 7892982e..4b55c8ba 100644 --- a/cmd/obol/agent_crd.go +++ b/cmd/obol/agent_crd.go @@ -6,6 +6,7 @@ import ( "github.com/ObolNetwork/obol-stack/internal/agentcrd" "github.com/ObolNetwork/obol-stack/internal/config" + "github.com/ObolNetwork/obol-stack/internal/hermes" "github.com/ObolNetwork/obol-stack/internal/kubectl" "github.com/ObolNetwork/obol-stack/internal/ui" "github.com/urfave/cli/v3" @@ -132,6 +133,9 @@ func createCRDAgent(cfg *config.Config, u *ui.UI, opts createCRDAgentOptions) er } u.Successf("Agent %s/%s %s", agentcrd.Namespace(opts.Name), opts.Name, action) u.Infof("Reconciler will provision: namespace → %s deployment → status updates", "hermes") + // Host-side chown for agent-/hermes-data (#475). Master-agent Sync + // already does this for hermes-obol-agent; CRD children skip helm Sync. + hermes.EnsureCRDAgentHermesPVCOwnership(cfg, opts.Name, u) u.Infof("Inspect: kubectl get agent %s -n %s -o yaml", opts.Name, agentcrd.Namespace(opts.Name)) return nil } diff --git a/cmd/obol/sell_agent.go b/cmd/obol/sell_agent.go index 80b77315..24d3b041 100644 --- a/cmd/obol/sell_agent.go +++ b/cmd/obol/sell_agent.go @@ -348,6 +348,7 @@ func runAgentBackedDemo( } u.Successf("Agent %s/%s created (skills: %s)", agentcrd.Namespace(agentName), agentName, strings.Join(spec.Agent.Skills, ", ")) + hermes.EnsureCRDAgentHermesPVCOwnership(cfg, agentName, u) } // 2. Build and apply the agent-typed ServiceOffer. diff --git a/internal/agentcrd/agent.go b/internal/agentcrd/agent.go index 02ca080e..ddbd1f00 100644 --- a/internal/agentcrd/agent.go +++ b/internal/agentcrd/agent.go @@ -31,12 +31,19 @@ func Namespace(name string) string { return "agent-" + name } +// HermesDataPVCHostPath is the host-side directory backing the child agent's +// hermes-data PVC (/agent-/hermes-data). local-path-provisioner +// maps the PVC here; HostHomePath writes under .../hermes-data/.hermes. +func HermesDataPVCHostPath(cfg *config.Config, name string) string { + desc := agentruntime.Describe(agentruntime.Hermes) + return filepath.Join(cfg.DataDir, Namespace(name), desc.DataPVCName) +} + // HostHomePath is where the agent's .hermes data lives on the host. The // cluster mounts this into the Hermes pod via hostPath; writing // SOUL.md/skills here puts them inside the pod automatically. func HostHomePath(cfg *config.Config, name string) string { - desc := agentruntime.Describe(agentruntime.Hermes) - return filepath.Join(cfg.DataDir, Namespace(name), desc.DataPVCName, desc.HomeDir) + return filepath.Join(HermesDataPVCHostPath(cfg, name), agentruntime.Describe(agentruntime.Hermes).HomeDir) } // HostSkillsPath is the per-agent skills dir. OBOL_SKILLS_DIR points here diff --git a/internal/agentcrd/agent_test.go b/internal/agentcrd/agent_test.go index a2aa5973..636a563a 100644 --- a/internal/agentcrd/agent_test.go +++ b/internal/agentcrd/agent_test.go @@ -114,6 +114,19 @@ func TestBuildAgent_Populated(t *testing.T) { } } +func TestHermesDataPVCHostPath(t *testing.T) { + cfg := &config.Config{DataDir: "/data/obol"} + got := HermesDataPVCHostPath(cfg, "quant-rc4") + want := "/data/obol/agent-quant-rc4/hermes-data" + if got != want { + t.Fatalf("HermesDataPVCHostPath = %q, want %q", got, want) + } + home := HostHomePath(cfg, "quant-rc4") + if home != want+"/.hermes" { + t.Fatalf("HostHomePath = %q, want %q/.hermes", home, want) + } +} + func TestSeedHostFiles_FreshAgent(t *testing.T) { dir := t.TempDir() cfg := &config.Config{DataDir: dir} diff --git a/internal/embed/skills/agent-factory/SKILL.md b/internal/embed/skills/agent-factory/SKILL.md index 1808e70e..4d6d4335 100644 --- a/internal/embed/skills/agent-factory/SKILL.md +++ b/internal/embed/skills/agent-factory/SKILL.md @@ -40,3 +40,16 @@ python3 scripts/factory.py create medical-advisor \ - The profile seed Secret is named `hermes-profile-seed` and contains `profile.tar.gz`. - Runtime environment overrides go in the optional `hermes-env` Secret. - The factory intentionally writes deterministic resource names only. + +### Linux k3d: fix Hermes PVC permissions after create + +On Linux k3d (`KubeletInUserNamespace`), the child Hermes pod may crash-loop with +`Permission denied` under `/data/.hermes` until the host-side data directory is +chowned. The factory runs in-cluster and cannot do that itself — run from the host: + +```bash +obol agent repair-perms +``` + +`obol agent new` and `obol sell demo quant` run this automatically; factory-only +creates need the one-liner above (same workaround as rc1 manual `docker exec` chown). diff --git a/internal/hermes/hermes.go b/internal/hermes/hermes.go index 59e8c66a..d64039b7 100644 --- a/internal/hermes/hermes.go +++ b/internal/hermes/hermes.go @@ -12,6 +12,7 @@ import ( "path/filepath" "strings" + "github.com/ObolNetwork/obol-stack/internal/agentcrd" "github.com/ObolNetwork/obol-stack/internal/agentruntime" "github.com/ObolNetwork/obol-stack/internal/config" "github.com/ObolNetwork/obol-stack/internal/dns" @@ -1306,10 +1307,12 @@ func fixRuntimeVolumeOwnership(cfg *config.Config, hostPath string, u *ui.UI) { owner := fmt.Sprintf("%d:%d", containerUID, containerGID) switch backendName { case "k3d": - if err := k3dNodeExec(cfg, hostPath, "chown -R "+owner+" {}"); err != nil && u != nil { + shellCmd := fmt.Sprintf("mkdir -p {} && chown -R %s {}", owner) + if err := k3dNodeExec(cfg, hostPath, shellCmd); err != nil && u != nil { u.Warnf("Failed to fix volume ownership for %s: %v", hostPath, err) } default: + _ = os.MkdirAll(hostPath, 0o755) _ = os.Chown(hostPath, containerUID, containerGID) } } @@ -1333,11 +1336,10 @@ func hermesPVCPaths(cfg *config.Config, id string) []string { } } -// ensureHermesPVCOwnership host-side chowns the Hermes PVC backing directories -// to containerUID:containerGID so the agent's init containers can write under -// /data on the first start. +// EnsureHermesDataPVCOwnership host-side chowns a Hermes data PVC backing dir +// to containerUID:containerGID so the agent can write under /data on first start. // -// Why this is needed (issue #475): +// Why this is needed (issue #475, extended to agent-* namespaces in #481 follow-up): // - The embedded k3d config (internal/embed/k3d-config.yaml) sets // KubeletInUserNamespace=true. Pod "root" maps to a host subuid that // lacks chown authority over the host bind-mount path provisioned by @@ -1352,38 +1354,28 @@ func hermesPVCPaths(cfg *config.Config, id string) []string { // k3d server container runs at the host Docker daemon's authority, which is // real root and is not subject to the kubelet's user-namespacing. // -// Best-effort. Waits up to 60s for each PVC to be Bound (local-path uses -// WaitForFirstConsumer, so the host dir doesn't exist until the consuming +// Best-effort. Waits up to 60s for the PVC to be Bound (local-path uses +// WaitForFirstConsumer, so the host dir may not exist until the consuming // pod is scheduled). On non-k3d backends fixRuntimeVolumeOwnership falls // back to a plain os.Chown. // // If a Hermes pod is currently stuck in Init:CrashLoopBackOff because of the // pre-fix permissions, deletes it so kubelet re-creates with the corrected // perms immediately rather than after exponential backoff (up to ~5 min). -// Skips the delete when no pod is stuck so repeated `Sync` calls -// (e.g. `obol model sync` after `obol model prefer`) do not gratuitously +// Skips the delete when no pod is stuck so repeated calls do not gratuitously // restart a healthy agent. -func ensureHermesPVCOwnership(cfg *config.Config, id string, u *ui.UI) { - namespace := agentruntime.Namespace(agentruntime.Hermes, id) +func EnsureHermesDataPVCOwnership(cfg *config.Config, namespace, hostPath string, u *ui.UI) { kubeconfigPath := filepath.Join(cfg.ConfigDir, "kubeconfig.yaml") kubectlBin := filepath.Join(cfg.BinDir, "kubectl") - // Wait only for the PVCs hermesPVCPaths chowns. remote-signer-keystores - // is intentionally NOT in this loop — see the doc comment on - // hermesPVCPaths for why. - for _, pvc := range []string{ - agentruntime.Describe(agentruntime.Hermes).DataPVCName, - } { - waitCmd := exec.Command(kubectlBin, - "wait", "--for=jsonpath={.status.phase}=Bound", - "--timeout=60s", "pvc/"+pvc, "-n", namespace) - waitCmd.Env = append(os.Environ(), "KUBECONFIG="+kubeconfigPath) - _ = waitCmd.Run() // best-effort; continue even on timeout - } + pvcName := agentruntime.Describe(agentruntime.Hermes).DataPVCName + waitCmd := exec.Command(kubectlBin, + "wait", "--for=jsonpath={.status.phase}=Bound", + "--timeout=60s", "pvc/"+pvcName, "-n", namespace) + waitCmd.Env = append(os.Environ(), "KUBECONFIG="+kubeconfigPath) + _ = waitCmd.Run() // best-effort; continue even on timeout - for _, p := range hermesPVCPaths(cfg, id) { - fixRuntimeVolumeOwnership(cfg, p, u) - } + fixRuntimeVolumeOwnership(cfg, hostPath, u) if hermesInitStuck(cfg, namespace) { deleteCmd := exec.Command(kubectlBin, @@ -1397,6 +1389,18 @@ func ensureHermesPVCOwnership(cfg *config.Config, id string, u *ui.UI) { } } +// EnsureCRDAgentHermesPVCOwnership applies EnsureHermesDataPVCOwnership for +// serviceoffer-controller child agents (namespace agent-). Call after +// `obol agent new`, `obol sell demo quant`, or when repairing factory-spawned +// children that crash-loop on Linux k3d. +func EnsureCRDAgentHermesPVCOwnership(cfg *config.Config, agentName string, u *ui.UI) { + EnsureHermesDataPVCOwnership(cfg, agentcrd.Namespace(agentName), agentcrd.HermesDataPVCHostPath(cfg, agentName), u) +} + +func ensureHermesPVCOwnership(cfg *config.Config, id string, u *ui.UI) { + EnsureHermesDataPVCOwnership(cfg, agentruntime.Namespace(agentruntime.Hermes, id), hermesPVCPaths(cfg, id)[0], u) +} + // hermesInitStuck reports whether at least one Hermes pod has an init // container in CrashLoopBackOff or an Error waiting state — the signature of // the perm-denied symptom this fix targets. Returns false on any kubectl diff --git a/plans/crd-agent-hermes-pvc-chown.md b/plans/crd-agent-hermes-pvc-chown.md new file mode 100644 index 00000000..0964ad2d --- /dev/null +++ b/plans/crd-agent-hermes-pvc-chown.md @@ -0,0 +1,44 @@ +# CRD child agent Hermes PVC ownership (Linux k3d) + +## Problem + +PR #481 (`ensureHermesPVCOwnership`) runs only after `hermes.Sync` for the **master** +agent (`hermes-obol-agent`). Child agents provisioned by the serviceoffer-controller +use namespace `agent-` and never go through helm Sync. + +On Linux k3d with `KubeletInUserNamespace`, local-path-provisioner creates +`/agent-/hermes-data` owned `1000:1000`. Hermes runs as `10000:10000` +and cannot write `/data/.hermes/home` → Init crash-loop. + +Factory-spawned children (`agent-factory` skill, in-cluster API) hit the same gap: +they never invoke host-side `obol` seed + chown. + +## Fix (branch `fix/crd-agent-hermes-pvc-chown`) + +| Call site | When | +|-----------|------| +| `hermes.EnsureCRDAgentHermesPVCOwnership` | Shared: wait PVC Bound → `mkdir -p` + `chown 10000:10000` via k3d `docker exec` | +| `obol agent new` | After Agent CR apply | +| `obol sell demo quant` | After new Agent CR apply | +| `obol agent repair-perms ` | Manual repair after `agent-factory` create | + +Master agent path unchanged: `ensureHermesPVCOwnership` → `EnsureHermesDataPVCOwnership` for `hermes-`. + +## Operator workaround (pre-fix / factory-only) + +```bash +obol agent repair-perms +# or legacy rc1 style: +docker exec k3d-obol-stack--server-0 \ + chown -R 10000:10000 /data/agent-/hermes-data +``` + +## Verify + +```bash +kubectl get pods -n agent- +# hermes should reach Running, not Init:CrashLoopBackOff + +obol sell status -n agent- +# Ready=True once Hermes is healthy +```