v0.21.0: agent reliability overhaul — plan integrity, target validation, observability pipelines, learning loop

P0 — stop the bleeding:
- prevent premature complete_task(success) when goal involves reachability
- validate run targets: block host-only commands (qm/pct/pvesh) on LXC/VM
- bump MCP client timeout 30s→120s to stop 'context deadline exceeded'

P1 — fix the plan system:
- add replaced_reason column to session_plan_steps (migration 030)
- track WHY steps are replaced (wrong_diagnosis/scope_change/superseded/etc)
- force fresh propose_plan on session resume (reopenSession marks old plan)

P2 — cognitive guardrails:
- SOUL.md scope-gate rule: ask before chasing unrelated subsystems
- auto-upsert knowledge entry on every session close

P3 — observability (all were empty/NULL):
- populate agent_activity.token_count from LLM usage (was always NULL)
- populate nomos_plan_executions linking executions to sessions
- write plan_completion_rate metric on task close

P4 — learning loop (all were empty/NULL):
- auto-classify every run call → classifications table (was 0 rows)
- auto-feedback on session close (was 0 rows)

internal/mcp/server.go

@@ -647,6 +647,19 @@ func classifyAndGate(ctx context.Context, pool *db.Pool, agentID, targetID uuid.
 		return textResult("No plan for this session. Call set_goal then propose_plan before run — even read-only tasks require a one-step plan. A one-step plan (\"Inspect X, report, write back\") is fine for trivial questions; the gate is about ordering, not approval. Read-only commands still auto-execute once a plan exists.")
 	}
 
+	// Target validation: host-only commands (qm, pct, pvesh, iptables) must
+	// not be dispatched against lxc:/vm: targets — those aren't Proxmox hosts
+	// and don't have these tools. Caught live 2026-08-04: the agent ran
+	// `qm stop 100` against lxc:dns, wasting a turn.
+	if cmdPrefix, hostOnly := hostOnlyCommand(command); hostOnly && !strings.HasPrefix(targetSlug, "host:") {
+		hostSuggestion := resolveProxmoxHostSlug(ctx, pool, targetSlug, "")
+		if hostSuggestion == "" {
+			hostSuggestion = "host:hubris or host:strong"
+		}
+		return textResult(fmt.Sprintf("Cannot run %q on %s — %s is a Proxmox host command. Use target %s instead.",
+			cmdPrefix, targetSlug, cmdPrefix, hostSuggestion))
+	}
+
 	// Dedup: an identical pending command (same target, command, and
 	// purpose) blocks a re-request — stops a tool-calling loop from queuing
 	// the same approval repeatedly.
@@ -721,8 +734,38 @@ func classifyAndGate(ctx context.Context, pool *db.Pool, agentID, targetID uuid.
 				SELECT 1 FROM relationships
 				WHERE source_id = t.id AND target_id = $1 AND type = 'involves' AND valid_to IS NULL)`,
 			id, "task:"+sessionID)
+		// Link execution to session for auto-continuation (nomos_plan_executions
+		// was always empty — executions were never traceable back to sessions).
+		if sid, serr := uuid.Parse(sessionID); serr == nil {
+			pool.Exec(ctx, `
+				INSERT INTO nomos_plan_executions (execution_id, session_id)
+				VALUES ($1, $2) ON CONFLICT (execution_id) DO NOTHING`, id, sid)
+		}
 	}
 
+	// Auto-classify: write the classification decision to the classifications
+	// table (was always empty — 0 rows despite 1,884 executions). The route
+	// matches the auto-run vs queue-for-approval decision below.
+	classRoute := "escalate"
+	if riskClass == policy.RiskReadOnly || riskClass == policy.RiskReversibleLow {
+		classRoute = "auto-act"
+	} else if riskClass == policy.RiskConfigMutation && assentWindowActive(ctx, pool, agentID, sessionID) {
+		classRoute = "auto-act"
+	} else if riskClass == policy.RiskDestructive && destructiveWindowActive(ctx, pool, agentID, targetSlug, sessionID) {
+		classRoute = "auto-act"
+	}
+	classReason, _ := json.Marshal(map[string]string{
+		"command": command, "purpose": purpose, "target": targetSlug, "declared_risk": declaredRisk,
+	})
+	classID, _ := uuid.NewV7()
+	pool.Exec(ctx, `INSERT INTO entities (id, slug, type, name, attributes) VALUES ($1, $2, 'classification', $3, '{}')`,
+		classID, "classification:"+classID.String(), "classification for "+execSlug)
+	pool.Exec(ctx, `INSERT INTO classifications (entity_id, action, risk_class, route, reasoning, correlation_id)
+		VALUES ($1, $2, $3, $4, $5, $6)`,
+		classID, actionCol, riskClass, classRoute, classReason, correlationID)
+	// Link classification to execution.
+	pool.Exec(ctx, `UPDATE executions SET classification_id = $2 WHERE entity_id = $1`, id, classID)
+
 	// read_only and reversible_low both run unattended, as seeds/policy.yaml
 	// and .agents/OIKOS.md declare ("reversible_low — restart, cache clear,
 	// sync pull. Unattended + ledger.").
@@ -840,6 +883,41 @@ func executeApprovedViaAPI(ctx context.Context, execID uuid.UUID, targetSlug, ac
 // for config_mutation auto-run now. See sessionHasPlan for the plan-existence
 // check used by the P1 plan-first gate.
 
+// hostOnlyCommands maps command prefixes that are only valid on Proxmox host
+// targets (not LXCs or VMs). Running these against an lxc: or vm: target
+// always fails with "command not found" and wastes a turn.
+var hostOnlyCommands = map[string]bool{
+	"qm":    true,
+	"pct":   true,
+	"pvesh": true,
+}
+
+// hostOnlyCommand checks whether the leading word of cmd is a host-only
+// command. Returns the command word and true if the command can only run on
+// a host: target.
+func hostOnlyCommand(cmd string) (string, bool) {
+	trimmed := strings.TrimSpace(cmd)
+	parts := strings.Fields(trimmed)
+	if len(parts) == 0 {
+		return "", false
+	}
+	first := parts[0]
+	// Check for shell wrappers: bash -c 'actual_cmd', sh -c 'actual_cmd'
+	if (first == "bash" || first == "sh") && len(parts) >= 3 && parts[1] == "-c" {
+		// The actual command is inside the -c argument; extract the first word.
+		// This handles `bash -c 'qm stop 100'` but not deeply nested wrappers.
+		actual := strings.Trim(strings.Join(parts[2:], " "), "'\"")
+		if inner := strings.Fields(actual); len(inner) > 0 {
+			first = inner[0]
+		}
+	}
+	// Strip path: /usr/sbin/qm → qm
+	if idx := strings.LastIndexByte(first, '/'); idx >= 0 {
+		first = first[idx+1:]
+	}
+	return first, hostOnlyCommands[first]
+}
+
 // sessionHasPlan reports whether this nomos session has any plan step on
 // record (any generation, any status). Used by the P1 plan-first gate in
 // classifyAndGate to refuse `run` before `propose_plan` has been called.